Loading…
AI news
New models, price changes, the live leaderboard, open-weight releases, research and the rules governments publish. Everything is detected from public data and shows its numbers and its source.
Last change detected 2 h ago · 120 in the last 7 days
Live from the Arena leaderboard · checked 5 min ago
Ranked by the Arena text leaderboard, where people vote on blind, head-to-head answers. Published 2026-10-28, re-read every 30 minutes.
Variants of one model, such as different reasoning-effort settings, are merged into the highest-scoring entry. Scores carry a margin of error, so nearby ranks are effectively ties, and entries marked prelim have few votes. Arena measures chat preference, not every kind of capability. Ranking data: LMArena (opens in a new tab), licensed CC BY 4.0.
| # | Model | Arena score | Access | Price / 1M tokens | Context and size | Run it locally? |
|---|---|---|---|---|---|---|
| 1 | Gemini 4 CarbonGoogle · released n/s | 1540 ±9 | Proprietary | n/s | n/s · size not disclosed | Cloud API only. The weights are not published, so it can't run locally. |
| 2 | Gemini 4 ArgonGoogle · released Sep 30, 2026 | 1529 ±6 | Proprietary | $2 / $10 | 1M · size not disclosed | Cloud API only. The weights are not published, so it can't run locally. |
| 3 | Claude Opus 5.5Anthropic · released Sep 22, 2026 | 1516 ±7 | Proprietary | $4 / $20 | 1M · size not disclosed | Cloud API only. The weights are not published, so it can't run locally. |
| 4 | Claude Fable 5.1Anthropic · released Sep 1, 2026 | 1514 ±6 | Proprietary | $10 / $50 | 1M · size not disclosed | Cloud API only. The weights are not published, so it can't run locally. |
| 5 | Claude Opus 5Anthropic · released Jul 24, 2026 | 1508 ±5 | Proprietary | $5 / $25 | 1M · size not disclosed | Cloud API only. The weights are not published, so it can't run locally. |
| 6 | Claude Opus 4.6Anthropic · released Feb 5, 2026 | 1502 ±3 | Proprietary | $5 / $25 | 1M · size not disclosed | Cloud API only. The weights are not published, so it can't run locally. |
| 7 | Muse WonderMeta · released Oct 12, 2026 | 1499 ±7 | Proprietary | $4 / $20 | 1M · size not disclosed | Cloud API only. The weights are not published, so it can't run locally. |
| 8 | Gemini 3.8 FlashGoogle · released Sep 2, 2026 | 1498 ±4 | Proprietary | $0.83 / $4.13 | 1M · size not disclosed | Cloud API only. The weights are not published, so it can't run locally. |
| 9 | Claude Fable 5Anthropic · released Jun 9, 2026 | 1492 ±4 | Proprietary | $10 / $50 | 1M · size not disclosed | Cloud API only. The weights are not published, so it can't run locally. |
| 10 | MiMo-V2.6-ProXiaomi · released Sep 21, 2026MIT licence. | 1491 ±9 | Open weights | $0.43 / $0.87 | 1.1M · 1.02T total / 42B active (MoE) | Not practical locally. About 573 GB of 8-bit weights; the model card uses two multi-GPU nodes (16-way tensor parallel). |
| 11 | Muse Spark 1.3Meta · released Sep 2, 2026Meta says it plans open-weight Muse releases, none confirmed yet. | 1491 ±5 | Proprietary | $1.25 / $4.25 | 1M · size not disclosed | Cloud API only. The weights are not published, so it can't run locally. |
| 12 | Claude Opus 4.7Anthropic · released Apr 16, 2026 | 1490 ±4 | Proprietary | $5 / $25 | 1M · size not disclosed | Cloud API only. The weights are not published, so it can't run locally. |
| 13 | Gemini 3.7 FlashGoogle · released Aug 13, 2026 | 1489 ±5 | Proprietary | $0.75 / $3.75 | 1M · size not disclosed | Cloud API only. The weights are not published, so it can't run locally. |
| 14 | Qwen3.8 MaxAlibaba · released Sep 2, 2026A reduced open-weight Max-class checkpoint (2.4T total / 95B active, text only) also exists under a custom licence. | 1483 ±5 | Proprietary | $2 / $6 | 1M · ~2.4T | Cloud API only. The weights are not published, so it can't run locally. |
