Choose AI / Coding

The best AI for coding

Which AI best solves real coding tasks (installing, debugging, compiling…), where to use it and how much it costs.

Our picks

Three ways to get it right: the best one regardless of price, the best value and the best one without paying.

The most powerful

Claude Sonnet 5.5

Anthropic

Score 64/100

It solves the most real coding tasks in independent tests.

You'll find it inClaude (web and app)

Pro: ≈ €18/mo$20/mo≈ €15$17/mo paid yearlyIt also has free use with limits in the app, in addition to the paid plans.

Just releasedNewModerately used

The best value

GPT-6.1 Sol

OpenAI

Score 52/100

It scores 52 versus 64 for the most powerful one, and it costs about 15 times less to use.

You'll find it inChatGPT (Plus plan)

Plus: ≈ €18/mo$20/mo

Score with its «high» reasoning mode, the best value; in its most powerful mode it reaches 56.

Just releasedProvenWidely used

The best free one (open source)

MiMo-V2.6-Pro

Xiaomi

Score 35/100

It's open source: you download it and use it without paying and without usage limits. It scores 35 versus 64 for the most powerful one.

You'll find it inDownload it (open source) · How to use it

Free (open source)

Needs a computer with about 625 GB of video memory (VRAM).

Open sourceNewModerately used

The top 10

One model per row (its best version). The variants (max, high, medium…) are in «For the data nerds», at the end.

1
Claude Sonnet 5.5Anthropic · You'll find it in: Claude (web and app)It also has free use with limits in the app, in addition to the paid plans.
Score 64/100~€11.29$12.65 per taskPro: ≈ €18/mo$20/mo
2
Claude Opus 5.5Anthropic · You'll find it in: Claude (Pro plan)
Score 60/100~€11.70$13.11 per taskPro: ≈ €18/mo$20/mo
3
GPT-6 AstraOpenAI · You'll find it in: ChatGPT (Plus plan)
Score 60/100~€5.23$5.86 per taskPlus: ≈ €18/mo$20/mo
4
Gemini 4 ArgonGoogle · You'll find it in: API only (for developers)
Score 57/100~€6.05$6.78 per taskDevelopers only
5
The best valueGPT-6.1 SolOpenAI · You'll find it in: ChatGPT (Plus plan)
Score 56/100~€1.63$1.82 per taskPlus: ≈ €18/mo$20/mo
6
Claude Fable 5.1Anthropic · You'll find it in: Claude (Pro plan)
Score 55/100~€14.08$15.78 per taskPro: ≈ €18/mo$20/mo
7
GLM-5.3Z AI · You'll find it in: Download it (open source) · How to use it
Score 42/100~€7.20$8.06 per taskFree (open source)
8
Qwen3.8 Max (0902)Alibaba · You'll find it in: API only (for developers)
Score 39/100~€16.67$18.68 per taskDevelopers only
9
MiMo-V2.6-ProXiaomi · You'll find it in: Download it (open source) · How to use it
Score 35/100~€0.39$0.44 per taskFree (open source)
10
Muse Spark 1.3Meta · You'll find it in: Meta AI (web and app)It also has free use with limits in the app, in addition to the paid plans.
Score 33/100~€6.19$6.94 per taskFree use with limits in the app

AI to decide, classify and search

For developers: models that don't write text, but pick an option, give a score or find the most similar item. They're used to filter, label or match.

Our picks for deciding

Measured with the Decision Index (real decision tasks, corrected for chance). The table has 17 decision models, 6 embedding models and 3 rerankers.

The most powerful

Perplexity Decider v1.1 27B

Perplexity

Score 63/100

It's the best at deciding in the Decision Index.

You'll find it inDownload it (open source) · How to use it

Free (open source)

⚠ Ojo: Barely tested: no independent ranking has confirmed it yet; Little used: almost nobody uses it yet. That's why we don't recommend it by default.

Just releasedOpen sourceBarely testedLittle used

The best value

Jev 1.13 (System One)

TypeSafe AI

Score 60/100

It scores 60 versus 63 for the most powerful one.

You'll find it inAPI only (for developers)

Developers only

NewWidely used

The best free one (open source)

Rune 26B-A4B v3

Surogate

Score 57/100

It's open source: you download it and use it without paying and without usage limits. It scores 57 versus 63 for the most powerful one.

You'll find it inDownload it (open source) · How to use it

Free (open source)

Needs a computer with about 20.0 GB of video memory (VRAM).

