The evolving landscape of models
Which model is best at what, checked daily against public leaderboards.
Atlas is being built to choose a model for each task; this page tracks the public evidence for that kind of choice.
Rankings reflect public leaderboards; they do not state which models Atlas runs.
Best at
Top 3 among ranked models, plus the best open-weight model. LiveBench 2.0 suites, release Oct 13, LiveBench. The overall leader can be off this list. Tasks without a public benchmark aren't shown.
| Task type, among ranked models | #1 | #2 | #3 | Best open-weight, among ranked models |
|---|---|---|---|---|
| Coding | Claude Fable 5.1 (97.7) | Claude Opus 5.5 (96.3) | GPT-6.1 Sol (96.0) | DeepSeek V4.1 Flash (85.4) |
| Agentic | Claude Opus 5.5 (85.0) | GPT-6 Astra (82.4) | Claude Fable 5.1 (80.8) | DeepSeek V4.1 Flash (43.8) |
| Reasoning | Gemini 4 Argon (95.1) | GPT-6 Astra (92.7) | GPT-6.1 Sol (92.6) | Kimi K3 (90.7) |
| Browser Use | Claude Fable 5.1 (100.0) | Gemini 4 Argon (92.0) | GPT-6.1 Sol (88.0) | GLM-5.3-Flash (84.0) |
| Data Analysis | Gemini 4 Argon (85.2) | GPT-6 Astra (83.0) | GPT-6.1 Sol (82.7) | DeepSeek V4.1 Flash (79.3) |
| Research | Grok 4.7 (86.4) | DeepSeek V4.1 Flash (85.3) | GPT-6 Astra (85.2) | DeepSeek V4.1 Flash (85.3) |
| Design | Claude Opus 5.5 (86.9) | Claude Sonnet 5.5 (85.0) | Claude Fable 5.1 (83.4) | GLM-5.3 (77.3) |
Intelligence and execution
| Rank | Model | Maker | Open-weight | Atlas composite | Publishers | Artificial AnalysisOct 13 | LMArenaOct 13 | LiveBench CodingOct 13 | LiveBench AgenticOct 13 | Share of tokens among listed models (OpenRouter, week of Oct 12) |
|---|---|---|---|---|---|---|---|---|---|---|
| 1 | Claude Opus 5.5 | Anthropic | No | 94.1 | 3 of 3 | 58 | 1508±8 | 96.3 | 85.0 | 6.8% |
| 2 | Claude Fable 5.1 | Anthropic | No | 82.4 | 3 of 3 | 53 | 1501±6 | 97.7 | 80.8 | not in top 20 |
| 3 | Gemini 4 Argon | No | 81.7 | 3 of 3 | 53 | 1525±9 | 89.7 | 69.9 | not in top 20 | |
| 4 | Claude Sonnet 5.5 | Anthropic | No | 77.5 | 3 of 3 | 56 | 1475±8 | 94.7 | 78.8 | 3.4% |
| 5 | GPT-6 Astra | OpenAI | No | 73.6 | 3 of 3 | 53 | 1477±6 | 93.7 | 82.4 | not in top 20 |
| 6 | GPT-6.1 Sol | OpenAI | No | 71.9 | 3 of 3 | 52 | 1484±8 | 96.0 | 73.0 | 2.8% |
| 7 | Muse Spark 1.3 | Meta | No | 53.4 | 3 of 3 | 48 | 1494±6 | 90.2 | 40.4 | not listed |
| 8 | GLM-5.3 | Zhipu AI (Z.ai) | license · weights | 38.1 | 3 of 3 | 45 | 1479±5 | 82.8 | 35.9 | 5.0% |
| 9 | Grok 4.7 | SpaceXSI | No | 35.1 | 3 of 3 | 46 | 1444±7 | 84.5 | 50.1 | not in top 20 |
| 10 | GPT-5.6 Terra | OpenAI | No | 35.1 | 3 of 3 | 42 | 1466±4 | 86.2 | 48.4 | not in top 20 |
| 11 | Gemini 3.8 Flash | No | 33.9 | 3 of 3 | 41 | 1496±5 | 82.2 | 30.1 | 2.6% | |
| 12 | DeepSeek V4.1 Flash | DeepSeek | license · weights | 30.5 | 3 of 3 | 39 | 1475±6 | 85.4 | 43.8 | 51.4% |
| 13 | Qwen3.8 Max | Alibaba (Qwen) | No | 29.6 | 3 of 3 | 45 | 1482±5 | 60.3 | 42.3 | not in top 20 |
| 14 | Kimi K3 | Moonshot AI | license · weights | 29.3 | 3 of 3 | 44 | 1488±5 | 59.3 | 42.8 | 2.9% |
| 15 | GLM-5.3-Flash | Zhipu AI (Z.ai) | license · weights | 27.5 | 3 of 3 | 42 | 1474±5 | 80.6 | 28.4 | 16.1% |
| 16 | GPT-6 Luna | OpenAI | No | 20.1 | 3 of 3 | 38 | 1443±5 | 84.5 | 48.1 | 9.0% |
Scores within about 10 points are effectively tied.