| 15 | Muse Spark 1.2Meta · released 2026 | 1483 ±6 | Proprietary | $1.25 / $4.25 | 1M · size not disclosed | Cloud API only. The weights are not published, so it can't run locally. |
| 16 | Gemini 3.5 FlashGoogle · released May 19, 2026 | 1482 ±4 | Proprietary | $1.5 / $9 | 1M · size not disclosed | Cloud API only. The weights are not published, so it can't run locally. |
| 17 | Gemini 3.1 Pro PreviewGoogle · released Feb 19, 2026 | 1480 ±3 | Proprietary | $2 / $12 | 1M · size not disclosed | Cloud API only. The weights are not published, so it can't run locally. |
| 18 | Gemini 3.6 FlashGoogle · released Jul 21, 2026 | 1480 ±4 | Proprietary | $0.75 / $3.75 | 1M · size not disclosed | Cloud API only. The weights are not published, so it can't run locally. |
| 19 | Gemini 3 ProGoogle · released Nov 18, 2025The preview endpoint was shut down on Mar 9, 2026 in favour of 3.1 Pro. | 1479 ±4 | Proprietary | n/s | n/s · size not disclosed | Cloud API only. The weights are not published, so it can't run locally. |
| 20 | Muse Spark 1.1Meta · released 2026 | 1478 ±4 | Proprietary | $1.25 / $4.25 | 1M · size not disclosed | Cloud API only. The weights are not published, so it can't run locally. |
| 21 | Kimi K3Moonshot AI · released Jul 27, 2026Licence: Kimi K3 License (sources disagree on the exact terms). | 1476 ±4 | Open weights | $0.19 / $15 | 1M · 2.8T total / 104B active (MoE) | Not practical locally. About 1.4 TB for the native 4-bit weights; Moonshot recommends 64+ accelerators. Use the API. |
| 22 | Qwen3.7 Max PreviewAlibaba · released n/s | 1475 ±10 | Proprietary | n/s | n/s · size not disclosed | Cloud API only. The weights are not published, so it can't run locally. |
| 23 | Muse SparkMeta · released Apr 8, 2026 | 1474 ±6 | Proprietary | n/s | n/s · size not disclosed | Cloud API only. The weights are not published, so it can't run locally. |
| 24 | GLM-5.3Z.ai · released Aug 28, 2026Arena lists MIT; most sources say the flagship has a bespoke, near-MIT licence (the Flash variant is plain MIT). | 1474 ±5 | Open weights | $0.07 / $1.4 | 1M · ~753B total / ~39B active (MoE) | Server only. About 743 GB native FP8: the vLLM recipe uses 8× H200. Community builds: ~465 GB (NVFP4), ~281 GB (2-bit GGUF). |
| 25 | GLM 5.3 FlashZ.ai · released Aug 26, 2026 | 1473 ±5 | Open weights | $0.02 / $0.15 | 1M · size not disclosed | Open weights; the size isn't stated. Check the model card before planning hardware. |
| 26 | Claude Sonnet 5.5Anthropic · released Sep 28, 2026 | 1473 ±7 | Proprietary | $2 / $10 | 1M · size not disclosed | Cloud API only. The weights are not published, so it can't run locally. |
| 27 | GPT-5.5OpenAI · released Apr 2026 | 1472 ±4 | Proprietary | $5 / $30 | 1.1M · size not disclosed | Cloud API only. The weights are not published, so it can't run locally. |
| 28 | Qwen3.5 Max PreviewAlibaba · released n/s | 1470 ±5 | Proprietary | n/s | n/s · size not disclosed | Cloud API only. The weights are not published, so it can't run locally. |
| 29 | GLM-5.2Z.ai · released Jun 2026MIT licence per Arena. | 1470 ±4 | Open weights | $0.01 / $7 | 1M · size not disclosed | Large MoE, so expect a multi-GPU server. We couldn't verify a specific figure. |
| 30 | GPT-5.4OpenAI · released 2026 | 1469 ±4 | Proprietary | $2.5 / $15 | 1.1M · size not disclosed | Cloud API only. The weights are not published, so it can't run locally. |
n/s = not stated in the sources we checked. Prices are USD per million input / output tokens as listed by OpenRouter or the developer. Only 5 of the top 30 are open-weight (MiMo-V2.6-Pro, Kimi K3, GLM-5.3, GLM 5.3 Flash, GLM-5.2), and most need a server. For models you can run on a PC, see the Models view. Confirm prices and specs with the developer before you rely on them.
Detected automatically from public data; every item links to its source. How it works and every source · RSS feed · Daily briefing by email