33 language models ranked by independent benchmarks, with the price and speed next to every score. Updated September 2026.
Live LLM leaderboard with 33 language models ranked by independent benchmark scores, pricing, speed, and context window. Filter by task type or price tier to find the best model for your workload.
Claude Opus 5.5
Anthropic · High effort
53.6
$1.82 per task
Full answer in 41 s
Claude Fable 5.1
Anthropic · High effort
51.2
$3.91 per task
Full answer in 29 s
GPT-6 Astra
OpenAI · Medium effort
49.6
$1.54 per task
Full answer in 15 s
| # | Model | |||||
|---|---|---|---|---|---|---|
| 1 | Claude Opus 5.5Anthropic · High effort · Frontier | 53.6 | $1.82 | 41 s | $4.00 · $20.00 | 1.0M |
| 2 | Claude Fable 5.1Anthropic · High effort · Frontier | 51.2 | $3.91 | 29 s | $10.00 · $50.00 | 1.0M |
| 3 | Claude Opus 5Older versionAnthropic · Extra high effort · Frontier | 49.7 | $4.88 | 34 s | $5.00 · $25.00 | 1.0M |
| 4 | GPT-6 AstraOpenAI · Medium effort · Frontier | 49.6 | $1.54 | 15 s | $10.00 · $50.00 | 1.1M |
| 5 | Claude Fable 5Older versionAnthropic · Max effort · Frontier | 49.6 | $8.75 | 114 s | $10.00 · $50.00 | 1.0M |
| 6 | Muse Spark 1.3Meta · Max effort · Mid-Range | 48.1 | $1.60 | 36 s | $1.25 · $4.25 | 1.0M |
| 7 | Grok 4.7SpaceXAI · Extra high effort · Premium | 46.4 | $3.74 | 12 s | $2.00 · $6.00 | 500K |
| 8 | MiMo-V2.6-ProXiaomi · Standard effort · Budget | 46.3 | $0.133 | 61 s | $0.43 · $0.87 | 1.0M |
| 9 | Qwen3.8 MaxAlibaba · Max effort · Premium | 45.4 | $5.41 | 66 s | $2.00 · $6.00 | 984K |
| 10 | GLM-5.3Z AI · Max effort · Mid-Range | 44.8 | $2.01 | 39 s | $1.40 · $4.40 | 1.0M |
| 11 | GPT-6 SolOpenAI · Extra high effort · Premium | 44.1 | $0.532 | 43 s | $2.00 · $10.00 | 1.1M |
| 12 | GPT-5.6 SolOlder versionOpenAI · Extra high effort · Premium | 44 | $1.18 | 43 s | $4.00 · $20.00 | 1.1M |
| 13 | Grok 4.6Older versionSpaceXAI · Medium effort · Premium | 42.8 | $1.50 | 30 s | $2.00 · $6.00 | 500K |
| 14 | GLM-5.3 FlashZ AI · Standard effort · Budget | 41.8 | $0.253 | 61 s | $0.15 · $0.50 | 1.0M |
| 15 | Gemini 3.8 FlashGoogle · High effort · Mid-Range | 40.9 | $1.24 | 24 s | $0.75 · $3.75 | 1.0M |
| 16 | Gemini 3.7 FlashOlder versionGoogle · Medium effort · Mid-Range | 39.6 | – | 7.1 s | $0.75 · $3.75 | 1.0M |
| 17 | DeepSeek V4.1 FlashDeepSeek · Max effort · Budget | 39.5 | $0.265 | 12 s | $0.30 · $1.20 | 1.0M |
| 18 | DeepSeek V4 ProDeepSeek · Max effort · Mid-Range | 36 | $0.674 | 30 s | $1.32 · $3.96 | 1.0M |
| 19 | Kimi K3Moonshot AI · Low effort · Premium | 34.5 | – | 70 s | $3.00 · $15.00 | 1.0M |
| 20 | Claude Sonnet 5Anthropic · Extra high effort · Premium | 34.4 | $2.87 | 24 s | $2.00 · $10.00 | 1.0M |
| 21 | DeepSeek V4 FlashOlder versionDeepSeek · Max effort · Budget | 34.3 | $0.220 | 12 s | $0.44 · $1.32 | 1.0M |
| 22 | GPT-6 LunaOpenAI · Extra high effort · Budget | 33.9 | $0.042 | 23 s | $0.10 · $0.50 | 1.1M |
| 23 | GLM-5.2Older versionZ AI · Max effort · Mid-Range | 33.7 | $0.965 | 37 s | $1.40 · $4.40 | 128K |
| 24 | GPT-5.6 LunaOlder versionOpenAI · High effort · Budget | 32.1 | $0.044 | 13 s | $0.20 · $1.20 | 1.1M |
| 25 | Gemini 3.1 ProGoogle · Standard effort · Premium | 29.7 | $0.675 | 35 s | $2.00 · $12.00 | 1.0M |
| 26 | MiniMax M3MiniMax · Standard effort · Mid-Range | 29.2 | $0.508 | 15 s | $0.30 · $1.20 | 1.0M |
| 27 | Claude Sonnet 4.6Older versionAnthropic · No reasoning effort · Mid-Range | 24.7 | – | 13 s | $3.00 · $15.00 | 1.0M |
| 28 | MiniMax M2.7Older versionMiniMax · Standard effort · Mid-Range | 22.8 | $0.102 | 50 s | $0.30 · $1.20 | 205K |
| 29 | Gemini 3 ProOlder versionGoogle · Low effort · Premium | 22.3 | – | – | $2.00 · $12.00 | 1.0M |
| 30 | Claude Haiku 4.5Anthropic · Not measured · Budget | 17 | – | – | $1.00 · $5.00 | 200K |
| 31 | DeepSeek R1 (Free)DeepSeek · Not measured · Free | 11 | – | – | Free · Free | 64K |
| 32 | Llama 3.3 70B (Free)Meta · Not measured · Free | 8 | – | – | Free · Free | 128K |
| 33 | Qwen 2.5 VL 72B (Free)Alibaba · Not measured · Free | 8 | – | – | Free · Free | 128K |
Intelligence, cost and speed from Artificial Analysis. Everyday uses each model's strongest reasoning setting that gives a full answer within 45 seconds; Max uses its highest-scoring setting. Cost per task is what one Intelligence Index task cost at that setting, thinking tokens included, so it reflects how much a model actually writes. A dash means Artificial Analysis hasn't measured it.