Open sourceNewModerately used

Table of decision, embedding and reranker models

26 models and variants · click ▸ on a row to see sources, memory and notes

Sort

Model	Tipo	Perf.DI / MTEB ⇅	Price€/1M$/1M tok ⇅	WeightsWeights⇅	MaturityMad.⇅	Latency	Paramstotal / act.	Memory ⇅VRAM · disco	License
Perplexity Decider v1.1 27B Perplexity	Typed decision	62.8	self-host	OpenOpen	Barely testedBarely t. ⬇ 599/mo HF · 1 rankingLittle usedLittle used	104 ms	27.8B	49.0 GB official (weights) ▣ 1× 80 GB disk 52.2 GB 1× 80 GB (BF16, official)	Apache 2.0
Tipo: Typed decision Featured: ★ N.º 1 (open) Weights: OpenOpen 🤗 Hugging Face · ↗ Web Maturity: Barely testedBarely t. Very little use · 599 HF downloads/mo · 1 leaderboard (Decision Index 0.3) · 4 days · Adoption: Little used (⬇ 599/mo HF · 1 ranking) (leaderboards: Decision Index 0.3; released: 2026-10-05; data from 2026-10-09) Score: #1 in the Decision Index, above Jev. What it does: Typed decision (choice/score/yes-no with probabilities) on Qwen3.8-27B with the causal mask lifted and a 255-option decision head. Memory: disk 52.2 GB (BF16) · Min. VRAM 49.0 GB official (weights) · fits on: 1× 80 GB (BF16, official) Official BF16 52.2 GB GGUF Q8_0 (community) 27.2 GB NVFP4 (community) 25.2 GB GGUF Q5_K_M (community) 19.6 GB MLX 4-bit (community) 15.3 GB Requirement: Model card: «a CUDA GPU with room for roughly 49 GiB of weights plus working memory». source Quality: Decision Index 0.3: 62.75 (Full score, #1 of 115; Jev 60.11) · public index 62.25 · the card states 61.56 on 0.2.1 · Latency: Median 104.1 ms (Decision Index 0.3, 1× RTX PRO 6000). · Price: Self-host (your own GPU) v1 (= «AutoJev-27B», 56.4 on 0.2.1) → v1.1 (Oct 5): more data (tasksource) and no causal mask. Its own code (autojev); the quantizations are from the community. License: Apache 2.0 Source: huggingface.co/spaces/multimodalart/jev-decision-index · 2026-10-08
Fastino GLiDE (28B, no-thinking) Fastino	Typed decision	60.2	n/d	ClosedClosed	NewNew 1 DI ranking · weights not publicNo usage dataNo data	98 ms	—	—	—
Tipo: Typed decision Featured: ★ N.º 2 DI Weights: ClosedClosed ↗ Web Maturity: NewNew By default: in Decision Index 0.3.1 as #2 · weights not public yet · Adoption: No usage data (1 DI ranking · weights not public) (leaderboards: Decision Index 0.3.1; released: 2026-09-30; data from 2026-10-10) Score: #2 in the DI; weights «coming soon». What it does: «Thinking»/no-thinking decision model: choice, yes/no and score with probabilities; does not generate free text. Weights «coming soon» according to DI. Memory: Weights not public yet Quality: Decision Index 0.3.1: Full 60.21 (#2 of 116) · public 59.06. Only behind Perplexity Decider v1.1 (62.75). · Latency: Median 98.4 ms (Decision Index 0.3.1). · Price: n/a (API; public price not listed in DI) Entry from the official 0.3.1 leaderboard (generated 2026-10-10). DI metadata: «Open weights coming soon». Not to be confused with GLiNER. Source: huggingface.co/spaces/multimodalart/jev-decision-index · 2026-10-10
Jev 1.13 (System One) TypeSafe AI	Typed decision	60.1	€0.037$0.042/1M	ClosedClosed	NewNew OpenRouter top 20 · #9 · 1 rankingWidely usedWidely used	524 ms (API)	—	—	—
Tipo: Typed decision Featured: ★ Best value (widely used) Weights: ClosedClosed ↗ Web Maturity: NewNew Released 22 days ago · top 20 OpenRouter · 1 leaderboard (Decision Index 0.3) · Adoption: Widely used (OpenRouter top 20 · #9 · 1 ranking) (leaderboards: Decision Index 0.3; released: 2026-09-17; data from 2026-10-09) Score: The only API in the top; 64k context; can't be fine-tuned. What it does: Reads a state (text/JSON) and answers typed questions: choice (up to 255 options), score (2–10 levels) or yes/no, with probabilities and confidence. Without generating text. Memory: Hosted (not downloadable) Quality: Decision Index 0.3: 60.11 (Full score, #3 of 115) · public index 57.96 · Latency: Median 524 ms per request measured by the Decision Index (hosted API, network round trip); TypeSafe advertises 70–500 ms. · Price: $0.042/1M input tokens; output free In Decision Index 0.3 (Oct 7) it comes 3rd, tied with Fastino GLiDE (weights «coming soon») and behind Perplexity Decider v1.1 (open). 64k context; best in English; cannot be fine-tuned. Also on Vercel AI Gateway (typesafe-ai/jev). Source: huggingface.co/spaces/multimodalart/jev-decision-index · 2026-10-08
Torchcast Decision 27B Torchcast AI	Typed decision	59.9	self-host	OpenOpen	Barely testedBarely t. ⬇ 602/mo HF · 1 rankingLittle usedLittle used	100 ms	27.8B	20.0 GB approx. ▣ 1 consumer GPU ✓ Fits on your computer disk 54.7 GB 1× 24 GB (approx.)	CC BY-NC 4.0 (non-commercial)
Tipo: Typed decision Weights: OpenOpen 🤗 Hugging Face · ↗ Web Maturity: Barely testedBarely t. Very little use · 602 HF downloads/mo · 1 leaderboard (Decision Index 0.3) · 6 days · Adoption: Little used (⬇ 602/mo HF · 1 ranking) (leaderboards: Decision Index 0.3; released: 2026-10-03; data from 2026-10-09) Score: Best public index, but non-commercial license. What it does: Typed decision on Qwen3.8-27B (merged LoRA); 256k context. Memory: disk 54.7 GB (BF16) · Min. VRAM 20.0 GB approx. · fits on: 1× 24 GB (approx.) Official BF16 54.7 GB GGUF Q8_0 (community) 28.6 GB GGUF Q4_K_M (community) 16.6 GB Quality: Decision Index 0.3: 59.91 (Full, n.º 4) · public index 65.10 (#1) · the index guided the choice of checkpoint (declared) · Latency: Median 99.9 ms (Decision Index 0.3) · card: 33 ms per yes/no question on 1× H100. · Price: Self-host (your own GPU) Big public jump (65.1) that shrinks on the private tests (Full 59.9). There is also a 12B version. License: CC BY-NC 4.0 (non-commercial) Source: huggingface.co/spaces/multimodalart/jev-decision-index · 2026-10-08
Kev 27B Jared Palmer	Typed decision	58.8	self-host	OpenOpen	NewNew ⬇ 1.5k/mo HF · 1 rankingLittle usedLittle used	135 ms	27.8B	65.5 GB official ▣ 1× 80 GB disk 51.2 GB 1× 80 GB (H100/H200/B200, official)	Apache 2.0
Tipo: Typed decision Weights: OpenOpen 🤗 Hugging Face · ↗ Web Maturity: NewNew Released 15 days ago · 2k HF downloads/mo · 1 leaderboard (Decision Index 0.3) · Adoption: Little used (⬇ 1.5k/mo HF · 1 ranking) (leaderboards: Decision Index 0.3; released: 2026-09-24; data from 2026-10-09) Score: The best calibrated in the top; data with a commercial license. What it does: Typed decision (full fine-tune of Qwen3.8-27B + head); states of up to 64k tokens. Memory: disk 51.2 GB (BF16) · Min. VRAM 65.5 GB official · fits on: 1× 80 GB (H100/H200/B200, official) Official BF16 51.2 GB W4A16 (community) 15.2 GB Requirement: Model card: 51 GB weights, server ~65.5 GB resident; 64k states reach 87.1 GB (H200). source Quality: Decision Index 0.3: 58.78 (Full, n.º 6) · public 56.69 · best score on private new domains among open models (55.5) · ECE 0.022 · Latency: Median 135 ms (Decision Index 0.3). · Price: Self-host (your own GPU) Card: data from sources with licenses that allow commercial use, and contamination screening. English only. License: Apache 2.0 Source: huggingface.co/spaces/multimodalart/jev-decision-index · 2026-10-08
Blink v0.3 26B-A4B Pixilab.ai	Typed decision	57.8	self-host	OpenOpen	Barely testedBarely t. ⬇ 765/mo HF · 1 rankingLittle usedLittle used	25 ms	26B / 4B	23.0 GB approx. ▣ 1 consumer GPU ✓ Fits on your computer disk 18.8 GB 1× 32 GB (approx.)	Apache 2.0
Tipo: Typed decision Featured: ★ Jev level at 25 ms Weights: OpenOpen 🤗 Hugging Face · ↗ Web Maturity: Barely testedBarely t. Very little use · 765 HF downloads/mo · 1 leaderboard (Decision Index 0.3) · 7 days · Adoption: Little used (⬇ 765/mo HF · 1 ranking) (leaderboards: Decision Index 0.3; released: 2026-10-02; data from 2026-10-09) Score: Jev level at 25 ms on 1 GPU. What it does: Typed decision on Gemma 4 26B-A4B (MoE, 4B active), published already quantized (NVFP4 / FP8). Memory: disk 18.8 GB (NVFP4) · Min. VRAM 23.0 GB approx. · fits on: 1× 32 GB (approx.) Official NVFP4 (Blackwell) 18.8 GB Official FP8 (Hopper/Ada) 27.2 GB Requirement: Model card: «17.5 GB on disk, served on a single RTX PRO 5000 with 32k context». source Quality: Decision Index 0.3: 57.76 (Full, n.º 8) · public 58.21 · ECE 0.040 · Latency: Median 25.1 ms (Decision Index 0.3): the fastest in the top 10. · Price: Self-host (your own GPU) NVFP4 for Blackwell GPUs; FP8 build for Hopper/Ada. Card: 32k context on an RTX PRO 5000. License: Apache 2.0 Source: huggingface.co/spaces/multimodalart/jev-decision-index · 2026-10-08