Composite score: Artificial Analysis 35%, LMArena 25%, LiveBench Coding 20%, LiveBench Agentic 20%.
LiveBench counts as one source, and a model needs all three sources to be ranked.
OpenRouter shows usage share only, not quality.
More columns
Artificial Analysis · Oct 13. LMArena · Oct 13. LiveBench Coding · Oct 13. LiveBench Agentic · Oct 13. OpenRouter · Oct 12.
- Claude Opus 5.5: AA 58 · Intelligence Index; LMArena 1508±8 · Text Arena; LiveBench Coding 96.3 · Coding; LiveBench Agentic 85.0 · Agentic; usage share 6.8%
- Claude Fable 5.1: AA 53 · Intelligence Index; LMArena 1501±6 · Text Arena; LiveBench Coding 97.7 · Coding; LiveBench Agentic 80.8 · Agentic; usage share not in top 20
- Gemini 4 Argon: AA 53 · Intelligence Index; LMArena 1525±9 · Text Arena; LiveBench Coding 89.7 · Coding; LiveBench Agentic 69.9 · Agentic; usage share not in top 20
- Claude Sonnet 5.5: AA 56 · Intelligence Index; LMArena 1475±8 · Text Arena; LiveBench Coding 94.7 · Coding; LiveBench Agentic 78.8 · Agentic; usage share 3.4%
- GPT-6 Astra: AA 53 · Intelligence Index; LMArena 1477±6 · Text Arena; LiveBench Coding 93.7 · Coding; LiveBench Agentic 82.4 · Agentic; usage share not in top 20
- GPT-6.1 Sol: AA 52 · Intelligence Index; LMArena 1484±8 · Text Arena; LiveBench Coding 96.0 · Coding; LiveBench Agentic 73.0 · Agentic; usage share 2.8%
- Muse Spark 1.3: AA 48 · Intelligence Index; LMArena 1494±6 · Text Arena; LiveBench Coding 90.2 · Coding; LiveBench Agentic 40.4 · Agentic; usage share not listed
- GLM-5.3: AA 45 · Intelligence Index; LMArena 1479±5 · Text Arena; LiveBench Coding 82.8 · Coding; LiveBench Agentic 35.9 · Agentic; usage share 5.0%
- Grok 4.7: AA 46 · Intelligence Index; LMArena 1444±7 · Text Arena; LiveBench Coding 84.5 · Coding; LiveBench Agentic 50.1 · Agentic; usage share not in top 20
- GPT-5.6 Terra: AA 42 · Intelligence Index; LMArena 1466±4 · Text Arena; LiveBench Coding 86.2 · Coding; LiveBench Agentic 48.4 · Agentic; usage share not in top 20
- Gemini 3.8 Flash: AA 41 · Intelligence Index; LMArena 1496±5 · Text Arena; LiveBench Coding 82.2 · Coding; LiveBench Agentic 30.1 · Agentic; usage share 2.6%
- DeepSeek V4.1 Flash: AA 39 · Intelligence Index; LMArena 1475±6 · Text Arena; LiveBench Coding 85.4 · Coding; LiveBench Agentic 43.8 · Agentic; usage share 51.4%
- Qwen3.8 Max: AA 45 · Intelligence Index; LMArena 1482±5 · Text Arena; LiveBench Coding 60.3 · Coding; LiveBench Agentic 42.3 · Agentic; usage share not in top 20
- Kimi K3: AA 44 · Intelligence Index; LMArena 1488±5 · Text Arena; LiveBench Coding 59.3 · Coding; LiveBench Agentic 42.8 · Agentic; usage share 2.9%
- GLM-5.3-Flash: AA 42 · Intelligence Index; LMArena 1474±5 · Text Arena; LiveBench Coding 80.6 · Coding; LiveBench Agentic 28.4 · Agentic; usage share 16.1%
- GPT-6 Luna: AA 38 · Intelligence Index; LMArena 1443±5 · Text Arena; LiveBench Coding 84.5 · Coding; LiveBench Agentic 48.1 · Agentic; usage share 9.0%
Share of tokens among listed models (OpenRouter, week of Oct 12)
- Claude Opus 5.5: 6.8%
- Claude Fable 5.1: not in top 20
- Gemini 4 Argon: not in top 20
- Claude Sonnet 5.5: 3.4%
- GPT-6 Astra: not in top 20
- GPT-6.1 Sol: 2.8%
- Muse Spark 1.3: not listed
- GLM-5.3: 5.0%
- Grok 4.7: not in top 20
- GPT-5.6 Terra: not in top 20
- Gemini 3.8 Flash: 2.6%
- DeepSeek V4.1 Flash: 51.4%
- Qwen3.8 Max: not in top 20
- Kimi K3: 2.9%
- GLM-5.3-Flash: 16.1%
- GPT-6 Luna: 9.0%
Cost efficiency
Each estimate counts the task's input and output tokens at the company's list price, so real costs usually run higher.
