Best performance
The strongest model, whatever the price.
Claude Fable 5 Anthropic · FrontierTop- Per 1M tokens
- $18
About $810/mo at regular usage
Which AI model should I use?
Compare models by human preference ratings, benchmark scores and API price. Find the cheapest option within each capability level. How this works · Updated 8 hours ago
I need a model for
Providers
Best value
A Top model at the lowest token price.
Gemini 3.8 Flash Google · StandardTop- Per 1M tokens
- $1.49
About $67/mo at regular usage
Routine work
The cheapest model that does the job.
Qwen3.5-Flash Qwen · SmallCapable- Per 1M tokens
- $0.10
About $4.7/mo at regular usage
Capability against cost for coding tasks
Similar scores, cheapest first. Prices per 1M tokens, blended.
Top As good as the best
- Gemini 3.8 Flash$1.49/1MVotes1530Benchmark76.3
- Claude Sonnet 5.5$3.60/1MVotes1529
- Gemini 4 Argon$3.60/1MVotes1550
- GPT-6.1 Sol$3.60/1MVotes1541
- Claude Opus 5.5$7.20/1MVotes1551
- GPT-5.6 Sol$7.20/1MVotes1533Benchmark78.3
- Claude Opus 4.6$9.00/1MVotes1546
- Claude Opus 4.7$9.00/1MVotes1546Benchmark73.6
- Claude Opus 4.8$9.00/1MVotes1529Benchmark74.3
- Claude Opus 5$9.00/1MVotes1534Benchmark78.0
- Claude Fable 5$18/1MVotes1550Benchmark76.5
- Claude Fable 5.1$18/1MVotes1537Benchmark80.7
- GPT-6 Astra$18/1MVotes1542Benchmark76.9
Strong Close to the best
- DeepSeek V4.1 Flash$0.14/1MVotes1523
- GPT-6 Luna$0.18/1MVotes1506
- DeepSeek V4 Flash Vision Exp$0.30/1MBenchmark65.0
- GPT-5.6 Luna$0.40/1MVotes1502Benchmark71.4
- Qwen3.8 27B$0.49/1MVotes1498Benchmark68.1
- DeepSeek V4 Pro 0813$0.92/1MVotes1500Benchmark68.8
- Gemini 3 Flash Preview$1.00/1MVotes1505
- Kimi K2.7 Code$1.18/1MBenchmark60.8
- Kimi K2.6$1.20/1MVotes1516Benchmark61.8
- Gemini 3.6 Flash$1.35/1MVotes1518Benchmark69.2
- Gemini 3.7 Flash$1.35/1MVotes1520Benchmark76.1
- Qwen3.7 Max$2.06/1MVotes1523Benchmark66.0
- Qwen3.8 2.4T A95B$2.80/1MBenchmark71.9
- Qwen3.8 Max (0902)$2.80/1MVotes1525Benchmark76.2
- Gemini 3.5 Flash$3.00/1MVotes1508Benchmark70.1
- Kimi K3$3.15/1MVotes1540Benchmark72.0
- Claude Sonnet 5$3.60/1MVotes1520Benchmark71.5
- GPT-6 Sol$3.60/1MVotes1519
- Gemini 3.1 Pro Preview$4.00/1MVotes1521Benchmark68.8
- GPT-5.6 Terra$4.00/1MVotes1517Benchmark76.7
- GPT-5.2 Chat$4.20/1MVotes1515
- GPT-5.4$5.00/1MVotes1511Benchmark71.1
- Claude Sonnet 4.6$5.40/1MVotes1528Benchmark63.0