Rune 26B-A4B v3 Surogate	Typed decision	57.4	self-host	OpenOpen	NewNew ⬇ 5.2k/mo HF · 1 rankingModerately usedModerately used	121 ms	26B total / 4B active	20.0 GB approx. ▣ 1 consumer GPU ✓ Fits on your computer disk 51.6 GB 1× 24 GB (approx.)	Apache 2.0 (repo gated)
Tipo: Typed decision Weights: OpenOpen ↗ Web Maturity: NewNew Released 18 days ago · 5k HF downloads/mo · 1 leaderboard (Decision Index 0.3) · Adoption: Moderately used (⬇ 5.2k/mo HF · 1 ranking) (leaderboards: Decision Index 0.3; released: 2026-09-21; data from 2026-10-09) Score: MoE 4B active; community GGUF. What it does: Open typed-decision model (full fine-tune of Gemma 4 26B-A4B). Memory: disk 51.6 GB (n/d) · Min. VRAM 20.0 GB approx. · fits on: 1× 24 GB (approx.) BF16 (gated) 51.6 GB GGUF Q8_0 (community) 26.9 GB GGUF Q4_K_M (community) 16.8 GB Quality: Decision Index 0.3: 57.43 (Full score, #10 of 115) · public index 58.28 · Latency: Median 120.5 ms (Decision Index 0.3, 1× RTX PRO 6000, BF16, sequential client). · Price: Self-host (your own GPU) Almost Jev's level, open; in BF16 it needs 1× 80 GB, quantized 1× 24–32 GB (approx.). License: Apache 2.0 (repo gated) Source: huggingface.co/spaces/multimodalart/jev-decision-index · 2026-10-08
Bespoke Nimble 9B v3 Bespoke Labs	Typed decision	54.7	self-host	OpenOpen	Barely testedBarely t. ⬇ 189/mo HF · 1 rankingLittle usedLittle used	74 ms	9.65B	9.0 GB approx. ▣ Edge · ≤10 GB ✓ Fits on your computer disk 17.9 GB 1× 8–12 GB (approx.)	CC BY-NC 4.0 (non-commercial)
Tipo: Typed decision Weights: OpenOpen 🤗 Hugging Face · ↗ Web Maturity: Barely testedBarely t. Very little use · 189 HF downloads/mo · 1 leaderboard (Decision Index 0.3) · 9 days · Adoption: Little used (⬇ 189/mo HF · 1 ranking) (leaderboards: Decision Index 0.3; released: 2026-09-30; data from 2026-10-09) Score: Best ≤10B; non-commercial license. What it does: Typed-decision LoRA adapter for Qwen3.5-9B (choice or yes/no with a probability per option). Memory: disk 17.9 GB (LoRA 0.69 GB + base) · Min. VRAM 9.0 GB approx. · fits on: 1× 8–12 GB (approx.) GGUF BF16 fusionado (ggml-org) 17.9 GB GGUF Q8_0 (ggml-org) 9.5 GB GGUF Q4_K_M (ggml-org) 6.3 GB Official LoRA (without base) 0.7 GB Disk = merged BF16 GGUF; the official adapter is 0.69 GB and needs the Qwen3.5-9B base. Quality: Decision Index 0.3: 54.67 (Full, #18; best ≤10B) · public 57.19 · Latency: Median 73.9 ms (Decision Index 0.3). · Price: Self-host (your own GPU) The official repo is only the LoRA (0.69 GB) on Qwen3.5-9B (Apache 2.0); ggml-org publishes merged GGUFs. License: CC BY-NC 4.0 (non-commercial) Source: huggingface.co/spaces/multimodalart/jev-decision-index · 2026-10-08
Drex v1.5 Nace.AI	Typed decision	50.8	€0.036$0.040/1M	OpenOpen	NewNew HF ~98 downloads/mo · OpenRouter · 1 rankingLittle usedLittle used	54 ms	8.95B	24.0 GB official ▣ 1 consumer GPU ✓ Fits on your computer disk 17.9 GB 1× GPU 24 GB (A10G bf16, test Nace) · Q8_0 ~9.5 GB on Metal/CPU	Nace.AI Open RAIL-M (restricted use; not Apache)
Tipo: Typed decision Featured: ★ Open ≤10B (public) Weights: OpenOpen 🤗 Hugging Face · ↗ Web Maturity: NewNew By default: released ~1 day ago · in Decision Index 0.3.1 · open weights + API · Adoption: Little used (HF ~98 downloads/mo · OpenRouter · 1 ranking) (leaderboards: Decision Index 0.3.1; released: 2026-10-09; data from 2026-10-10) Score: Open ≤10B with the best public index; Full drops to 50.8. RAIL-M license. What it does: Typed decisions (choice / noul yes-no / score) over a text/JSON state; one pass per question, without generating text. Compatible with Jev's System One (/v1/systemone). Memory: disk 17.9 GB (BF16) · Min. VRAM 24.0 GB official · fits on: 1× GPU 24 GB (A10G bf16, test Nace) · Q8_0 ~9.5 GB on Metal/CPU Official BF16 17.9 GB Q8_0 GGUF (convertir; card) 9.5 GB Requirement: Model card: bf16 ~18 GB on a CUDA GPU (tested on A10G 24 GB); Q8_0 GGUF ~9.5 GB on Apple silicon or CPU. source Q8_0 isn't available as a ready GGUF repo: it's converted with Nace's llama.cpp fork (drex-v1.5 branch). Q8 GB according to the model card. Quality: Decision Index 0.3.1: Full 50.84 (#24 of 116) · public index 58.08 (best open ≤10B on public; 0.9-pt tie with Jev/Nimble on public). JevBench 86.2% (Jev 87.0%). · Latency: Median 53.5 ms (Decision Index 0.3.1, official board). OpenRouter P50 ~0.18 s (DeepInfra). Model card: ~0.65 s (8k–32k) · ~2.0 s (32k–128k) on long docs. · Price: Self-host · OpenRouter API $0.04/1M input (output free) Weights on HF (2026-10-09 on OpenRouter). Base XiaomiMiMo/MiMo-V2.6-Distill-Qwen-9B. 16k context by default, up to 131k. Own API drex.nace.ai and OpenRouter nace-ai/drex-v1.5 at $0.04/1M in. In DI Full it falls far behind Decider/GLiDE/Jev on private tests (weak on knowledge: e.g. GPQA). Its own forks of llama.cpp/Ollama. License: Nace.AI Open RAIL-M (restricted use; not Apache) Source: huggingface.co/spaces/multimodalart/jev-decision-index · 2026-10-10
ezjev 4B s2 everettjf	Typed decision	47.0	self-host	OpenOpen	Barely testedBarely t. ⬇ 83/mo HF · 1 rankingLittle usedLittle used	36 ms	4.66B	5.0 GB approx. ▣ Edge · ≤10 GB ✓ Fits on your computer disk 9.1 GB 1× 8–12 GB (approx.)	Other (MIT code; data with NC sources)
Tipo: Typed decision Weights: OpenOpen 🤗 Hugging Face · ↗ Web Maturity: Barely testedBarely t. Very little use · 83 HF downloads/mo · 1 leaderboard (Decision Index 0.3) · 6 days · Adoption: Little used (⬇ 83/mo HF · 1 ranking) (leaderboards: Decision Index 0.3; released: 2026-10-03; data from 2026-10-09) Score: Best ~4B; check licenses before commercial use. What it does: Typed decision on Qwen3.5-4B (merged LoRA), served with vLLM. Memory: disk 9.1 GB (BF16) · Min. VRAM 5.0 GB approx. · fits on: 1× 8–12 GB (approx.) Official BF16 9.1 GB GGUF Q8_0 (community) 4.5 GB GGUF Q4_K_M (community) 2.7 GB Quality: Decision Index 0.3: 46.95 (Full, #35; best ~4B) · public 50.82 · trained with train/dev splits of ~17 benchmarks from the index (declared) · Latency: Median 35.5 ms (Decision Index 0.3, 1× RTX PRO 6000). · Price: Self-host (your own GPU) The author now recommends ezjev-4b-s3 (no score on the board). Some training sources are non-commercial (e.g. ANLI). License: Other (MIT code; data with NC sources) Source: huggingface.co/spaces/multimodalart/jev-decision-index · 2026-10-08
Decider 4B Mapika	Typed decision	40.3	self-host	OpenOpen	NewNew ⬇ 25k/mo HF · 1 rankingModerately usedModerately used	13 ms	4.2B	11.0 GB approx. ▣ 1 consumer GPU ✓ Fits on your computer disk 8.4 GB 1× 16 GB (approx.)	Apache 2.0
Tipo: Typed decision Weights: OpenOpen ↗ Web Maturity: NewNew Released 17 days ago · 25k HF downloads/mo · 1 leaderboard (Decision Index 0.3) · Adoption: Moderately used (⬇ 25k/mo HF · 1 ranking) (leaderboards: Decision Index 0.3; released: 2026-09-22; data from 2026-10-09) Score: 13 ms; open recipe (Mapika). What it does: Typed decisions with calibrated probabilities in one pass (decider family 2B/4B/35B-A3B). Memory: disk 8.4 GB (n/d) · Min. VRAM 11.0 GB approx. · fits on: 1× 16 GB (approx.) BF16 8.4 GB Quality: Decision Index 0.3: 40.33 (Full score, #50 of 115) · public index 41.19 · Latency: Median 12.6 ms (Decision Index 0.3, FP8 + CUDA graphs, 1× RTX PRO 6000). · Price: Self-host (your own GPU) Decider family 2B / 4B / 35B-A3B (35B-A3B: Full 48.5). Permissive license; recipe at github.com/Mapika/decider. License: Apache 2.0 Source: huggingface.co/spaces/multimodalart/jev-decision-index · 2026-10-08
Liquid AI d1-3B Liquid AI	Typed decision	40.1	self-host	OpenOpen	NewNew ⬇ 5.4k/mo HF · 1 rankingModerately usedModerately used	16 ms	3.12B	4.0 GB approx. ▣ Edge · ≤10 GB ✓ Fits on your computer disk 6.2 GB Any 8 GB GPU, Jetson or Mac	LFM Open License 1.0
Tipo: Typed decision Featured: ★ Best edge Weights: OpenOpen ↗ Web Maturity: NewNew Released 4 days ago · 5k HF downloads/mo · 1 leaderboard (Decision Index 0.3) · Adoption: Moderately used (⬇ 5.4k/mo HF · 1 ranking) (leaderboards: Decision Index 0.3; released: 2026-10-05; data from 2026-10-09) Score: Edge: 6 GB, 16 ms; text + image. What it does: Decision model: a single pass, zero output tokens; choice/score/yes-no with confidence and probabilities. Text + image. Memory: disk 6.2 GB (BF16) · Min. VRAM 4.0 GB approx. · fits on: Any 8 GB GPU, Jetson or Mac Official BF16 6.2 GB GGUF Q8_0 2.9 GB GGUF Q4_K_M 1.7 GB Requirement: Model card: 8 ms per decision on an RTX 4090 (bf16); also Jetson Orin Nano (50 ms) and Apple M5 Pro (30 ms). source Quality: Decision Index 0.3: 40.08 (Full score, #51 of 115) · public index 48.99 · Latency: Median 16.1 ms (Decision Index 0.3; Liquid AI submission on AMD MI325X) · model card: 8 ms on an RTX 4090, 30 ms Apple M5 Pro, 50 ms Jetson Orin Nano. · Price: Self-host (your own GPU) It scores 49.0 on the public index, but drops on the private tests (Full 40.1). 32k context. License: LFM Open License 1.0 Source: huggingface.co/spaces/multimodalart/jev-decision-index · 2026-10-08