Prices: official vendor pricing pages, checked Oct 13, 2026.
Estimated cost per result (lower bound; lower is cheaper)
Read a 20-page PDF and convert it to structured Markdown (30,000 in, 8,000 out)
Under $0.01: GLM-5.3-Flash, GPT-6 Luna
$0.01–$0.10: Muse Spark 1.3, GLM-5.3, Gemini 3.8 Flash, DeepSeek V4.1 Flash
Over $0.10: Claude Opus 5.5, Claude Fable 5.1, Gemini 4 Argon, Claude Sonnet 5.5, GPT-6 Astra, GPT-6.1 Sol, Grok 4.7, GPT-5.6 Terra, Qwen3.8 Max, Kimi K3
| Model | Estimated cost | Tier used | composite points per dollar |
|---|---|---|---|
| GLM-5.3-Flash | $0.0085 | Standard rate | 3235.3 |
| GPT-6 Luna | $0.007 | Short context | 2871.4 |
| DeepSeek V4.1 Flash | $0.0186 | Peak rate | 1639.8 |
| Muse Spark 1.3 | $0.0715 | Standard rate | 746.9 |
| Gemini 3.8 Flash | $0.0525* | Standard rate | 645.7 |
| Gemini 4 Argon | $0.14* | Standard rate | 583.6 |
| Claude Sonnet 5.5 | $0.14 | Standard rate | 553.6 |
| GPT-6.1 Sol | $0.14 | Short context | 513.6 |
| GLM-5.3 | $0.0772 | Standard rate | 493.5 |
| Claude Opus 5.5 | $0.28 | Standard rate | 336.1 |
| Grok 4.7 | $0.108 | Short context | 325.0 |
| Qwen3.8 Max | $0.108 | Standard rate | 274.1 |
| GPT-5.6 Terra | $0.156 | Short context | 225.0 |
| Kimi K3 | $0.21 | Standard rate | 139.5 |
| Claude Fable 5.1 | $0.7 | Standard rate | 117.7 |
| GPT-6 Astra | $0.7 | Short context | 105.1 |
Update one recurring document from new notes (15,000 in, 3,000 out)
Under $0.01: DeepSeek V4.1 Flash, GLM-5.3-Flash, GPT-6 Luna
$0.01–$0.10: Gemini 4 Argon, Claude Sonnet 5.5, GPT-6.1 Sol, Muse Spark 1.3, GLM-5.3, Grok 4.7, GPT-5.6 Terra, Gemini 3.8 Flash, Qwen3.8 Max, Kimi K3
Over $0.10: Claude Opus 5.5, Claude Fable 5.1, GPT-6 Astra
| Model | Estimated cost | Tier used | composite points per dollar |
|---|---|---|---|
| GLM-5.3-Flash | $0.0037 | Standard rate | 7333.3 |
| GPT-6 Luna | $0.003 | Short context | 6700.0 |
| DeepSeek V4.1 Flash | $0.0081 | Peak rate | 3765.4 |
| Muse Spark 1.3 | $0.0315 | Standard rate | 1695.2 |
| Gemini 3.8 Flash | $0.0225* | Standard rate | 1506.7 |
| Gemini 4 Argon | $0.06* | Standard rate | 1361.7 |
| Claude Sonnet 5.5 | $0.06 | Standard rate | 1291.7 |
| GPT-6.1 Sol | $0.06 | Short context | 1198.3 |
| GLM-5.3 | $0.0342 | Standard rate | 1114.0 |
| Claude Opus 5.5 | $0.12 | Standard rate | 784.2 |
| Grok 4.7 | $0.048 | Short context | 731.3 |
| Qwen3.8 Max | $0.048 | Standard rate | 616.7 |
| GPT-5.6 Terra | $0.066 | Short context | 531.8 |
| Kimi K3 | $0.09 | Standard rate | 325.6 |
| Claude Fable 5.1 | $0.3 | Standard rate | 274.7 |
| GPT-6 Astra | $0.3 | Short context | 245.3 |
Fix one bug in a mid-size codebase (60,000 in, 6,000 out)
Under $0.01: GPT-6 Luna
$0.01–$0.10: Gemini 3.8 Flash, DeepSeek V4.1 Flash, GLM-5.3-Flash
Over $0.10: Claude Opus 5.5, Claude Fable 5.1, Gemini 4 Argon, Claude Sonnet 5.5, GPT-6 Astra, GPT-6.1 Sol, Muse Spark 1.3, GLM-5.3, Grok 4.7, GPT-5.6 Terra, Qwen3.8 Max, Kimi K3
| Model | Estimated cost | Tier used | composite points per dollar |
|---|---|---|---|
| GLM-5.3-Flash | $0.012 | Standard rate | 2291.7 |
| GPT-6 Luna | $0.009 | Short context | 2233.3 |