- Claude Opus 4.5$9.00/1MVotes1522
- GPT-5.5$10/1MVotes1510Benchmark74.9
Capable Fine for routine work
- Qwen3.5-Flash$0.10/1MVotes1435
- DeepSeek V4 Flash 0423$0.11/1MVotes1484Benchmark56.2
- DeepSeek V4 Flash 0731$0.14/1MVotes1484Benchmark69.1
- Claude Haiku 5.5$0.18/1MVotes1489
- GPT-5.4 Nano$0.41/1MVotes1460Benchmark56.1
- Gemini 3.1 Flash Lite$0.50/1MVotes1458
- Qwen3.7 Plus$0.51/1MVotes1500Benchmark55.9
- Qwen3.5-122B-A10B$0.62/1MVotes1458Benchmark45.7
- Qwen3.6 Plus$0.65/1MVotes1496Benchmark54.5
- Qwen3.5-27B$0.73/1MVotes1451
- Gemini 3.5 Flash Lite$0.74/1MVotes1502Benchmark49.3
- Qwen3.6 27B$0.90/1MBenchmark53.7
- Kimi K2.5$0.90/1MVotes1504Benchmark46.8
- Qwen3.5 397B A17B$1.12/1MVotes1491Benchmark48.2
- DeepSeek V4 Pro 0423$1.14/1MVotes1500Benchmark59.4
- GPT-5.4 Mini$1.50/1MVotes1495Benchmark56.1
- GPT-5.1$3.00/1MVotes1476Benchmark49.4
- GPT-5.2$4.20/1MVotes1482
Basic Noticeably weaker
- Gemma 4 26B A4B$0.13/1MVotes1482Benchmark39.3
- Qwen3.5-9B$0.19/1MBenchmark28.7
- Gemma 4 31B$0.24/1MVotes1502Benchmark43.4
- Qwen3 Coder Next$0.26/1MBenchmark36.2
- Qwen3.6 35B A3B$0.26/1MBenchmark41.9
- DeepSeek V3.2$0.28/1MVotes1469Benchmark44.2
- Qwen3.5-35B-A3B$0.31/1MVotes1435Benchmark37.0
- Gemini 3.1 Flash Lite Preview$0.50/1MVotes1458Benchmark34.7
- Kimi K2 Thinking$0.98/1MBenchmark21.0
28 current models have no score for this task yet and are not placed.
Equivalents across providers
The newest model at each place in a provider's range, with its level for coding tasks.
Frontier
- Claude Fable 5.1Top
- GPT-6 AstraTop
- Gemini 4 ArgonTop
- DeepSeekNone
- QwenNone
- Moonshot AINone
Large
- Claude Opus 5.5Top
- GPT-6.1 SolTop
- Gemini 3.1 Pro PreviewStrong
- DeepSeek V4 Pro 0813Strong
- Qwen3.8 Max (0902)Strong
- Kimi K3.1 ProNo score
Standard
- Claude Sonnet 5.5Top
- GPT-5.6 TerraStrong
- Gemini 3.8 FlashTop
- DeepSeek V3.2Basic
- Qwen3.7 PlusCapable
- Kimi K3Strong
Capability by provider for coding tasks
Each provider's current models, ranked by Human votes.
Anthropic
- Claude Opus 5.5TopVotes1551
- Claude Fable 5TopVotes1550Benchmark76.5
- Claude Opus 4.6TopVotes1546
- Claude Opus 4.7TopVotes1546Benchmark73.6
- Claude Fable 5.1TopVotes1537Benchmark80.7
- Claude Opus 5TopVotes1534Benchmark78.0
- Claude Opus 4.8TopVotes1529Benchmark74.3
- Claude Sonnet 5.5TopVotes1529
- Claude Sonnet 4.6StrongVotes1528Benchmark63.0
- Claude Opus 4.5StrongVotes1522