Claude Opus 5.5 took the top spot on September 22. At the settings people actually run, it scores 53.6 on the Intelligence Index (57.6 at max effort), ahead of Claude Fable 5.1 at 51.2 and GPT-6 Astra at 49.6. It also costs less than half as much per task as Fable 5.1, and its $4/$20 list price is 20% below Opus 5. Fable 5.1 still leads on agentic knowledge work. For almost everything else, Opus 5.5 is the pick.
If the scores look lower than last month, that's the index, not the models. Artificial Analysis re-based it three times in September. Version 4.3 added Terminal-Bench 4.0 and AutomationBench-AA, and v4.3.2 re-anchored GDPval-AA. Top scores that sat around 60 now land in the 50s, so compare models on the same version, not against old screenshots.
The value story is GPT-6. Sol scores 48 (GPT-5.6 Sol was 47) and Luna matches its predecessor at 37, each at half the price: $2/$10 for Sol, $0.10/$0.50 for Luna. Meta's Muse Spark 1.3 also scores 48, at $1.25/$4.25 and around 219 tokens per second. Below that, Grok 4.7 (46 at $2/$6) and Xiaomi's open-weights MiMo-V2.6-Pro (46 at $0.43/$0.87) are where agents and batch jobs get cheap.
For deeper analysis of each model, see our Best LLM in 2026 guide, coding model rankings, GPT-5.6 breakdown, or cost comparison. Use the Model Selector for personalized recommendations.
Most LLM leaderboards show a single metric. Quality score, or ELO rating, or some proprietary number. That tells you which model is smartest but nothing about whether you can afford to use it.
This leaderboard puts quality (the Artificial Analysis Intelligence Index v4.3.2), real cost per task and time to a full answer side by side, with list prices and context windows next to them. All of it comes from Artificial Analysis.
We rank each model at the reasoning setting people actually run: its strongest setting that gives a full answer within 45 seconds. Benchmark headlines use maximum effort, which often costs three times as much and can take minutes per answer. Switch the table to Max effort to see those numbers. Cost per task is what one benchmark task cost at that setting, thinking tokens included, so a model that thinks a lot can cost more than its list price suggests.
We exclude deprecated models, extremely limited-availability models, and models that don't offer meaningful advantages over alternatives in their price range. We keep the list to models worth choosing between.
An LLM leaderboard ranks large language models by objective performance metrics. This leaderboard uses quality scores from Artificial Analysis Intelligence Index, combined with pricing, speed, and context window data to give a complete picture of each model's capabilities and cost.
Claude Opus 5.5. At the settings people actually run it scores 53.6 on the Artificial Analysis Intelligence Index (57.6 at max effort), ahead of Claude Fable 5.1 at 51.2 and GPT-6 Astra at 49.6. For value, MiMo-V2.6-Pro scores 46.3 at about $0.13 per task and GPT-6 Sol scores 44.1 at $0.53. On a tight budget, GPT-6 Luna scores 33.9 at about $0.04 per task.
Quality scores come from the Artificial Analysis Intelligence Index v4.3.2, an independent index built from 10 evaluations covering agents, coding, general knowledge, and scientific reasoning. We score each model at its strongest reasoning setting that answers within 45 seconds, not at maximum effort. Cost per task and response times come from the same Artificial Analysis tests.
Claude Opus 5.5 leads the Artificial Analysis Coding Agent Index v1.5 at 66. Claude Fable 5.1 and GPT-6 Astra tie at 62. GPT-6 Sol (57) and Grok 4.7 (56) get close for $2 per million input tokens. For budget coding work, GPT-6 Luna scores 41 at $0.10/$0.50.
Because few people run them that way. Max effort adds a few points on benchmarks but often triples the cost and can take minutes per answer: GPT-6 Sol at max needs almost three minutes. Effort labels also mean different things at different labs, so we pick each model's strongest setting that answers within 45 seconds. The Max effort switch shows the benchmark numbers if you want them.
Everything from model picks to CSV converters, with no account needed.