Which model is actually ahead?
Arena scores come from blind head-to-head votes between two anonymous models. We show each score with its confidence band, so a lead the votes have not settled looks unsettled.
Short answer
Google’s gemini-4-argon-high led the Arena text board on 15 October 2026 at 1524 ±9, 15 points ahead of Meta’s muse-wonder-max at 1509 ±11, with Anthropic’s claude-opus-5.5-high third at 1508 ±7. Anthropic models held 7 of the top 12 places; OpenAI’s best, gpt-5.6-sol-xhigh, was 21st. The best-placed open-licence model, Moonshot’s kimi-k3-max, was 18th at 1487 ±5.
The leader’s score rests on 4,904 votes, Arena marks it preliminary, and its band overlaps the second model’s band. When you choose a model to code with, read the band and the vote count before the rank.
Text leaderboard 18
| Arena # | Model | Weights | Score | Score with band (1420 to 1540) | Source |
|---|---|---|---|---|---|
| 1 | gemini-4-argon-highGoogle · 4,904 votes · preliminary | Proprietary | 1524 ±9 | Arena text board, 15 Oct 2026 | |
| 2 | muse-wonder-maxMeta · 3,247 votes | Proprietary | 1509 ±11 | Arena text board, 15 Oct 2026 | |
| 3 | claude-opus-5.5-highAnthropic · 8,300 votes | Proprietary | 1508 ±7 | Arena text board, 15 Oct 2026 | |
| 4 | claude-fable-5-highAnthropic · 42,004 votes | Proprietary | 1504 ±4 | Arena text board, 15 Oct 2026 | |
| 5 | claude-opus-4-6-highAnthropic · 81,394 votes | Proprietary | 1503 ±3 | Arena text board, 15 Oct 2026 | |
| 6 | claude-fable-5.1-maxAnthropic · 14,590 votes | Proprietary | 1502 ±6 | Arena text board, 15 Oct 2026 | |
| 7 | claude-opus-4-7-highAnthropic · 65,116 votes | Proprietary | 1501 ±4 | Arena text board, 15 Oct 2026 | |
| 8 | claude-opus-4-6Anthropic · 86,071 votes | Proprietary | 1497 ±3 | Arena text board, 15 Oct 2026 | |
| 9 | gemini-3.8-flash-highGoogle · 37,454 votes · preliminary | Proprietary | 1496 ±5 | Arena text board, 15 Oct 2026 | |
| 10 | muse-spark-1.3-maxMeta · 17,561 votes | Proprietary | 1494 ±6 | Arena text board, 15 Oct 2026 | |
| 11 | claude-opus-4-7Anthropic · 66,216 votes | Proprietary | 1494 ±4 | Arena text board, 15 Oct 2026 | |
| 12 | muse-spark-1.2 (xHigh)Meta · 8,248 votes | Proprietary | 1493 ±7 | Arena text board, 15 Oct 2026 | |
| 18 | kimi-k3-maxMoonshot · Kimi K3 license · 32,795 votes | Open | 1487 ±5 | Arena text board, 15 Oct 2026 | |
| 27 | mimo-v2.6-proXiaomi · MIT · 4,075 votes | Open | 1479 ±9 | Arena text board, 15 Oct 2026 | |
| 28 | glm-5.3-maxZ.ai · MIT · 22,543 votes | Open | 1479 ±5 | Arena text board, 15 Oct 2026 | |
| 32 | glm-5.2-maxZ.ai · MIT · 49,820 votes | Open | 1476 ±4 | Arena text board, 15 Oct 2026 | |
| 38 | deepseek-v4.1-flash-maxDeepSeek · MIT · 12,793 votes | Open | 1475 ±6 | Arena text board, 15 Oct 2026 | |
| 41 | glm-5.3-flashZ.ai · MIT · 32,925 votes | Open | 1474 ±5 | Arena text board, 15 Oct 2026 |
Arena # is the model’s position among all 416 models on the 15 October 2026 board (8,834,211 votes). The first 12 rows are the board’s top 12, in order and with no gaps. None of them has an open licence, so the last six rows are the six best-placed open-licence models on the board, shown with their true positions (18 to 41); the full board is the first source below. Licence labels are as Arena lists them: Open means Arena lists an open or model-specific licence, Proprietary means it lists none. Vote counts sit under each model name.
Head to head
Pick two models. The score gap converts to how often voters would be expected to prefer one over the other, using the base-10, 400-point scale in Arena’s published rating code.
gemini-4-argon-high would be preferred in about 55% of head-to-heads.
Gap 37 points. Expected share = 1 / (1 + 10^(−gap/400)), the scale in Arena’s published rating code. The formula ignores the ± bands, so a close result is not settled.
How to read an Arena score
- Arena score
- A Bradley-Terry rating fitted to blind votes between two anonymous models. Arena counts only votes cast while the models are still anonymous. Compare a score only with other scores on the same board and view.
- ± band
- The confidence interval around the score. Since July 2025 Arena computes it with a closed-form method instead of bootstrapping. If two bands overlap, the votes have not clearly separated those models.
- 100 points
- On the 400-point scale, a 100-point gap means the higher model is expected to win about 64% of head-to-heads. A 200-point gap is about 76%.
- Preliminary
- Arena marks a score preliminary when a model tested anonymously is then released publicly, until enough fresh votes arrive. The 15 October leader carries this tag.
- Votes
- New models start with few votes and wide bands. On 15 October the leader had 4,904 votes; gemini-3.1-pro-preview, ranked 19th, had 127,649 and a band of ±3.
For vibe coding
The text board measures general chat preference, not whether code runs. Arena’s WebDev board, which scores front-end web tasks including multi-step agentic work, orders models differently: on 13 October 2026 Anthropic’s claude-opus-5.5-max led it at 1811 (2,597 votes), ahead of OpenAI’s gpt-6-astra-max at 1780 (5,746 votes), while gemini-4-argon-high was 9th at 1678 (preliminary, 1,772 votes), and the best-placed open-licence model was kimi-k3-max, 14th at 1655.
A rank says nothing about the licence, security or correctness of the code a model writes for you, so review its output either way. The prompt shelf collects prompts the editors wrote for common coding jobs; they have not yet been run against a named model.