Liquid AI d1-omni-600M Liquid AI	Typed decision	9.4	self-host	OpenOpen	NewNew ⬇ 8.5k/mo HF · 1 rankingModerately usedModerately used	11 ms	587M	2.0 GB approx. ▣ Edge · ≤10 GB ✓ Fits on your computer disk 2.4 GB 1× 8–12 GB (approx.)	LFM Open License 1.0
Tipo: Typed decision Weights: OpenOpen ↗ Web Maturity: NewNew Released 4 days ago · 8k HF downloads/mo · 1 leaderboard (Decision Index 0.3) · Adoption: Moderately used (⬇ 8.5k/mo HF · 1 ranking) (leaderboards: Decision Index 0.3; released: 2026-10-05; data from 2026-10-09) Score: Text, image and audio; for edge. What it does: Mini version of d1: decisions on text, image and audio with zero output tokens. Memory: disk 2.4 GB (n/d) · Min. VRAM 2.0 GB approx. · fits on: 1× 8–12 GB (approx.) F32 2.4 GB GGUF Q8_0 0.4 GB Quality: Decision Index 0.3: 9.45 (Full score, #97 of 115) · public index 17.90 · Latency: Median 10.9 ms (Decision Index 0.3, Liquid AI submission). · Price: Self-host (CPU/edge) For cheap mass filtering or on-device; 16k context. License: LFM Open License 1.0 Source: huggingface.co/spaces/multimodalart/jev-decision-index · 2026-10-08
Laya Convai Innovations	Typed decision	4.4	self-host	OpenOpen	NewNew ⬇ 36k/mo HF · 1 rankingWidely usedWidely used	5.8 ms	421M	2.0 GB approx. ▣ Edge · ≤10 GB ✓ Fits on your computer disk 0.8 GB Any GPU (CPU possible, slower)	Apache 2.0
Tipo: Typed decision Weights: OpenOpen 🤗 Hugging Face · ↗ Web Maturity: NewNew Released 21 days ago · 36k HF downloads/mo · 1 leaderboard (Decision Index 0.3) · Adoption: Widely used (⬇ 36k/mo HF · 1 ranking) (leaderboards: Decision Index 0.3; released: 2026-09-18; data from 2026-10-09) Score: Super fast, but ~chance without fine-tuning: only useful fine-tuned with your labels. What it does: Non-autoregressive encoder (ModernBERT-large + decision head, trained with RLCD): choice/score/yes-no in one pass. Family: laya (English, 421M), laya-multilingual (mmBERT-base, 322M) and laya-typed-decisions (fine-tuned). Memory: disk 0.8 GB (F16) · Min. VRAM 2.0 GB approx. · fits on: Any GPU (CPU possible, slower) Official F16 (English) 0.8 GB Official multilingual F16 0.6 GB GGUF Q8_0 (ggml-org) 0.5 GB Quality: Decision Index 0.3: 4.43 (Full, #106 of 115) · public 6.19 · 0.2.1: 6.04 · card: 0.36 zero-shot accuracy (chance 0.32); 0.766 only the checkpoint fine-tuned on that same task · Latency: Median 5.8 ms (Decision Index 0.3, CUDA graphs, 1× RTX PRO 6000) · the model card states 33 ms per question on a T4 GPU. · Price: Self-host · free on Vercel AI Gateway until Oct 31 (promo, @vercel_dev Oct 1) The card's «better than Jev» comparisons use other samples and a checkpoint fine-tuned on that task; Jev was not measured on the same suite. Limited options (≲20). Servable with llama.cpp (/v1/systemone, ggml-org/Laya-GGUF). Hosted: free on Vercel AI Gateway until Oct 31 (https://x.com/vercel_dev/status/2105758650052849859); also on Dell Enterprise Hub. License: Apache 2.0 Source: huggingface.co/spaces/multimodalart/jev-decision-index · 2026-10-08
Qwen3-Embedding-8B / 0.6B Alibaba	Embeddings	MTEB-M 70.6 / 64.3	self-host	OpenOpen	ProvenProven ⬇ 11.2M/mo HF · 1 rankingWidely usedWidely used	n/d	8B / 0.6B	3.0 GB approx. ▣ Edge · ≤10 GB ✓ Fits on your computer disk 15.1 GB 0.6B: CPU or any GPU · 8B: 1× 24 GB (approx.)	Apache 2.0
Tipo: Embeddings · retrieve Featured: ★ Retrieve Weights: OpenOpen ↗ Web Maturity: ProvenProven Massive use · 11.2 M HF downloads/mo · 1 leaderboard (MTEB) · from 2025-06 · Adoption: Widely used (⬇ 11.2M/mo HF · 1 ranking) (leaderboards: MTEB; released: 2025-06-03; data from 2026-10-09) Score: Multilingual, 32k, configurable dimension. What it does: Multilingual embeddings (100+ languages) to retrieve the most similar texts by similarity; configurable dimension (MRL) and per-task instructions. Memory: disk 15.1 GB (n/d) · Min. VRAM 3.0 GB approx. · fits on: 0.6B: CPU or any GPU · 8B: 1× 24 GB (approx.) 8B BF16 15.1 GB 0.6B BF16 1.2 GB Quality: MTEB multilingual: 70.58 (8B) / 64.33 (0.6B) · Latency: n/d · Price: Self-host (your own GPU) 32k context (long documents). The 0.6B is the sweet spot for indexing millions of texts. License: Apache 2.0 Source: huggingface.co/Qwen/Qwen3-Embedding-8B · 2026-10-08
EmbeddingGemma-300m Google	Embeddings	Best <500M on MTEB (Google)	self-host	OpenOpen	ProvenProven ⬇ 3.6M/mo HF · 1 rankingWidely usedWidely used	<15 ms (EdgeTPU)	308M	0.2 GB official (RAM, QAT) ▣ Edge · ≤10 GB ✓ Fits on your computer disk 1.2 GB CPU / mobile (<200 MB of RAM quantized, QAT)	Gemma
Tipo: Embeddings · retrieve Weights: OpenOpen ↗ Web Maturity: ProvenProven Massive use · 3.6 M HF downloads/mo · 1 leaderboard (MTEB) · from 2025-07 · Adoption: Widely used (⬇ 3.6M/mo HF · 1 ranking) (leaderboards: MTEB; released: 2025-07-17; data from 2026-10-09) Score: On-device; 2k context (split long documents). What it does: Very light multilingual embeddings (308M) for devices or CPU; 768→128 dimensions (MRL). Memory: disk 1.2 GB (n/d) · Min. VRAM 0.2 GB official (RAM, QAT) · fits on: CPU / mobile (<200 MB of RAM quantized, QAT) F32 1.2 GB Requirement: Google: runs in less than 200 MB of RAM with quantization (QAT). source Quality: Best open multilingual embedding <500M on MTEB (according to Google) · Latency: <15 ms with 256 tokens (EdgeTPU) · Price: Self-host (CPU) 2k context: short for long documents (split them). License: Gemma Source: developers.googleblog.com/en/introducing-embeddinggemma/ · 2026-10-08
BGE-M3 BAAI	Embeddings	MTEB-M 59.6	self-host	OpenOpen	ProvenProven ⬇ 33.2M/mo HF · 1 rankingWidely usedWidely used	n/d	568M	4.0 GB approx. ▣ Edge · ≤10 GB ✓ Fits on your computer disk 2.3 GB 1× 8–12 GB (approx.)	MIT
Tipo: Embeddings · retrieve Weights: OpenOpen ↗ Web Maturity: ProvenProven Massive use · 33.2 M HF downloads/mo · 1 leaderboard (MTEB) · from 2024-01 · Adoption: Widely used (⬇ 33.2M/mo HF · 1 ranking) (leaderboards: MTEB; released: 2024-01-27; data from 2026-10-09) Score: Dense + lexical + ColBERT; 8k. What it does: Hybrid embedding: dense + lexical (sparse) + multi-vector (ColBERT) in one model; useful when exact terms matter (names, codes). Memory: disk 2.3 GB (n/d) · Min. VRAM 4.0 GB approx. · fits on: 1× 8–12 GB (approx.) F32 2.3 GB Quality: Multilingual MTEB: 59.56 (Qwen table, May 2025) · Latency: n/d · Price: Self-host (your own GPU or CPU) 8192 context. Veteran and well tested; lower score than Qwen3 but native hybrid search. License: MIT Source: huggingface.co/BAAI/bge-m3 · 2026-10-08
Voyage voyage-4 / 4-large Voyage AI	Embeddings	n/d	€0.054$0.060/1M	ClosedClosed	ProvenProven 1 rankingNo usage dataNo data	n/d	—	—	—
Tipo: Embeddings · retrieve Weights: ClosedClosed ↗ Web Maturity: ProvenProven Veteran · 1 leaderboard (MTEB) · from 2026-01 · Adoption: No usage data (1 ranking) (leaderboards: MTEB; released: 2026-01-15; data from 2026-10-09) Score: 200M tokens free; lite $0.02/1M. What it does: Embeddings via API with no infrastructure. Memory: Hosted Quality: n/d · Latency: n/d · Price: $0.06/1M (voyage-4) · $0.12/1M (4-large); 200M tokens free voyage-4-lite $0.02/1M. Batch API −33 %. Source: docs.voyageai.com/docs/pricing · 2026-10-08
Qwen3-Reranker-4B / 0.6B Alibaba	Rerankers	MMTEB-R 72.7 / 66.4	self-host	OpenOpen	ProvenProven ⬇ 3.9M/mo HF · 0 rankingsWidely usedWidely used	n/d	4B / 0.6B	3.0 GB approx. ▣ Edge · ≤10 GB ✓ Fits on your computer disk 8.0 GB 0.6B: any GPU · 4B: 1× 12–16 GB (approx.)	Apache 2.0