- Claude Sonnet 5StrongVotes1520Benchmark71.5
- Claude Haiku 5.5CapableVotes1489
OpenAI
- GPT-6 AstraTopVotes1542Benchmark76.9
- GPT-6.1 SolTopVotes1541
- GPT-5.6 SolTopVotes1533Benchmark78.3
- GPT-6 SolStrongVotes1519
- GPT-5.6 TerraStrongVotes1517Benchmark76.7
- GPT-5.2 ChatStrongVotes1515
- GPT-5.4StrongVotes1511Benchmark71.1
- GPT-5.5StrongVotes1510Benchmark74.9
- GPT-6 LunaStrongVotes1506
- GPT-5.6 LunaStrongVotes1502Benchmark71.4
- GPT-5.4 MiniCapableVotes1495Benchmark56.1
- GPT-5.2CapableVotes1482
- GPT-5.1CapableVotes1476Benchmark49.4
- GPT-5.4 NanoCapableVotes1460Benchmark56.1
- Gemini 4 ArgonTopVotes1550
- Gemini 3.8 FlashTopVotes1530Benchmark76.3
- Gemini 3.1 Pro PreviewStrongVotes1521Benchmark68.8
- Gemini 3.7 FlashStrongVotes1520Benchmark76.1
- Gemini 3.6 FlashStrongVotes1518Benchmark69.2
- Gemini 3.5 FlashStrongVotes1508Benchmark70.1
- Gemini 3 Flash PreviewStrongVotes1505
- Gemma 4 31BBasicVotes1502Benchmark43.4
- Gemini 3.5 Flash LiteCapableVotes1502Benchmark49.3
- Gemma 4 26B A4BBasicVotes1482Benchmark39.3
- Gemini 3.1 Flash LiteCapableVotes1458
- Gemini 3.1 Flash Lite PreviewBasicVotes1458Benchmark34.7
DeepSeek
- DeepSeek V4.1 FlashStrongVotes1523
- DeepSeek V4 Pro 0423CapableVotes1500Benchmark59.4
- DeepSeek V4 Pro 0813StrongVotes1500Benchmark68.8
- DeepSeek V4 Flash 0423CapableVotes1484Benchmark56.2
- DeepSeek V4 Flash 0731CapableVotes1484Benchmark69.1
- DeepSeek V3.2BasicVotes1469Benchmark44.2
- DeepSeek V4 Flash Vision ExpStrongBenchmark65.0
Qwen
- Qwen3.8 Max (0902)StrongVotes1525Benchmark76.2
- Qwen3.7 MaxStrongVotes1523Benchmark66.0
- Qwen3.7 PlusCapableVotes1500Benchmark55.9
- Qwen3.8 27BStrongVotes1498Benchmark68.1
- Qwen3.6 PlusCapableVotes1496Benchmark54.5
- Qwen3.5 397B A17BCapableVotes1491Benchmark48.2
- Qwen3.5-122B-A10BCapableVotes1458Benchmark45.7
- Qwen3.5-27BCapableVotes1451
- Qwen3.5-FlashCapableVotes1435
- Qwen3.5-35B-A3BBasicVotes1435Benchmark37.0
- Qwen3 Coder NextBasicBenchmark36.2
- Qwen3.5-9BBasicBenchmark28.7
- Qwen3.6 27BCapableBenchmark53.7
- Qwen3.6 35B A3BBasicBenchmark41.9
- Qwen3.8 2.4T A95BStrongBenchmark71.9
Moonshot AI
- Kimi K3StrongVotes1540Benchmark72.0
- Kimi K2.6StrongVotes1516Benchmark61.8
- Kimi K2.5CapableVotes1504Benchmark46.8
- Kimi K2 ThinkingBasicBenchmark21.0
- Kimi K2.7 CodeStrongBenchmark60.8
Raw data for coding tasks
Source scores and model prices, as published.
