Best LLM for Coding: The Arena Coding Leaderboard
Anthropic has had the top coding model on every Arena snapshot for 9 months. Two lighter models from other labs are now inside its margin of error.
The signal
- LaunchesGLM-5.3-Flash (26 Aug 2026): Z.ai's light model is the best non-Anthropic coder on this snapshot, #7.
- LaunchesGemini 3.8 Flash (2 Sept 2026) enters the coding board at #8 with 1,456 votes, so its interval is still wide.
- ArenaOn the 2 Sept 2026 coding board, 6 Anthropic models rate above the best model from any other lab. OpenAI's best, gpt-5.4-high, is #21.
claude-opus-4-6-high leads Arena coding at 1535. Anthropic has held #1 for 280 days, but Z.ai and Google sit within the interval, 19 points back.
Best coding model from each lab in the Arena coding top 50, 2 Sept 2026. Bar = Arena points above the #50 coding model (kimi-k2.5-thinking, 1474).
Arena coding is a category of the Arena text board: blind human votes on prompts that involve code. It measures which answer people prefer, not whether the code runs or passes tests. "Interval overlaps #1" means the model's 95% upper bound reaches the leader's lower bound; those ranks are not settled. Light-tier models with few votes have wide intervals. Latest snapshot on or before 9 Sept 2026. One row per lab (its best model); other variants of the same lab are omitted.
Show the data (11 rows)
| Rank | Model | Lab | Rating | 95% CI | Votes |
|---|---|---|---|---|---|
| 1 | claude-opus-4-6-high | Anthropic | 1535 | 1530–1541 | 18,800 |
| 7 | glm-5.3-flash | Z.ai | 1516 | 1499–1533 | 1,274 |
| 8 | gemini-3.8-flash-high | 1516 | 1500–1532 | 1,456 | |
| 10 | kimi-k3-max | Moonshot | 1511 | 1502–1520 | 4,719 |
| 16 | mimo-v2.5-pro | Xiaomi | 1501 | 1495–1507 | 15,813 |
| 18 | qwen3.8-max | Alibaba | 1498 | 1489–1508 | 3,867 |
| 20 | muse-spark-1.2 (xHigh) | Meta | 1496 | 1477–1516 | 935 |
| 21 | gpt-5.4-high | OpenAI | 1495 | 1490–1501 | 16,433 |
| 28 | ernie-5.1 | Baidu | 1489 | 1482–1496 | 10,299 |
| 36 | amazon-nova-experimental-chat-26-02-10 | Amazon | 1485 | 1465–1505 | 835 |
| 41 | grok-4.5 | xAI | 1483 | 1475–1490 | 7,383 |
The Mode read
Two light models launched in late August and early September: Z.ai's GLM-5.3-Flash and Google's Gemini 3.8 Flash. Both went straight into the Arena coding top 8. So which LLM is best for coding right now, and how sure can we be?
Anthropic holds the top spot with claude-opus-4-6-high and has led the coding board on every snapshot for 280 days. It holds 7 of the top 10. But the best Z.ai model is only 19 points behind, so it still wins about 47% of head-to-head votes.
A 9-month lead is long by Arena standards, and it was earned with developers. Developers are also the users who try a new model fastest. When a cheap challenger is inside the margin of error, the lead will not hold through rankings. It holds through the tools where developers already work every day.
In coding, the lead has lasted 9 months. The challenger is already inside the margin of error.
What we would back: Coding products that own the review, test and deploy loop, where the model is one part of the job.