Tipo: Rerankers · rerank Featured: ★ Rerank Weights: OpenOpen ↗ Web Maturity: ProvenProven Massive use · 3.9 M HF downloads/mo · since 2025-05 · no independent leaderboard in the sources consulted · Adoption: Widely used (⬇ 3.9M/mo HF · 0 rankings) (released: 2025-05-29; data from 2026-10-09) Score: Cross-encoder with instructions; 32k. What it does: Cross-encoder: scores each (query, document) pair by reading both at once; reranks the top-k returned by the embedding. Memory: disk 8.0 GB (n/d) · Min. VRAM 3.0 GB approx. · fits on: 0.6B: any GPU · 4B: 1× 12–16 GB (approx.) 4B BF16 8.0 GB 0.6B BF16 1.2 GB Quality: MMTEB-R: 72.74 (4B) / 66.36 (0.6B); BGE-reranker-v2-m3: 58.36 · Latency: n/d · Price: Self-host (your own GPU) 32k context; accepts an instruction («does this text answer the question?»). License: Apache 2.0 Source: huggingface.co/Qwen/Qwen3-Reranker-0.6B · 2026-10-08
bge-reranker-v2-m3 BAAI	Rerankers	MMTEB-R 58.4	self-host	OpenOpen	ProvenProven ⬇ 16.6M/mo HF · 0 rankingsWidely usedWidely used	n/d	568M	4.0 GB approx. ▣ Edge · ≤10 GB ✓ Fits on your computer disk 2.3 GB 1× 8–12 GB (approx.)	Apache 2.0
Tipo: Rerankers · rerank Weights: OpenOpen ↗ Web Maturity: ProvenProven Massive use · 16.6 M HF downloads/mo · since 2024-03 · no independent leaderboard in the sources consulted · Adoption: Widely used (⬇ 16.6M/mo HF · 0 rankings) (released: 2024-03-15; data from 2026-10-09) Score: Partner of BGE-M3. jina-reranker-v3 is non-commercial. What it does: Lightweight multilingual cross-encoder; natural partner of BGE-M3. Memory: disk 2.3 GB (n/d) · Min. VRAM 4.0 GB approx. · fits on: 1× 8–12 GB (approx.) F32 2.3 GB Quality: MMTEB-R: 58.36 (Qwen table) · Latency: n/d · Price: Self-host (your own GPU or CPU) Alternative: jina-reranker-v3 (0.6B) is CC-BY-NC, not suitable for commercial use. License: Apache 2.0 Source: huggingface.co/BAAI/bge-reranker-v2-m3 · 2026-10-08
Voyage rerank-3 / rerank-3-lite Voyage AI	Rerankers	n/d	€0.045$0.050/1M	ClosedClosed	Barely testedBarely t. 0 rankingsNo usage dataNo data	n/d	—	—	—
Tipo: Rerankers · rerank Weights: ClosedClosed ↗ Web Maturity: Barely testedBarely t. No independent benchmark · 38 days · Adoption: No usage data (0 rankings) (released: 2026-09-01; data from 2026-10-09) Score: ~$0.0025 per query of 100 docs. What it does: Reranker via API. Memory: Hosted Quality: n/d · Latency: n/d · Price: $0.05/1M tokens (rerank-3) · $0.02/1M (lite); ~$0.0025 per request of 100 docs Billed tokens = query tokens × no. of docs + doc tokens. 200M free. Source: docs.voyageai.com/docs/pricing · 2026-10-08
OpenAI Decisions API (GPT-6 Luna) OpenAI	Typed decision	n/a (not in the DI)	€0.089$0.10/1M	ClosedClosed	Barely testedBarely t. 0 rankingsNo usage dataNo data	n/d	—	—	—
Tipo: Typed decision Weights: ClosedClosed ↗ Web Maturity: Barely testedBarely t. No independent benchmark · 3 days · Adoption: No usage data (0 rankings) (released: 2026-10-06; data from 2026-10-09) Score: Direct Jev rival (public beta Oct 6); accepts images. What it does: /v1/decisions endpoint: predicate (yes/no), choice or score questions over text or images, with a structured answer; uses GPT-6 Luna. Memory: Hosted Quality: Not in Decision Index 0.3. · Latency: OpenAI: up to 10× faster than GPT-6 Luna via the Responses API. No independent measurement. · Price: $0.10/1M input tokens; output and cache free Jev costs $0.042/1M input and only accepts text. OpenAI acknowledges that Jev «inspired» the product. Source: fortune.com/2026/10/08/jev-an-ai-for-making-quick-decisions-has-been-a · 2026-10-09
Microsoft-Decision-1 Microsoft	Typed decision	n/a (not in DI yet)	€0.037$0.042/1M	ClosedClosed	NewNew OpenRouter (new) · 0 rankings DINo usage dataNo data	~0.30 s	—	—	—
Tipo: Typed decision Featured: ★ New MS API Weights: ClosedClosed ↗ Web Maturity: NewNew By default: released 1 day ago · no Decision Index yet · API on OpenRouter · Adoption: No usage data (OpenRouter (new) · 0 rankings DI) Score: Same price as Jev ($0.042/1M in); weights not public. What it does: Typed decisions (yes/no, choice, score/rubric) with calibrated probabilities; post-trained on Qwen3.5-9B. Decisions API (Foundry + OpenRouter), does not generate text. Memory: Hosted Quality: Not in Decision Index 0.3.1 (snapshot 2026-10-09). Microsoft: best accuracy on its pack of 36 blind benchmarks (~150k questions) versus H2O-Lightning-4B, Rune, Quyet, GPT-6 Luna Decisions, etc. · Latency: OpenRouter P50 ~0.30 s (Azure). Microsoft: P50 ~35× faster than GPT-6 Sol on its internal bench. · Price: $0.042/1M input; output free (same as Jev) Launched 2026-10-09 on Foundry and OpenRouter. Closed weights (open base Qwen3.5-9B). Xbox Research: competitive with GPT-6 Sol at >14× the speed and ~200× lower cost in labelling. Ideal for matching/routing/classification. Source: commandline.microsoft.com/microsoft-decision-1-model-foundry/ · 2026-10-10
JEV-27B-VL (AutoTrust) AutoTrust AI	Typed decision	DI vision 69.8 (#1)	self-host	OpenOpen	NewNew ⬇ 1.5M/mo HF · 1 rankingWidely usedWidely used	264 ms (vision)	27.8B	2.0 GB approx. ▣ Edge · ≤10 GB ✓ Fits on your computer disk 55.6 GB 1× 8–12 GB (approx.)	Apache 2.0 (adapters); base under its own license
Tipo: Typed decision Weights: OpenOpen 🤗 Hugging Face · ↗ Web Maturity: NewNew Released 9 days ago · 1.5 M HF downloads/mo · 1 leaderboard (Decision Index (vision)) · Adoption: Widely used (⬇ 1.5M/mo HF · 1 ranking) (leaderboards: Decision Index (vision); released: 2026-09-30; data from 2026-10-09) Score: The best in the Decision Index with images. What it does: Typed decision over images + text (LoRA adapter on a 27B base); /v1/decide endpoint in vLLM. Memory: disk 55.6 GB (BF16) · Min. VRAM 2.0 GB approx. · fits on: 1× 8–12 GB (approx.) Official BF16 55.6 GB Adaptador vLLM 0.5 GB AutoTrust says there are quantized variants; not verified here. Quality: Decision Index vision board: 69.82 Full (#1), 72.78 public; 2nd Solomon v1.1 with 67.66. Its sibling GEV-26B-Decide gets 56.17 on the general 0.3 index. · Latency: Median 263.9 ms per row on the Decision Index vision board. · Price: Self-host (your own GPU) License: Apache 2.0 (adapters); base under its own license Source: huggingface.co/spaces/multimodalart/jev-decision-index · 2026-10-09
EmbeddingGemma 2 Google	Embeddings	MTEB Code 78.7	self-host	OpenOpen	Barely testedBarely t. ⬇ 21k/mo HF · 0 rankingsWidely usedWidely used	~191–567 MB RAM (QAT)	740M	0.6 GB official (RAM, quantized) ▣ Edge · ≤10 GB ✓ Fits on your computer disk 1.5 GB CPU / mobile (~191 MB of RAM text only, ~567 MB multimodal, quantized)	Apache 2.0
Tipo: Embeddings · retrieve Weights: OpenOpen 🤗 Hugging Face · ↗ Web Maturity: Barely testedBarely t. No independent benchmark · 21k HF downloads/mo · 25 days · Adoption: Widely used (⬇ 21k/mo HF · 0 rankings) (released: 2026-09-14; data from 2026-10-09) Score: Successor to the 300m: Apache 2.0, 8k context (4×). What it does: Multimodal embeddings (text, code, image, video and audio) in a single space; 768→128 dimensions (MRL); 8k context. Memory: disk 1.5 GB (BF16) · Min. VRAM 0.6 GB official (RAM, quantized) · fits on: CPU / mobile (~191 MB of RAM text only, ~567 MB multimodal, quantized) Official BF16 1.5 GB GGUF Q8_0 0.3 GB GGUF UD-Q4_K_XL 0.2 GB Requirement: Google: ~191 MB of active RAM text only and ~567 MB multimodal, quantized, on a Pixel 11 Pro. source Quality: Google: <1B multimodal leader on MTEB Code and MAEB; MTEB Code 68.76 → 78.68 versus EmbeddingGemma 1. · Latency: On-device; Google gives no latency. · Price: Self-host (CPU/mobile) License: Apache 2.0 Source: blog.google/innovation-and-ai/technology/developers-tools/embeddinggem · 2026-10-09
pplx-embed-v2-late 0.6B / 9B Perplexity	Embeddings	ViDoRe v3 62.3 / 65.2	self-host	OpenOpen	NewNew ⬇ 1k/mo HF · 1 rankingLittle usedLittle used	n/d	0.6B / 9B	5.0 GB approx. ▣ Edge · ≤10 GB ✓ Fits on your computer disk 2.4 GB 1× 8–12 GB (approx.)	MIT
Tipo: Embeddings · retrieve Weights: OpenOpen 🤗 Hugging Face · ↗ Web Maturity: NewNew Only 1 leaderboard (MTEB) · 1k HF downloads/mo · 81 days · Adoption: Little used (⬇ 1k/mo HF · 1 ranking) (leaderboards: MTEB; released: 2026-07-20; data from 2026-10-09) Score: Multi-vector: more accurate, but the index takes up more space. What it does: Multi-vector retrieval (ColBERT, MaxSim) of text, images and PDFs without OCR; both sizes share the space (index with 9B, query with 0.6B). Memory: disk 2.4 GB (F32) · Min. VRAM 5.0 GB approx. · fits on: 1× 8–12 GB (approx.) 0.6B F32 2.4 GB 9B F32 33.6 GB Weights published in F32; in BF16 they'd be half. Quality: ViDoRe v3 nDCG@10 (image): 62.3 (0.6B) and 65.2 (9B), according to the model cards. · Latency: n/d · Price: Self-host (your own GPU) License: MIT Source: aicoder.com/news/news-20261008-perplexity-pplx-embed-v2-late-multimoda · 2026-10-09