| Human votesArena. Which answer people prefer. Higher is better. | Benchmark scoreArtificial Analysis indices. Higher is better. | OpenRouterPrice in dollars per million tokens. | |||||||
|---|---|---|---|---|---|---|---|---|---|
| Claude Opus 5.5 | Anthropic | 1551 | – | $4.00 | $20 | $0.20 | $7.20 | 1M | Sep 22, 2026 |
| Claude Fable 5 | Anthropic | 1550 | 76.5 | $10 | $50 | $1.00 | $18 | 1M | Jun 9, 2026 |
| Gemini 4 Argon | 1550 | – | $2.00 | $10 | $0.10 | $3.60 | 1.0M | Oct 13, 2026 | |
| Claude Opus 4.6 | Anthropic | 1546 | – | $5.00 | $25 | $0.50 | $9.00 | 1M | Feb 4, 2026 |
| Claude Opus 4.7 | Anthropic | 1546 | 73.6 | $5.00 | $25 | $0.50 | $9.00 | 1M | Apr 16, 2026 |
| GPT-6 Astra | OpenAI | 1542 | 76.9 | $10 | $50 | $1.00 | $18 | 1.1M | Sep 4, 2026 |
| GPT-6.1 Sol | OpenAI | 1541 | – | $2.00 | $10 | $0.10 | $3.60 | 1.1M | Sep 29, 2026 |
| Kimi K3 | Moonshot AI | 1540 | 72.0 | $0.19 | $15 | $0.18 | $3.15 | 1.0M | Jul 16, 2026 |
| Claude Fable 5.1 | Anthropic | 1537 | 80.7 | $10 | $50 | $0.25 | $18 | 1M | Sep 1, 2026 |
| Claude Opus 5 | Anthropic | 1534 | 78.0 | $5.00 | $25 | $0.50 | $9.00 | 1M | Jul 24, 2026 |
| GPT-5.6 Sol | OpenAI | 1533 | 78.3 | $4.00 | $20 | $0.40 | $7.20 | 1.1M | Jul 9, 2026 |
| Gemini 3.8 Flash | 1530 | 76.3 | $0.82 | $4.13 | $0.08 | $1.49 | 1.0M | Sep 2, 2026 | |
| Claude Opus 4.8 | Anthropic | 1529 | 74.3 | $5.00 | $25 | $0.50 | $9.00 | 1M | May 27, 2026 |
| Claude Sonnet 5.5 | Anthropic | 1529 | – | $2.00 | $10 | $0.10 | $3.60 | 1M | Sep 28, 2026 |
| Claude Sonnet 4.6 | Anthropic | 1528 | 63.0 | $3.00 | $15 | $0.30 | $5.40 | 1M | Feb 17, 2026 |
| Qwen3.8 Max (0902) | Qwen | 1525 | 76.2 | $2.00 | $6.00 | $0.25 | $2.80 | 1M | Sep 3, 2026 |
| Qwen3.7 Max | Qwen | 1523 | 66.0 | $1.48 | $4.42 | $0.29 | $2.06 | 1M | May 21, 2026 |
| DeepSeek V4.1 Flash | DeepSeek | 1523 | – | $0.04 | $0.50 | $0.01 | $0.14 | 1.0M | Sep 10, 2026 |
| Claude Opus 4.5 | Anthropic | 1522 | – | $5.00 | $25 | $0.50 | $9.00 | 200K | Nov 24, 2025 |
| Gemini 3.1 Pro Preview | 1521 | 68.8 | $2.00 | $12 | $0.20 | $4.00 | 1.0M | Feb 19, 2026 | |
| Claude Sonnet 5 | Anthropic | 1520 | 71.5 | $2.00 | $10 | $0.20 | $3.60 | 1M | Jun 30, 2026 |
| Gemini 3.7 Flash | 1520 | 76.1 | $0.75 | $3.75 | $0.07 | $1.35 | 1.0M | Aug 13, 2026 | |
| GPT-6 Sol | OpenAI | 1519 | – | $2.00 | $10 | $0.20 | $3.60 | 1.1M | Sep 22, 2026 |
| Gemini 3.6 Flash | 1518 | 69.2 | $0.75 | $3.75 | $0.07 | $1.35 | 1.0M | Jul 21, 2026 | |
| GPT-5.6 Terra | OpenAI | 1517 | 76.7 | $2.00 | $12 | $0.20 | $4.00 | 1.1M | Jul 9, 2026 |