| DeepSeek V4.1 Flash | $0.0252 | Peak rate | 1210.3 |
| Muse Spark 1.3 | $0.1005 | Standard rate | 531.3 |
| Gemini 3.8 Flash | $0.0675* | Standard rate | 502.2 |
| Gemini 4 Argon | $0.18* | Standard rate | 453.9 |
| Claude Sonnet 5.5 | $0.18 | Standard rate | 430.6 |
| GPT-6.1 Sol | $0.18 | Short context | 399.4 |
| GLM-5.3 | $0.1104 | Standard rate | 345.1 |
| Claude Opus 5.5 | $0.36 | Standard rate | 261.4 |
| Grok 4.7 | $0.156 | Short context | 225.0 |
| Qwen3.8 Max | $0.156 | Standard rate | 189.7 |
| GPT-5.6 Terra | $0.192 | Short context | 182.8 |
| Kimi K3 | $0.27 | Standard rate | 108.5 |
| Claude Fable 5.1 | $0.9 | Standard rate | 91.6 |
| GPT-6 Astra | $0.9 | Short context | 81.8 |
* Gemini 4 Argon: promotional price until Dec 1, 2026
* Gemini 3.8 Flash: promotional price until Jan 1, 2027
Muse Spark 1.3: Priced in Standard mode. Contributor mode trains on your data.
No published price: Claude Haiku 5.5, DeepSeek V4 Flash Vision, Muse Wonder Full, Muse Wonder, MiMo-V2.6-Pro, Step 5 Preview, Ling 3.1 Flash, Qwen3.8 2.4T A95B, Qwen3.8-Flash-Next, Mistral Large 4 Preview, MiMo-V2.6-Flash, DeepSeek V4 Pro 0813.
Open-weight only
Protect and train on your proprietary data. 4 of the 16 ranked models are open-weight: GLM-5.3, Kimi K3, DeepSeek V4.1 Flash and GLM-5.3-Flash.
- GLM-5.3 (38.1) license · weights
- DeepSeek V4.1 Flash (30.5) license · weights
- Kimi K3 (29.3) license · weights
- GLM-5.3-Flash (27.5) license · weights
Listed but not ranked: MiMo-V2.6-Pro (2 publishers; HF MIT tag, no LICENSE file).
Not ranked
A model needs 3 sources from 3 different publishers.
Full model records and availability
- Claude Haiku 5.5 (Anthropic): 2 of 3 publishers.
- DeepSeek V4 Flash Vision (DeepSeek): 2 of 3 publishers.
- Muse Wonder Full (Meta): 1 of 3 publishers.
- Muse Wonder (Meta): 2 of 3 publishers.
- MiMo-V2.6-Pro (Xiaomi): 2 of 3 publishers.
- Step 5 Preview (StepFun): 2 of 3 publishers.
- Ling 3.1 Flash (inclusionAI (Ant Group)): 1 of 3 publishers.
- Qwen3.8 2.4T A95B (Alibaba (Qwen)): 1 of 3 publishers.
- Qwen3.8-Flash-Next (Alibaba (Qwen)): 1 of 3 publishers.
- Mistral Large 4 Preview (Mistral AI): 1 of 3 publishers.
- MiMo-V2.6-Flash (Xiaomi): 2 of 3 publishers.
- DeepSeek V4 Pro 0813 (DeepSeek): 2 of 3 publishers.
Mistral Large 4 (Mistral AI)
Mistral AI released Mistral Large 4 in public preview on October 6, 2026, a general-purpose multimodal model with a mixture-of-experts design and listed context 1M.
- Public preview, not general availability.
- Field news about a vendor release. It does not mean Atlas uses or integrates this model.
Method and sources
Each source is scaled 0–100 across the ranked models, using min-max. Published values are stored unchanged. If a model is missing one LiveBench suite, its weights are re-scaled. Ties go to the model with the higher LMArena low-end score. SWE-bench Verified, Epoch AI and Hugging Face boards are not in this rubric.
- Artificial Analysis · Intelligence Index 2026-10-13 ~14:00 PT
- LMArena · Text Arena 2026-10-13 (page date)
- LiveBench · LiveBench 2.0 2026-10-13 ~14:15 PT
- OpenRouter · rankings week view, bucket 2026-10-12, read ~14:10 PT