Decision Index 0.3 (official board multimodalart/jev-decision-index, data from Oct 7): «Full score» = 20% public benchmarks + 50% private tests of the same skills + 30% private tasks from new domains, corrected for chance (0 = chance, 100 = perfect). Latency = median per request measured by the board (1× RTX PRO 6000, sequential client; Jev = API over the network). MTEB/MMTEB-R: tables from the Qwen model cards.

Data zone

For the data nerds

Quality vs price chart, full table with filters and all the variants of each model (max, high, medium…). Click ▸ on a row to see sources, memory and notes.

Sources publish API prices in dollars. With the €/$ button at the top you choose the currency; in euros they are converted at €1 = $1.1206 (ECB reference rate of October 9, 2026).

Terminal-Bench 4.0 vs cost per task

Up = better · left = cheaper · the dashed line is the efficient frontier.

38 models · 9 on the frontier · hover over or tap a point to see its name and values

ClosedOpen weightsEfficient frontier

Swipe to see the whole chart

No cost or no data: Gemma 4 31B, Mellum2.1.

Full table

40 models and variants · click ▸ on a row to see sources, memory and notes

Sort

Model	TB 4.0% solved ⇅	Costper task ⇅	WeightsWeights⇅	MaturityMad.⇅	€/1M$/1Min / out	AgentsAA	Paramstotal / act.	Memory ⇅VRAM · disco	License
Claude Sonnet 5.5 (max) Anthropic	63.6	€11.29$12.65	ClosedClosed	NewNew on OpenRouter · 5 rankingsModerately usedModerately used	€1.78$2.00 / €8.92$10.00	68.4 Claude Code	—	—	—
Featured: on the efficient frontier (nothing gives more performance for less money) Weights: ClosedClosed Maturity: NewNew Released 11 days ago · on OpenRouter · 5 leaderboards · Adoption: Moderately used (on OpenRouter · 5 rankings) (leaderboards: AA Intelligence Index, AA Terminal-Bench 4.0, AA Coding Agents, LMArena Text, LMArena WebDev; released: 2026-09-28; data from 2026-10-09) AA Coding Agents Index: 68.4 with “Claude Code - Sonnet 5.5 (max)” at €12.66$14.19/task Source: artificialanalysis.ai/evaluations/terminalbench-4-0 · 2026-10-08
Claude Opus 5.5 (max) Anthropic	59.6	€11.70$13.11	ClosedClosed	ProvenProven OpenRouter top 20 · #2 · 5 rankingsWidely usedWidely used	€3.57$4.00 / €17.85$20.00	66.0 Claude Code	—	—	—
Weights: ClosedClosed Maturity: ProvenProven top 20 OpenRouter · 5 leaderboards · 17 days · Adoption: Widely used (OpenRouter top 20 · #2 · 5 rankings) (leaderboards: AA Intelligence Index, AA Terminal-Bench 4.0, AA Coding Agents, LMArena Text, LMArena WebDev; released: 2026-09-22; data from 2026-10-09) AA Coding Agents Index: 66.0 with “Claude Code - Opus 5.5 (max)” at €11.63$13.04/task Source: artificialanalysis.ai/evaluations/terminalbench-4-0 · 2026-10-08
Claude Opus 5.5 (xhigh) Anthropic	59.6	€7.83$8.78	ClosedClosed	ProvenProven OpenRouter top 20 · #2 · 5 rankingsWidely usedWidely used	€3.57$4.00 / €17.85$20.00	n/d	—	—	—
Weights: ClosedClosed Maturity: ProvenProven top 20 OpenRouter · 5 leaderboards · 17 days · Adoption: Widely used (OpenRouter top 20 · #2 · 5 rankings) (leaderboards: AA Intelligence Index, AA Terminal-Bench 4.0, AA Coding Agents, LMArena Text, LMArena WebDev; released: 2026-09-22; data from 2026-10-09) Source: artificialanalysis.ai/evaluations/terminalbench-4-0 · 2026-10-08
GPT-6 Astra (xhigh) OpenAI	59.6	€5.23$5.86	ClosedClosed	ProvenProven on OpenRouter · 5 rankingsModerately usedModerately used	€8.92$10.00 / €44.62$50.00	n/d	—	—	—
Featured: on the efficient frontier (nothing gives more performance for less money) Weights: ClosedClosed Maturity: ProvenProven on OpenRouter · 5 leaderboards · 36 days · Adoption: Moderately used (on OpenRouter · 5 rankings) (leaderboards: AA Intelligence Index, AA Terminal-Bench 4.0, AA Coding Agents, LMArena Text, LMArena WebDev; released: 2026-09-03; data from 2026-10-09) Source: artificialanalysis.ai/evaluations/terminalbench-4-0 · 2026-10-08
GPT-6 Astra (max) OpenAI	59.1	€7.58$8.50	ClosedClosed	ProvenProven on OpenRouter · 5 rankingsModerately usedModerately used	€8.92$10.00 / €44.62$50.00	61.6 Codex	—	—	—
Weights: ClosedClosed Maturity: ProvenProven on OpenRouter · 5 leaderboards · 36 days · Adoption: Moderately used (on OpenRouter · 5 rankings) (leaderboards: AA Intelligence Index, AA Terminal-Bench 4.0, AA Coding Agents, LMArena Text, LMArena WebDev; released: 2026-09-03; data from 2026-10-09) AA Coding Agents Index: 61.6 with “Codex - GPT-6 Astra (max)” at €6.66$7.47/task Source: artificialanalysis.ai/evaluations/terminalbench-4-0 · 2026-10-08
Claude Sonnet 5.5 (xhigh) Anthropic	57.1	€5.42$6.08	ClosedClosed	NewNew on OpenRouter · 5 rankingsModerately usedModerately used	€1.78$2.00 / €8.92$10.00	62.9 Claude Code	—	—	—
Weights: ClosedClosed Maturity: NewNew Released 11 days ago · on OpenRouter · 5 leaderboards · Adoption: Moderately used (on OpenRouter · 5 rankings) (leaderboards: AA Intelligence Index, AA Terminal-Bench 4.0, AA Coding Agents, LMArena Text, LMArena WebDev; released: 2026-09-28; data from 2026-10-09) AA Coding Agents Index: 62.9 with “Claude Code - Sonnet 5.5 (xhigh)” at €2.97$3.33/task Source: artificialanalysis.ai/evaluations/terminalbench-4-0 · 2026-10-08
Gemini 4 Argon (high) Google	57.1	€6.05$6.78	ClosedClosed	NewNew 5 rankingsNo usage dataNo data	€1.78$2.00 / €8.92$10.00	63.8 Antigravity CLI	—	—	—
Weights: ClosedClosed Maturity: NewNew Released 9 days ago · 5 leaderboards · Adoption: No usage data (5 rankings) (leaderboards: AA Intelligence Index, AA Terminal-Bench 4.0, AA Coding Agents, LMArena Text, LMArena WebDev; released: 2026-09-30; data from 2026-10-09) Score: In the agents index it is listed as unavailable. AA Coding Agents Index: 63.8 with “Antigravity CLI - Gemini 4 Argon (high)” at €5.21$5.84/task (marked as unavailable) Source: artificialanalysis.ai/evaluations/terminalbench-4-0 · 2026-10-08
Claude Opus 5.5 (high) Anthropic	56.6	€4.57$5.12	ClosedClosed	ProvenProven OpenRouter top 20 · #2 · 5 rankingsWidely usedWidely used	€3.57$4.00 / €17.85$20.00	n/d	—	—	—
Featured: on the efficient frontier (nothing gives more performance for less money) Weights: ClosedClosed Maturity: ProvenProven top 20 OpenRouter · 5 leaderboards · 17 days · Adoption: Widely used (OpenRouter top 20 · #2 · 5 rankings) (leaderboards: AA Intelligence Index, AA Terminal-Bench 4.0, AA Coding Agents, LMArena Text, LMArena WebDev; released: 2026-09-22; data from 2026-10-09) Source: artificialanalysis.ai/evaluations/terminalbench-4-0 · 2026-10-08
GPT-6.1 Sol (max) OpenAI	56.1	€1.63$1.82	ClosedClosed	ProvenProven OpenRouter top 20 · #1 · 5 rankingsWidely usedWidely used	€1.78$2.00 / €8.92$10.00	60.1 Codex	—	—	—
Featured: on the efficient frontier (nothing gives more performance for less money) Weights: ClosedClosed Maturity: ProvenProven top 20 OpenRouter · 5 leaderboards · 10 days · Adoption: Widely used (OpenRouter top 20 · #1 · 5 rankings) (leaderboards: AA Intelligence Index, AA Terminal-Bench 4.0, AA Coding Agents, LMArena Text, LMArena WebDev; released: 2026-09-29; data from 2026-10-09) AA Coding Agents Index: 60.1 with “Codex - GPT-6.1 Sol (max)” at €1.39$1.55/task Source: artificialanalysis.ai/evaluations/terminalbench-4-0 · 2026-10-08
Claude Fable 5.1 (xhigh) Anthropic	55.1	€14.08$15.78	ClosedClosed	ProvenProven on OpenRouter · 5 rankingsModerately usedModerately used	€8.92$10.00 / €44.62$50.00	n/d	—	—	—
Weights: ClosedClosed Maturity: ProvenProven on OpenRouter · 5 leaderboards · 38 days · Adoption: Moderately used (on OpenRouter · 5 rankings) (leaderboards: AA Intelligence Index, AA Terminal-Bench 4.0, AA Coding Agents, LMArena Text, LMArena WebDev; released: 2026-09-01; data from 2026-10-09) Source: artificialanalysis.ai/evaluations/terminalbench-4-0 · 2026-10-09
GPT-6.1 Sol (xhigh) OpenAI	54.0	€0.92$1.03	ClosedClosed	ProvenProven OpenRouter top 20 · #1 · 5 rankingsWidely usedWidely used	€1.78$2.00 / €8.92$10.00	62.9 Codex	—	—	—
Featured: on the efficient frontier (nothing gives more performance for less money) Weights: ClosedClosed Maturity: ProvenProven top 20 OpenRouter · 5 leaderboards · 10 days · Adoption: Widely used (OpenRouter top 20 · #1 · 5 rankings) (leaderboards: AA Intelligence Index, AA Terminal-Bench 4.0, AA Coding Agents, LMArena Text, LMArena WebDev; released: 2026-09-29; data from 2026-10-09) AA Coding Agents Index: 62.9 with “Codex - GPT-6.1 Sol (xhigh)” at €0.93$1.04/task Source: artificialanalysis.ai/evaluations/terminalbench-4-0 · 2026-10-08
Claude Fable 5.1 (max) Anthropic	52.0	€17.15$19.22	ClosedClosed	ProvenProven on OpenRouter · 5 rankingsModerately usedModerately used	€8.92$10.00 / €44.62$50.00	62.2 Claude Code	—	—	—
Weights: ClosedClosed Maturity: ProvenProven on OpenRouter · 5 leaderboards · 38 days · Adoption: Moderately used (on OpenRouter · 5 rankings) (leaderboards: AA Intelligence Index, AA Terminal-Bench 4.0, AA Coding Agents, LMArena Text, LMArena WebDev; released: 2026-09-01; data from 2026-10-09) AA Coding Agents Index: 62.2 with “Claude Code - Fable 5.1 (max) (with fallback)” at €11.05$12.39/task Source: artificialanalysis.ai/evaluations/terminalbench-4-0 · 2026-10-09
GPT-6.1 Sol (high) OpenAI	51.5	€0.74$0.83	ClosedClosed	ProvenProven OpenRouter top 20 · #1 · 5 rankingsWidely usedWidely used	€1.78$2.00 / €8.92$10.00	60.1 Codex	—	—	—
Featured: ★ Best value · on the efficient frontier (nothing gives more performance for less money) Weights: ClosedClosed Maturity: ProvenProven top 20 OpenRouter · 5 leaderboards · 10 days · Adoption: Widely used (OpenRouter top 20 · #1 · 5 rankings) (leaderboards: AA Intelligence Index, AA Terminal-Bench 4.0, AA Coding Agents, LMArena Text, LMArena WebDev; released: 2026-09-29; data from 2026-10-09) AA Coding Agents Index: 60.1 with “Codex - GPT-6.1 Sol (high)” at €0.79$0.89/task Source: artificialanalysis.ai/evaluations/terminalbench-4-0 · 2026-10-08
GPT-6.1 Sol (medium) OpenAI	48.0	€0.55$0.61	ClosedClosed	ProvenProven OpenRouter top 20 · #1 · 5 rankingsWidely usedWidely used	€1.78$2.00 / €8.92$10.00	61.4 Codex	—	—	—
Featured: on the efficient frontier (nothing gives more performance for less money) Weights: ClosedClosed Maturity: ProvenProven top 20 OpenRouter · 5 leaderboards · 10 days · Adoption: Widely used (OpenRouter top 20 · #1 · 5 rankings) (leaderboards: AA Intelligence Index, AA Terminal-Bench 4.0, AA Coding Agents, LMArena Text, LMArena WebDev; released: 2026-09-29; data from 2026-10-09) AA Coding Agents Index: 61.4 with “Codex - GPT-6.1 Sol (medium)” at €0.63$0.70/task Source: artificialanalysis.ai/evaluations/terminalbench-4-0 · 2026-10-08