| Kimi K2.6 | Moonshot AI | 1516 | 61.8 | $0.65 | $3.41 | $0.15 | $1.20 | 262K | Apr 20, 2026 |
| GPT-5.2 Chat | OpenAI | 1515 | – | $1.75 | $14 | $0.17 | $4.20 | 128K | Dec 10, 2025 |
| GPT-5.4 | OpenAI | 1511 | 71.1 | $2.50 | $15 | $0.25 | $5.00 | 1.1M | Mar 5, 2026 |
| GPT-5.5 | OpenAI | 1510 | 74.9 | $5.00 | $30 | $0.50 | $10 | 1.1M | Apr 24, 2026 |
| Gemini 3.5 Flash | 1508 | 70.1 | $1.50 | $9.00 | $0.15 | $3.00 | 1.0M | May 19, 2026 | |
| GPT-6 Luna | OpenAI | 1506 | – | $0.10 | $0.50 | $0.01 | $0.18 | 1.1M | Sep 22, 2026 |
| Gemini 3 Flash Preview | 1505 | – | $0.50 | $3.00 | $0.05 | $1.00 | 1.0M | Dec 17, 2025 | |
| Kimi K2.5 | Moonshot AI | 1504 | 46.8 | $0.50 | $2.50 | $0.15 | $0.90 | 262K | Jan 27, 2026 |
| Gemma 4 31B | 1502 | 43.4 | $0.20 | $0.40 | – | $0.24 | 262K | Apr 2, 2026 | |
| Gemini 3.5 Flash Lite | 1502 | 49.3 | $0.30 | $2.50 | $0.03 | $0.74 | 1.0M | Jul 21, 2026 | |
| GPT-5.6 Luna | OpenAI | 1502 | 71.4 | $0.20 | $1.20 | $0.02 | $0.40 | 1.1M | Jul 9, 2026 |
| DeepSeek V4 Pro 0423 | DeepSeek | 1500 | 59.4 | $0.95 | $1.90 | $0.08 | $1.14 | 1.0M | Apr 24, 2026 |
| DeepSeek V4 Pro 0813 | DeepSeek | 1500 | 68.8 | $0.66 | $1.98 | $0.02 | $0.92 | 1.0M | Aug 12, 2026 |
| Qwen3.7 Plus | Qwen | 1500 | 55.9 | $0.32 | $1.28 | $0.06 | $0.51 | 1M | Jun 3, 2026 |
| Qwen3.8 27B | Qwen | 1498 | 68.1 | $0.15 | $1.88 | $0.04 | $0.49 | 1M | Aug 14, 2026 |
| Qwen3.6 Plus | Qwen | 1496 | 54.5 | $0.33 | $1.95 | – | $0.65 | 1M | Apr 2, 2026 |
| GPT-5.4 Mini | OpenAI | 1495 | 56.1 | $0.75 | $4.50 | $0.07 | $1.50 | 400K | Mar 17, 2026 |
| Qwen3.5 397B A17B | Qwen | 1491 | 48.2 | $0.50 | $3.60 | $0.30 | $1.12 | 262K | Feb 16, 2026 |
| Claude Haiku 5.5 | Anthropic | 1489 | – | $0.10 | $0.50 | $0.01 | $0.18 | 1M | Oct 7, 2026 |
| DeepSeek V4 Flash 0423 | DeepSeek | 1484 | 56.2 | $0.09 | $0.18 | $0.02 | $0.11 | 1.0M | Apr 24, 2026 |
| DeepSeek V4 Flash 0731 | DeepSeek | 1484 | 69.1 | $0.12 | $0.24 | $0.02 | $0.14 | 1.0M | Jul 31, 2026 |
| Gemma 4 26B A4B | 1482 | 39.3 | $0.09 | $0.30 | $0.05 | $0.13 | 262K | Apr 3, 2026 | |
| GPT-5.2 | OpenAI | 1482 | – | $1.75 | $14 | $0.17 | $4.20 | 400K | Dec 10, 2025 |
| GPT-5.1 | OpenAI | 1476 | 49.4 | $1.25 | $10 | $0.13 | $3.00 | 400K | Nov 13, 2025 |
| DeepSeek V3.2 | DeepSeek | 1469 | 44.2 | $0.26 | $0.38 | $0.13 | $0.28 | 164K | Dec 1, 2025 |
| GPT-5.4 Nano | OpenAI | 1460 | 56.1 | $0.20 | $1.25 | $0.02 | $0.41 | 400K | Mar 17, 2026 |
| Qwen3.5-122B-A10B | Qwen | 1458 | 45.7 | $0.26 | $2.08 | – | $0.62 | 262K | Feb 25, 2026 |
| Gemini 3.1 Flash Lite | 1458 | – | $0.25 | $1.50 | $0.03 | $0.50 | 1.0M | May 7, 2026 | |