GLM-5.3 (max) Z AI	41.9	€7.20$8.06	OpenOpen	ProvenProven OpenRouter top 20 · #4 · 5 rankingsWidely usedWidely used	€1.25$1.40 / €3.93$4.40	53.6 Opencode	753B / 40B	833 GB approx. ▣ 1 node 8 GPU disk 756 GB 1 node 8× H200 (FP8)	GLM-5.3 (condition >$10,000M MaaS)
Weights: OpenOpen 🤗 Hugging Face Maturity: ProvenProven 1.5 M HF downloads/mo · top 20 OpenRouter · 5 leaderboards · 52 days · Adoption: Widely used (OpenRouter top 20 · #4 · 5 rankings) (leaderboards: AA Intelligence Index, AA Terminal-Bench 4.0, AA Coding Agents, LMArena Text, LMArena WebDev; released: 2026-08-18; data from 2026-10-09) Memory: disk 756 GB (FP8) · Min. VRAM 833 GB approx. · fits on: 1 node 8× H200 (FP8) Official FP8 756 GB Requirement: GenAI Brief: ~750 GB in FP8; fits on an 8× H200 node. source Repo zai-org/GLM-5.3-FP8 returns 401 (not accessible). AA Coding Agents Index: 53.6 with “Opencode - GLM-5.3” at €3.78$4.24/task License: GLM-5.3 (condition >$10,000M MaaS) Source: artificialanalysis.ai/evaluations/terminalbench-4-0 · 2026-10-08
Qwen3.8 Max (0902) Alibaba	38.9	€16.67$18.68	ClosedClosed	ProvenProven on OpenRouter · 4 rankingsModerately usedModerately used	€1.78$2.00 / €5.35$6.00	43.3 Claude Code	—	—	—
Weights: ClosedClosed Maturity: ProvenProven on OpenRouter · 4 leaderboards · 37 days · Adoption: Moderately used (on OpenRouter · 4 rankings) (leaderboards: AA Intelligence Index, AA Terminal-Bench 4.0, AA Coding Agents, LMArena WebDev; released: 2026-09-02; data from 2026-10-09) AA Coding Agents Index: 43.3 with “Claude Code - Qwen3.8 Max” at €3.10$3.48/task Source: artificialanalysis.ai/evaluations/terminalbench-4-0 · 2026-10-09
MiMo-V2.6-Pro Xiaomi	34.8	€0.39$0.44	OpenOpen	NewNew ⬇ 93k/mo HF · 4 rankingsModerately usedModerately used	€0.39$0.43 / €0.78$0.87	n/d	1.0T / 42B	625 GB approx. ▣ 1 node 8 GPU disk 566 GB 1 node 8× Blackwell (2 nodes on Hopper)	MIT
Featured: ★ Best open value · on the efficient frontier (nothing gives more performance for less money) Weights: OpenOpen 🤗 Hugging Face Maturity: NewNew Released 18 days ago · 93k HF downloads/mo · on OpenRouter · 4 leaderboards · Adoption: Moderately used (⬇ 93k/mo HF · 4 rankings) (leaderboards: AA Intelligence Index, AA Terminal-Bench 4.0, LMArena Text, LMArena WebDev; released: 2026-09-21; data from 2026-10-09) Memory: disk 566 GB (FP4/FP8 mixed) · Min. VRAM 625 GB approx. · fits on: 1 node 8× Blackwell (2 nodes on Hopper) Official FP4/FP8 566 GB Requirement: Model card: vLLM with tensor-parallel 8 (SGLang tp 16). GenAI Brief: ~1 TB in FP8; one node of 8 Blackwell GPUs. source License: MIT Source: artificialanalysis.ai/evaluations/terminalbench-4-0 · 2026-10-08
Muse Spark 1.3 (max) Meta	33.3	€6.19$6.94	ClosedClosed	ProvenProven on OpenRouter · 5 rankingsModerately usedModerately used	€1.12$1.25 / €3.79$4.25	54.3 Muse Code	—	—	—
Weights: ClosedClosed Maturity: ProvenProven on OpenRouter · 5 leaderboards · 37 days · Adoption: Moderately used (on OpenRouter · 5 rankings) (leaderboards: AA Intelligence Index, AA Terminal-Bench 4.0, AA Coding Agents, LMArena Text, LMArena WebDev; released: 2026-09-02; data from 2026-10-09) AA Coding Agents Index: 54.3 with “Muse Code - Muse Spark 1.3 (max)” at €3.55$3.98/task Source: artificialanalysis.ai/evaluations/terminalbench-4-0 · 2026-10-08
Step 5 Preview StepFun	33.3	€4.62$5.18	ClosedClosed	NewNew on OpenRouter · 5 rankingsModerately usedModerately used	€0.89$1.00 / €2.41$2.70	51.2 Claude Code	—	—	—
Weights: ClosedClosed Maturity: NewNew Released 21 days ago · on OpenRouter · 5 leaderboards · Adoption: Moderately used (on OpenRouter · 5 rankings) (leaderboards: AA Intelligence Index, AA Terminal-Bench 4.0, AA Coding Agents, LMArena Text, LMArena WebDev; released: 2026-09-18; data from 2026-10-09) AA Coding Agents Index: 51.2 with “Claude Code - Step 5 Preview” at €1.48$1.66/task Source: artificialanalysis.ai/evaluations/terminalbench-4-0 · 2026-10-08
Ling 3.1 Flash InclusionAI	33.3	€4.90$5.49	ClosedClosed	NewNew on OpenRouter · 2 rankingsModerately usedModerately used	€0.27$0.30 / €0.80$0.90	n/d	—	—	—
Weights: ClosedClosed Maturity: NewNew Released 8 days ago · on OpenRouter · 2 leaderboards (AA Intelligence Index, AA Terminal-Bench 4.0) · Adoption: Moderately used (on OpenRouter · 2 rankings) (leaderboards: AA Intelligence Index, AA Terminal-Bench 4.0; released: 2026-10-01; data from 2026-10-09) Source: artificialanalysis.ai/evaluations/terminalbench-4-0 · 2026-10-09
Claude Haiku 5.5 (max) Anthropic	32.8	€0.58$0.65	ClosedClosed	NewNew on OpenRouter · 4 rankingsModerately usedModerately used	€0.089$0.10 / €0.45$0.50	36.5 Claude Code	—	—	—
Weights: ClosedClosed Maturity: NewNew Released 2 days ago · on OpenRouter · 4 leaderboards · Adoption: Moderately used (on OpenRouter · 4 rankings) (leaderboards: AA Intelligence Index, AA Terminal-Bench 4.0, AA Coding Agents, LMArena WebDev; released: 2026-10-07; data from 2026-10-09) AA Coding Agents Index: 36.5 with “Claude Code - Haiku 5.5 (max)” at €2.30$2.58/task Source: artificialanalysis.ai/evaluations/terminalbench-4-0 · 2026-10-08
GLM-5.3-Flash Z AI	32.8	€0.70$0.79	OpenOpen	ProvenProven OpenRouter top 20 · #11 · 4 rankingsWidely usedWidely used	€0.13$0.15 / €0.45$0.50	n/d	320B / 18B	363 GB approx. ▣ 1 node 8 GPU disk 328 GB 1 node 8× H100 (FP8)	MIT
Weights: OpenOpen 🤗 Hugging Face Maturity: ProvenProven 6.4 M HF downloads/mo · top 20 OpenRouter · 4 leaderboards · 44 days · Adoption: Widely used (OpenRouter top 20 · #11 · 4 rankings) (leaderboards: AA Intelligence Index, AA Terminal-Bench 4.0, LMArena Text, LMArena WebDev; released: 2026-08-26; data from 2026-10-09) Memory: disk 328 GB (FP8) · Min. VRAM 363 GB approx. · fits on: 1 node 8× H100 (FP8) Official FP8 328 GB GGUF Q8_0 341 GB GGUF UD-IQ1_S 93.1 GB Requirement: GenAI Brief: Flash fits comfortably on 8× H100. source License: MIT Source: artificialanalysis.ai/evaluations/terminalbench-4-0 · 2026-10-08
Claude Haiku 5.5 (xhigh) Anthropic	29.3	€0.41$0.46	ClosedClosed	NewNew on OpenRouter · 4 rankingsModerately usedModerately used	€0.089$0.10 / €0.45$0.50	41.4 Claude Code	—	—	—
Weights: ClosedClosed Maturity: NewNew Released 2 days ago · on OpenRouter · 4 leaderboards · Adoption: Moderately used (on OpenRouter · 4 rankings) (leaderboards: AA Intelligence Index, AA Terminal-Bench 4.0, AA Coding Agents, LMArena WebDev; released: 2026-10-07; data from 2026-10-09) AA Coding Agents Index: 41.4 with “Claude Code - Haiku 5.5 (xhigh)” at €0.54$0.61/task Source: artificialanalysis.ai/evaluations/terminalbench-4-0 · 2026-10-08
DeepSeek V4.1 Flash (max) DeepSeek	26.8	€1.03$1.16	OpenOpen	ProvenProven OpenRouter top 20 · #3 · 4 rankingsWidely usedWidely used	€0.27$0.30 / €1.07$1.20	n/d	552B / 16B	563 GB approx. ▣ 1 node 8 GPU disk 510 GB 1 node 8 GPU (FP8)	MIT
Weights: OpenOpen 🤗 Hugging Face Maturity: ProvenProven 1.3 M HF downloads/mo · top 20 OpenRouter · 4 leaderboards · 29 days · Adoption: Widely used (OpenRouter top 20 · #3 · 4 rankings) (leaderboards: AA Intelligence Index, AA Terminal-Bench 4.0, LMArena Text, LMArena WebDev; released: 2026-09-10; data from 2026-10-09) Memory: disk 510 GB (FP8/INT8) · Min. VRAM 563 GB approx. · fits on: 1 node 8 GPU (FP8) Official FP8 510 GB GGUF Q2 (antirez) 366 GB Requirement: GenAI Brief: ~550 GB in FP8; one 8-GPU node. source License: MIT Source: artificialanalysis.ai/evaluations/terminalbench-4-0 · 2026-10-08
Mistral Large 4 Preview Mistral	26.8	€5.18$5.81	ClosedClosed	NewNew on OpenRouter · 4 rankingsModerately usedModerately used	€1.21$1.36 / €3.73$4.18	n/d	—	—	—
Weights: ClosedClosed Maturity: NewNew Released 3 days ago · on OpenRouter · 4 leaderboards · Adoption: Moderately used (on OpenRouter · 4 rankings) (leaderboards: AA Intelligence Index, AA Terminal-Bench 4.0, LMArena Text, LMArena WebDev; released: 2026-10-06; data from 2026-10-09) Score: List price (preview at 50%). Source: artificialanalysis.ai/evaluations/terminalbench-4-0 · 2026-10-09
Qwen3.8-Flash-Next Alibaba	25.3	€0.78$0.88	OpenOpen	ProvenProven ⬇ 1.6M/mo HF · 3 rankingsWidely usedWidely used	€0.13$0.15 / €0.42$0.47	n/d	180B / 6B	124 GB approx. ▣ 2–4 GPU disk 360 GB 1× H200 (141 GB) or 2× 80 GB (approx.)	Qwen Community 1.0
Weights: OpenOpen 🤗 Hugging Face Maturity: ProvenProven 1.6 M HF downloads/mo · 3 leaderboards · 44 days · Adoption: Widely used (⬇ 1.6M/mo HF · 3 rankings) (leaderboards: AA Intelligence Index, AA Terminal-Bench 4.0, LMArena WebDev; released: 2026-08-26; data from 2026-10-09) Memory: disk 360 GB (BF16) · Min. VRAM 124 GB approx. · fits on: 1× H200 (141 GB) or 2× 80 GB (approx.) Official BF16 360 GB GGUF UD-Q4_K_XL 111 GB GGUF UD-IQ1_S 72.5 GB 180B total / 6B active: in Q4 it runs on 2× 80 GB or a Mac with 128 GB+ (approx.). License: Qwen Community 1.0 Source: artificialanalysis.ai/evaluations/terminalbench-4-0 · 2026-10-08
Grok 4.7 (high) xAI (SpaceXAI)	24.7	€7.98$8.95	ClosedClosed	NewNew on OpenRouter · 4 rankingsModerately usedModerately used	€1.78$2.00 / €5.35$6.00	n/d	—	—	—
Weights: ClosedClosed Maturity: NewNew Released 18 days ago · on OpenRouter · 4 leaderboards · Adoption: Moderately used (on OpenRouter · 4 rankings) (leaderboards: AA Intelligence Index, AA Terminal-Bench 4.0, LMArena Text, LMArena WebDev; released: 2026-09-21; data from 2026-10-09) Source: artificialanalysis.ai/evaluations/terminalbench-4-0 · 2026-10-08
MiMo-V2.6-Flash Xiaomi	22.7	€0.21$0.23	OpenOpen	ProvenProven OpenRouter top 20 · #8 · 4 rankingsWidely usedWidely used	€0.12$0.14 / €0.25$0.28	n/d	309B / 15B	186 GB approx. ▣ 2–4 GPU disk 173 GB 4× GPU 80 GB (TP 4)	MIT
Featured: ★ Ultra-cheap open · on the efficient frontier (nothing gives more performance for less money) Weights: OpenOpen 🤗 Hugging Face Maturity: ProvenProven 56k HF downloads/mo · top 20 OpenRouter · 4 leaderboards · 18 days · Adoption: Widely used (OpenRouter top 20 · #8 · 4 rankings) (leaderboards: AA Intelligence Index, AA Terminal-Bench 4.0, LMArena Text, LMArena WebDev; released: 2026-09-21; data from 2026-10-09) Memory: disk 173 GB (FP4/FP8 mixed) · Min. VRAM 186 GB approx. · fits on: 4× GPU 80 GB (TP 4) Official FP4/FP8 173 GB GGUF MXFP4 167 GB Requirement: Model card: vLLM example with tensor-parallel 4. source License: MIT Source: artificialanalysis.ai/evaluations/terminalbench-4-0 · 2026-10-08
Gemini 3.8 Flash (high) Google	19.7	€5.21$5.84	ClosedClosed	ProvenProven OpenRouter top 20 · #16 · 5 rankingsWidely usedWidely used	€0.67$0.75 / €3.35$3.75	41.9 Antigravity SDK	—	—	—