| Gemini 3.1 Flash Lite Preview | 1458 | 34.7 | $0.25 | $1.50 | $0.03 | $0.50 | 1.0M | Mar 3, 2026 | |
| Qwen3.5-27B | Qwen | 1451 | – | $0.26 | $2.60 | – | $0.73 | 262K | Feb 25, 2026 |
| Qwen3.5-Flash | Qwen | 1435 | – | $0.07 | $0.26 | – | $0.10 | 1M | Feb 25, 2026 |
| Qwen3.5-35B-A3B | Qwen | 1435 | 37.0 | $0.14 | $1.00 | $0.05 | $0.31 | 262K | Feb 25, 2026 |
| DeepSeek V4 Flash Vision Exp | DeepSeek | – | 65.0 | $0.22 | $0.65 | $0.01 | $0.30 | 1.0M | Aug 21, 2026 |
| DeepSeek V4.1 Flash 1017New | DeepSeek | – | – | $0.15 | $0.60 | $0.00 | $0.24 | 1.0M | Oct 17, 2026 |
| Gemini 3.1 Pro Preview Custom Tools | – | – | $2.00 | $12 | $0.20 | $4.00 | 1.0M | Feb 25, 2026 | |
| GPT Chat Latest | OpenAI | – | – | $5.00 | $30 | $0.50 | $10 | 400K | May 5, 2026 |
| GPT-5.1-Codex | OpenAI | – | – | $1.25 | $10 | $0.13 | $3.00 | 400K | Nov 13, 2025 |
| GPT-5.1-Codex-Max | OpenAI | – | – | $1.25 | $10 | $0.13 | $3.00 | 400K | Dec 4, 2025 |
| GPT-5.1-Codex-Mini | OpenAI | – | – | $0.25 | $2.00 | $0.03 | $0.60 | 400K | Nov 13, 2025 |
| GPT-5.2 Pro | OpenAI | – | – | $21 | $168 | – | $50.4 | 400K | Dec 10, 2025 |
| GPT-5.2-Codex | OpenAI | – | – | $1.75 | $14 | $0.17 | $4.20 | 400K | Jan 14, 2026 |
| GPT-5.3-Codex | OpenAI | – | – | $1.75 | $14 | $0.17 | $4.20 | 400K | Feb 24, 2026 |
| GPT-5.4 Pro | OpenAI | – | – | $30 | $180 | – | $60 | 1.1M | Mar 5, 2026 |
| GPT-5.5 Pro | OpenAI | – | – | $30 | $180 | – | $60 | 1.1M | Apr 24, 2026 |
| GPT-5.6 Luna Pro | OpenAI | – | – | $0.20 | $1.20 | $0.02 | $0.40 | 1.1M | Jul 9, 2026 |
| GPT-5.6 Sol Pro | OpenAI | – | – | $4.00 | $20 | $0.40 | $7.20 | 1.1M | Jul 9, 2026 |
| GPT-5.6 Terra Pro | OpenAI | – | – | $2.00 | $12 | $0.20 | $4.00 | 1.1M | Jul 9, 2026 |
| GPT-6 Astra Pro | OpenAI | – | – | $10 | $50 | $1.00 | $18 | 1.1M | Sep 4, 2026 |
| GPT-6 Luna Pro | OpenAI | – | – | $0.10 | $0.50 | $0.01 | $0.18 | 1.1M | Sep 22, 2026 |
| GPT-6 Sol Pro | OpenAI | – | – | $2.00 | $10 | $0.20 | $3.60 | 1.1M | Sep 22, 2026 |
| GPT-6.1 Sol Pro | OpenAI | – | – | $2.00 | $10 | $0.10 | $3.60 | 1.1M | Sep 29, 2026 |
| GPT-6.1 TerraNew | OpenAI | – | – | $1.00 | $5.00 | $0.05 | $1.80 | 1.1M | Oct 26, 2026 |
| GPT-6.1 Terra ProNew | OpenAI | – | – | $1.00 | $5.00 | $0.05 | $1.80 | 1.1M | Oct 26, 2026 |
| Kimi K2 Thinking | Moonshot AI | – | 21.0 | $0.60 | $2.50 | $0.15 | $0.98 | 262K | Nov 6, 2025 |
| Kimi K2.7 Code | Moonshot AI | – | 60.8 | $0.60 | $3.50 | $0.15 | $1.18 | 262K | Jun 12, 2026 |
| Kimi K3.1 ProNew | Moonshot AI | – | – | $2.00 | $10 | $0.20 | $3.60 | 1.0M | Oct 23, 2026 |