Weights: ClosedClosed Maturity: ProvenProven top 20 OpenRouter · 5 leaderboards · 37 days · Adoption: Widely used (OpenRouter top 20 · #16 · 5 rankings) (leaderboards: AA Intelligence Index, AA Terminal-Bench 4.0, AA Coding Agents, LMArena Text, LMArena WebDev; released: 2026-09-02; data from 2026-10-09) AA Coding Agents Index: 41.9 with “Antigravity SDK - Gemini 3.8 Flash (high)” at €2.20$2.47/task Source: artificialanalysis.ai/evaluations/terminalbench-4-0 · 2026-10-08
GPT-6 Luna (max) OpenAI	12.6	€0.21$0.24	ClosedClosed	ProvenProven OpenRouter top 20 · #7 · 5 rankingsWidely usedWidely used	€0.089$0.10 / €0.45$0.50	41.1 Codex	—	—	—
Weights: ClosedClosed Maturity: ProvenProven top 20 OpenRouter · 5 leaderboards · 17 days · Adoption: Widely used (OpenRouter top 20 · #7 · 5 rankings) (leaderboards: AA Intelligence Index, AA Terminal-Bench 4.0, AA Coding Agents, LMArena Text, LMArena WebDev; released: 2026-09-22; data from 2026-10-09) AA Coding Agents Index: 41.1 with “Codex - GPT-6 Luna (max)” at €0.16$0.18/task Source: artificialanalysis.ai/evaluations/terminalbench-4-0 · 2026-10-08
Kimi K3 (max) Moonshot AI	12.6	€4.54$5.09	OpenOpen	ProvenProven OpenRouter top 20 · #6 · 5 rankingsWidely usedWidely used	€2.68$3.00 / €13.39$15.00	51.9 Kimi Code CLI	2.8T / 104B	1719 GB approx. ▣ Multi-node disk 1561 GB Multi-node on Hopper; high-memory Blackwell node	Kimi K3 (attribution and MaaS agreement >$20M)
Weights: OpenOpen 🤗 Hugging Face Maturity: ProvenProven 1.2 M HF downloads/mo · top 20 OpenRouter · 5 leaderboards · 85 days · Adoption: Widely used (OpenRouter top 20 · #6 · 5 rankings) (leaderboards: AA Intelligence Index, AA Terminal-Bench 4.0, AA Coding Agents, LMArena Text, LMArena WebDev; released: 2026-07-16; data from 2026-10-09) Memory: disk 1561 GB (MXFP4/U8) · Min. VRAM 1719 GB approx. · fits on: Multi-node on Hopper; high-memory Blackwell node Official MXFP4 1561 GB GGUF UD-IQ1_S (community) 594 GB Requirement: GenAI Brief: ~1.4 TB in MXFP4; multi-node on Hopper. source AA Coding Agents Index: 51.9 with “Kimi Code CLI - Kimi K3” at €4.50$5.05/task License: Kimi K3 (attribution and MaaS agreement >$20M) Source: artificialanalysis.ai/evaluations/terminalbench-4-0 · 2026-10-08
Qwen3.8 2.4T A95B Alibaba	11.1	€10.25$11.49	OpenOpen	ProvenProven ⬇ 33k/mo HF · 2 rankingsWidely usedWidely used	€1.78$2.00 / €5.35$6.00	n/d	2.4T / 95B	5384 GB approx. ▣ Multi-node disk 4892 GB Multi-node (does not fit on a Hopper node)	Qwen3.8-Max (commercial terms)
Weights: OpenOpen 🤗 Hugging Face Maturity: ProvenProven 33k HF downloads/mo · on OpenRouter · 2 leaderboards (AA Intelligence Index, AA Terminal-Bench 4.0) · 58 days · Adoption: Widely used (⬇ 33k/mo HF · 2 rankings) (leaderboards: AA Intelligence Index, AA Terminal-Bench 4.0; released: 2026-08-12; data from 2026-10-09) Memory: disk 4892 GB (BF16) · Min. VRAM 5384 GB approx. · fits on: Multi-node (does not fit on a Hopper node) Official BF16 4892 GB Requirement: GenAI Brief: does not fit on a Hopper node at any published precision. source License: Qwen3.8-Max (commercial terms) Source: artificialanalysis.ai/evaluations/terminalbench-4-0 · 2026-10-08
Qwen3.8 27B (xhigh) Alibaba	5.6	€3.52$3.94	OpenOpen	ProvenProven ⬇ 6.8M/mo HF · 4 rankingsWidely usedWidely used	€0.45$0.50 / €2.68$3.00	n/d	27B / 27B (dense)	20.0 GB approx. ▣ 1 consumer GPU ✓ Fits on your computer disk 55.6 GB 1× 80 GB (BF16) · 1× 24–32 GB (Q4)	Apache 2.0
Weights: OpenOpen 🤗 Hugging Face Maturity: ProvenProven 6.8 M HF downloads/mo · on OpenRouter · 4 leaderboards · 56 days · Adoption: Widely used (⬇ 6.8M/mo HF · 4 rankings) (leaderboards: AA Intelligence Index, AA Terminal-Bench 4.0, LMArena Text, LMArena WebDev; released: 2026-08-14; data from 2026-10-09) Memory: disk 55.6 GB (BF16) · Min. VRAM 20.0 GB approx. · fits on: 1× 80 GB (BF16) · 1× 24–32 GB (Q4) Official BF16 55.6 GB Official FP8 30.9 GB GGUF UD-Q4_K_M 16.5 GB Requirement: GenAI Brief: 55.6 GB BF16 fits on an 80 GB GPU without quantizing; the default «single-GPU» model. source License: Apache 2.0 Source: artificialanalysis.ai/evaluations/terminalbench-4-0 · 2026-10-08
MiniMax-M3 MiniMax	2.0	€2.71$3.04	OpenOpen	ProvenProven ⬇ 204k/mo HF · 4 rankingsWidely usedWidely used	€0.27$0.30 / €1.07$1.20	n/d	428B / 23B	284 GB approx. ▣ 1 node 8 GPU disk 854 GB 1 node 8× H100 (8 bits)	MiniMax Community (commercial use with notice)
Weights: OpenOpen 🤗 Hugging Face Maturity: ProvenProven 204k HF downloads/mo · on OpenRouter · 4 leaderboards · from 2026-06 · Adoption: Widely used (⬇ 204k/mo HF · 4 rankings) (leaderboards: AA Intelligence Index, AA Terminal-Bench 4.0, LMArena Text, LMArena WebDev; released: 2026-06-01; data from 2026-10-09) Memory: disk 854 GB (BF16) · Min. VRAM 284 GB approx. · fits on: 1 node 8× H100 (8 bits) Official BF16 854 GB Official MXFP8 444 GB GGUF MXFP4_MOE 256 GB Requirement: GenAI Brief: at 8 bits the weights fit on an 8× H100 node. source License: MiniMax Community (commercial use with notice) Source: artificialanalysis.ai/evaluations/terminalbench-4-0 · 2026-10-08
Nemotron 3 Ultra NVIDIA	0.5	€2.73$3.06	OpenOpen	ProvenProven OpenRouter top 20 · #14 · 3 rankingsWidely usedWidely used	€0.54$0.60 / €2.23$2.50	n/d	550B / 55B	390 GB approx. ▣ 1 node 8 GPU disk 1121 GB 4× B200/GB200 (NVFP4) or 8× H100	OpenMDW-1.1
Weights: OpenOpen 🤗 Hugging Face Maturity: ProvenProven 760k HF downloads/mo · top 20 OpenRouter · 3 leaderboards · from 2026-06 · Adoption: Widely used (OpenRouter top 20 · #14 · 3 rankings) (leaderboards: AA Intelligence Index, AA Terminal-Bench 4.0, LMArena Text; released: 2026-06-04; data from 2026-10-09) Memory: disk 1121 GB (BF16) · Min. VRAM 390 GB approx. · fits on: 4× B200/GB200 (NVFP4) or 8× H100 Official BF16 1121 GB Official NVFP4 352 GB Requirement: NVFP4 model card: minimum GPU 4× GB200, 4× B200, 4× GB300, 4× B300 or 8× H100. source License: OpenMDW-1.1 Source: artificialanalysis.ai/evaluations/terminalbench-4-0 · 2026-10-08
Muse Glimmer (high) Meta	0.5	€0.24$0.27	OpenOpen	ProvenProven ⬇ 269k/mo HF · 4 rankingsWidely usedWidely used	€0.29$0.32 / €1.20$1.35	n/d	30B / n/d	19.0 GB approx. ▣ 1 consumer GPU ✓ Fits on your computer disk 59.6 GB 1× 24–32 GB (quantized)	Apache 2.0
Weights: OpenOpen 🤗 Hugging Face Maturity: ProvenProven 269k HF downloads/mo · on OpenRouter · 4 leaderboards · 60 days · Adoption: Widely used (⬇ 269k/mo HF · 4 rankings) (leaderboards: AA Intelligence Index, AA Terminal-Bench 4.0, LMArena Text, LMArena WebDev; released: 2026-08-10; data from 2026-10-09) Memory: disk 59.6 GB (BF16) · Min. VRAM 19.0 GB approx. · fits on: 1× 24–32 GB (quantized) Official BF16 59.6 GB GGUF Q8_0 29.6 GB GGUF UD-Q4_K_XL 15.9 GB Requirement: Meta (via GenAI Brief): runs on 24 or 32 GB consumer GPUs. source License: Apache 2.0 Source: artificialanalysis.ai/evaluations/terminalbench-4-0 · 2026-10-08
Gemma 4 31B Google	0.0	self-host	OpenOpen	ProvenProven ⬇ 9.6M/mo HF · 4 rankingsWidely usedWidely used	n/d	n/d	30.7B / 30.7B (dense)	22.0 GB approx. ▣ 1 consumer GPU ✓ Fits on your computer disk 62.5 GB 1× 24 GB (approx.)	Apache 2.0
Weights: OpenOpen 🤗 Hugging Face Maturity: ProvenProven 9.6 M HF downloads/mo · on OpenRouter · 4 leaderboards · from 2026-04 · Adoption: Widely used (⬇ 9.6M/mo HF · 4 rankings) (leaderboards: AA Intelligence Index, AA Terminal-Bench 4.0, LMArena Text, LMArena WebDev; released: 2026-04-02; data from 2026-10-09) Score: 0 tasks solved in Terminal-Bench 4.0. Memory: disk 62.5 GB (BF16) · Min. VRAM 22.0 GB approx. · fits on: 1× 24 GB (approx.) Official BF16 62.5 GB GGUF Q8_0 33.2 GB GGUF Q4_K_M 18.3 GB Requirement: Google: 12B/26B A4B/31B designed for consumer GPUs and workstations. source License: Apache 2.0 Source: artificialanalysis.ai/evaluations/terminalbench-4-0 · 2026-10-08
gpt-oss-120b (high) OpenAI	0.0	€0.057$0.064	OpenOpen	ProvenProven ⬇ 4.0M/mo HF · 3 rankingsWidely usedWidely used	€0.13$0.15 / €0.53$0.59	n/d	117B / 5.1B	74.0 GB approx. ▣ 1× 80 GB disk 65.2 GB 1× 80 GB (H100 / MI300X)	Apache 2.0
Weights: OpenOpen 🤗 Hugging Face Maturity: ProvenProven 4.0 M HF downloads/mo · on OpenRouter · 3 leaderboards · from 2025-08 · Adoption: Widely used (⬇ 4.0M/mo HF · 3 rankings) (leaderboards: AA Intelligence Index, AA Terminal-Bench 4.0, LMArena Text; released: 2025-08-05; data from 2026-10-09) Score: 0 tasks solved in Terminal-Bench 4.0. Memory: disk 65.2 GB (MXFP4) · Min. VRAM 74.0 GB approx. · fits on: 1× 80 GB (H100 / MI300X) Official MXFP4 65.2 GB Requirement: Model card: fits on a single 80 GB GPU (H100 or MI300X) thanks to MXFP4. source License: Apache 2.0 Source: artificialanalysis.ai/evaluations/terminalbench-4-0 · 2026-10-08
Mistral Large 3 Mistral	0.0	€0.12$0.13	OpenOpen	ProvenProven ⬇ 2k/mo HF · 4 rankingsModerately usedModerately used	€0.45$0.50 / €1.34$1.50	n/d	675B / 41B	445 GB approx. ▣ 1 node 8 GPU disk 682 GB 1 node 8× H200/B200 (FP8) · 1 node H100/A100 (NVFP4)	Apache 2.0
Weights: OpenOpen 🤗 Hugging Face Maturity: ProvenProven 2k HF downloads/mo · on OpenRouter · 4 leaderboards · from 2025-12 · Adoption: Moderately used (⬇ 2k/mo HF · 4 rankings) (leaderboards: AA Intelligence Index, AA Terminal-Bench 4.0, LMArena Text, LMArena WebDev; released: 2025-12-02; data from 2026-10-09) Score: 0 tasks solved in Terminal-Bench 4.0. Memory: disk 682 GB (FP8) · Min. VRAM 445 GB approx. · fits on: 1 node 8× H200/B200 (FP8) · 1 node H100/A100 (NVFP4) Official FP8 682 GB Official NVFP4 403 GB GGUF Q4_K_M 407 GB Requirement: Model card: FP8 on a B200 or H200 node; NVFP4 on an H100 or A100 node. source License: Apache 2.0 Source: artificialanalysis.ai/evaluations/terminalbench-4-0 · 2026-10-08
Mellum2.1 JetBrains	n/d	self-host	OpenOpen	Barely testedBarely t. ⬇ 110/mo HF · 0 rankingsLittle usedLittle used	n/d	n/d	12B / 2.5B	12.0 GB approx. ▣ 1 consumer GPU ✓ Fits on your computer disk 24.3 GB 1× 16 GB (approx.)	Apache 2.0
Weights: OpenOpen 🤗 Hugging Face Maturity: Barely testedBarely t. No independent benchmark · 110 HF downloads/mo · 19 days · Adoption: Little used (⬇ 110/mo HF · 0 rankings) (released: 2026-09-20; data from 2026-10-09) Score: JetBrains' coding model; AA has not measured it on Terminal-Bench 4.0 (only its own SWE-bench Verified figures). Memory: disk 24.3 GB (BF16) · Min. VRAM 12.0 GB approx. · fits on: 1× 16 GB (approx.) Official BF16 24.3 GB GGUF Q8_0 12.9 GB GGUF Q4_K_M 8.8 GB License: Apache 2.0 Source: blog.jetbrains.com/ai/2026/10/mellum2-1-gets-to-work-a-fast-open-model · 2026-10-08

Keep exploring

Tools for codingCursor, Copilot, Claude Code, Codex…How we measure itTerminal-Bench, cost per task and criteria.Coding locallyOpen-source models on your own GPU.