| Qwen3 Coder Next | Qwen | – | 36.2 | $0.12 | $0.80 | $0.07 | $0.26 | 262K | Feb 4, 2026 |
| Qwen3.5 Plus 2026-02-15 | Qwen | – | – | $0.26 | $1.56 | – | $0.52 | 1M | Feb 16, 2026 |
| Qwen3.5 Plus 2026-04-20 | Qwen | – | – | $0.30 | $1.80 | – | $0.60 | 1M | Apr 27, 2026 |
| Qwen3.5-9B | Qwen | – | 28.7 | $0.17 | $0.25 | – | $0.19 | 262K | Mar 10, 2026 |
| Qwen3.6 27B | Qwen | – | 53.7 | $0.32 | $3.20 | – | $0.90 | 262K | Apr 27, 2026 |
| Qwen3.6 35B A3B | Qwen | – | 41.9 | $0.10 | $0.90 | $0.03 | $0.26 | 262K | Apr 27, 2026 |
| Qwen3.6 Flash | Qwen | – | – | $0.19 | $1.13 | – | $0.38 | 1M | Apr 27, 2026 |
| Qwen3.7 Flash | Qwen | – | – | $0.03 | $0.13 | $0.01 | $0.05 | 1M | Jul 27, 2026 |
| Qwen3.8 2.4T A95B | Qwen | – | 71.9 | $2.00 | $6.00 | $0.25 | $2.80 | 1.0M | Aug 12, 2026 |
| Qwen3.8 Flash | Qwen | – | – | $0.15 | $0.47 | $0.02 | $0.21 | 1M | Aug 26, 2026 |
| Qwen3.8 Max Prime | Qwen | – | – | $4.00 | $12 | $0.50 | $5.60 | 1M | Sep 23, 2026 |
| Qwen3.8 Omni Flash | Qwen | – | – | $0.15 | $0.47 | $0.02 | $0.21 | 1M | Sep 21, 2026 |
How this works
Three kinds of scores
Arena (Human votes) is a blind preference test: people ask their own question, get two anonymous answers and vote for the better one. The result is an Elo rating, so only the gap between two models matters. Each rating carries an uncertainty margin (shown as ± on hover): two models whose margins overlap are tied. Ratings are style-controlled, which removes the advantage of longer or more formatted answers. When Arena lists a model under several settings, the default setting is used when it exists, otherwise the best-rated one; the entry used shows on hover of the score. For agents, Arena publishes a ranking rather than a rating, so only the rank is used. It reflects everyday use and picks up new models within days.
Artificial Analysis (Benchmark score) runs automated benchmarks: coding tasks, maths, science questions, agent work in a terminal. Scores are out of 100 and measure raw capability on hard problems. Domain indices for brand-new models can take a few weeks to appear.
Vals AI (Terminal test) measures how often a model completes Terminal-Bench 4.0 tasks with Mini-SWE-agent, the same agent harness for every model. Only the Agents tab uses it. Results average three runs. Reasoning settings can differ between models. Its success rates are only ever compared with other Vals AI results, never with a benchmark index. Vals AI also reports a cost per test, shown as "Per tested task" with the reasoning effort it was run at.
Prices and the model catalog come from OpenRouter. Blended price is 4 input tokens for 1 output token, the same profile as the monthly estimate.
Which scores each tab uses
| Tab | Scores |
|---|---|
| WritingDrafting, rewriting, answering, brainstorming | Human votes (Arena creative writing, style-controlled Elo) |
| CodingWriting and reviewing code in a chat or editor | Human votes (Arena coding, style-controlled Elo); Benchmark score (Artificial Analysis Coding Index) |
| AgentsLong tasks with tools, delegated to an agent | Benchmark score (Artificial Analysis Agentic Index); Human votes (Arena agent leaderboard rank); Terminal test (Vals Terminal-Bench 4.0 success rate (%)) |
| ReasoningAnalysis and hard problems | Benchmark score (Artificial Analysis Intelligence Index); Human votes (Arena hard prompts, style-controlled Elo) |
Coding or Agents?
Coding is you and the model, turn by turn: you ask in a chat or an editor, it answers, you read the code and decide what to keep. A good first answer is what counts, and that is what human votes measure. Agents is the model working on its own: it reads files, runs commands, calls tools, checks its results and keeps going until the job is done, as in Claude Code or an automated pipeline. What counts there is finishing long tasks without going off track, which is what the benchmark score and the terminal test measure; human votes rank agents too. Rule of thumb: if you will read every answer, look at Coding; if you will only read the final result, look at Agents.
Capability levels
Models are not ranked by score. Each one gets a level per task, from its distance to the best current model on each signal, among the providers you keep enabled: Top (as good as the best), Strong (close to the best), Capable (fine for routine work), Basic (noticeably weaker). Top is within 25 Elo points, 10% of an index or 5 places on the agent ranking. Strong is within 60 points, 25% or 12 places. Capable is within 120 points, 45% or 25 places. Each available signal gets its own level. The final level is the median of the available signal levels. When there are two middle levels, we use the lower one. The terminal test uses the same relative thresholds as an index: 90%, 75% and 55% of the best success rate. Models within each level are sorted by price. The best fit also depends on your workflow.
Best performance, best value, routine work
Best performance is the Top model closest to the best across the tab's signals. Each gap is measured in units of the Top threshold, so Elo points, index shares and ranking places add up; missing signals receive a gap of 1.0, equivalent to the Top threshold, so missing evidence is not treated as a perfect score. The smallest average wins, the cheaper model on a tie. Best value is the lowest-priced Top model. On the Agents tab it is the Top model with the lowest cost per tested task reported by Vals AI, a benchmark measurement rather than your usage; a model without a terminal test result cannot be Best value there. Terminal-Bench tasks are hard, multi-step jobs, so that cost describes that kind of work, not small everyday tasks. Routine work is the lowest token price among models rated Capable or better, on every tab. Scores are taken at the highest setting each source publishes; the setting behind each score shows in the model's details. Actual costs depend on token usage, including reasoning tokens. Monthly estimates assume 4,000 input and 1,000 output tokens per call, 300 calls a day, over 30 days. On the Agents tab, Best performance and Best value are priced by the tested task instead when the model has a Vals AI cost: 10 tasks a day, over 30 days.
Lineups, freshness, raw data
Small, Standard, Large and Frontier mark where a model sits in its provider's lineup, from the cheapest option to the flagship; they say nothing about capability. Current models are those released in the last 12 months, plus each provider's three highest Intelligence Index scores whatever their age; older ones stay in the raw data behind "Show older models". Capability by provider ranks each provider's models on one source at a time, as published; the bar runs from the lowest to the highest score among the models shown. Data is refreshed twice a day; the Arena leaderboard used here comes from Arena's public dataset, which can list fewer models than the live site, and was published on Oct 28, 2026. Terminal test results were published on Oct 29, 2026. The raw data table shows the sources' numbers as published, without any of the above applied.