Best LLM for Coding (October 2026)
Claude Opus 5.5 is the best LLM for coding on the LMArena WebDev leaderboard (1807 vs 1774 for GPT-6 Astra, snapshot Oct 22, 2026) and tops the official Terminal-Bench 4.0 board at 64.8%, at $4 / $20 per 1M input / output tokens. Best value: Claude Sonnet 5.5 ($2 / $10). Best open-weight: DeepSeek-V4.1-Flash-20261017 ($0.15 / $0.60). All reported from the sources below, not run by us.
Updated · Reported from the sources below, not run by us · Markdown · All models · New models this week
What is the best AI model for coding right now? Best for each job
| Pick | Model | $ per 1M in / out | Why (sourced below) |
|---|---|---|---|
| Best overall | Claude Opus 5.5 | $4 / $20 | #1 on LMArena WebDev (1807, Oct 22, 2026), 64.8% on Terminal-Bench 4.0 (official board, Oct 26, 2026), 89.9% on SWE-bench Pro (vendor-reported). |
| Best value | Claude Sonnet 5.5 | $2 / $10 | 98% of the top LMArena WebDev score (1769 vs 1807) at 50% of its output price; 61.8% on Terminal-Bench 4.0; 81.3% on SWE-bench Pro (vendor-reported). |
| Best open-weight | DeepSeek-V4.1-Flash-20261017 | $0.15 / $0.60 | Highest-ranked open-weights row on LMArena WebDev (rank 14, 1665), MIT licence. DeepSeek reports 42.3% on Terminal-Bench 4.0 for it (vendor-run). Kimi K3.1 Pro ranks higher but its weights were unpublished at the snapshot. |
| Best for terminal agents | Claude Opus 5.5 | $4 / $20 | #1 on the official Terminal-Bench 4.0 board at 64.8% (Claude Code, max), ahead of Claude Sonnet 5.5 at 61.8%. Artificial Analysis measured Kimi K3.1 Pro at 67% in its own run, a different harness and not on the official board. |
Price is what LMArena lists next to each row; the vendor’s own price is on the model page. The picks follow written rules over the tables below, so they move when the data does.
LMArena WebDev leaderboard
Independent source · snapshot · source
Crowd-voted preference on front-end web development tasks, including agentic coding workflows. Score is an Elo-style rating with a confidence margin; the page showed 151 models and 923,854 votes. Score (Elo-style) ± margin, votes. Price is the $ per 1M input / output tokens LMArena lists next to each model.
| Rank | Model | Score | Votes | $ per 1M in / out | Licence (as listed) |
|---|---|---|---|---|---|
| 1 | Claude Opus 5.5 (Max) | 1807 ± 13 | 3,140 | $4 / $20 | Proprietary |
| 2 | GPT-6 Astra (Max) | 1774 ± 10 | 6,579 | $10 / $50 | Proprietary |
| 3 | Claude Sonnet 5.5 (xHigh) | 1769 ± 12 | 3,822 | $2 / $10 | Proprietary |
| 4 | GPT-6.1 Sol (Max) | 1749 ± 12 | 3,469 | $2 / $10 | Proprietary |
| 5 | Claude Fable 5.1 (Max) | 1742 ± 9 | 8,098 | $10 / $50 | Proprietary |
| 6 | Muse Wonder Full (Max) | 1726 ± 14 | 2,253 | N/A | Proprietary |
| 7 | Claude Sonnet 5.5 (High) | 1709 ± 10 | 5,632 | $2 / $10 | Proprietary |
| 8 | Kimi K3.1 Pro (Max) | 1692 ± 14 (preliminary) | 2,451 | $2 / $10 | Proprietary (per LMArena; weights unpublished at snapshot) |
| 9 | Claude Opus 5 (Max) | 1691 ± 6 | 18,995 | $5 / $25 | Proprietary |
| 10 | GPT-6 Sol (Max) | 1681 ± 10 | 5,008 | $2 / $10 | Proprietary |
| 14 | DeepSeek-V4.1-Flash-20261017 (Max) | 1665 ± 13 | 2,500 | $0.15 / $0.60 | MIT |
| 17 | Kimi K3 (Max) | 1655 ± 6 | 20,312 | $3 / $15 | Kimi K3 license |
| 18 | Muse Wonder (Max) | 1651 ± 13 | 2,733 | $4 / $20 | Proprietary |
| 34 | MiniMax-M3.1-Flash | 1589 ± 13 (preliminary) | 2,212 | $0.15 / $0.60 | MiniMax Community License |
We show the top 10 rows plus the highest-ranked entries from our model list below that. Rows marked preliminary are labelled so by LMArena. LMArena ranks several effort settings of one model separately.
Terminal-Bench 4.0 official leaderboard
Independent source · snapshot · source
Resolution rate on the 66 Terminal-Bench 4.0 terminal tasks, five trials per task, hosted by Stanford, Harbor and the Laude Institute. Each row is a model run inside a named agent harness; we show each model's best listed setting. Resolution rate, %
| Model | Score |
|---|---|
| Claude Opus 5.5 (Claude Code, max) | 64.8% |
| Claude Sonnet 5.5 (Claude Code, max) | 61.8% |
| GPT-6 Astra (Codex, max) | 58.2% |
| GPT-6.1 Sol (Codex, max) | 58.2% |
| Claude Fable 5.1 (Claude Code, max) | 57.9% |
| GPT-6 Sol (Codex, max) | 49.4% |
| GLM-5.3 (Claude Code, max) | 41.8% |
| Grok 4.7 (Grok Build, xhigh) | 37.6% |
| GPT-6 Luna (Codex, max) | 16.4% |
Read from the leaderboard page on Oct 26, 2026 and rounded as the page displays them. Scores depend on the harness shown in brackets and on reasoning effort. Kimi K3.1 Pro and DeepSeek-V4.1-Flash-1017 were not on this board at the snapshot; their Terminal-Bench figures below are from other sources.
SWE-bench Pro (Public), as reported by the vendors
Vendor-reported, not independently reproduced · snapshot · source
Share of SWE-bench Pro Public tasks resolved, as each vendor reported it in its own system card or launch post. The registry we read lists them with a link to each vendor document. Public tasks resolved, %
| Model | Score |
|---|---|
| Claude Opus 5.5 (max) | 89.9% |
| Claude Sonnet 5.5 (max) | 81.3% |
| Claude Haiku 5.5 (max) | 64.8% |
Vendor-reported and not independently reproduced by us. We show only the models in our list that have a vendor-reported figure on the registry at the snapshot; GPT-6 Astra, Gemini 4 Argon and Kimi K3.1 Pro have none there, and we left out third-party estimates.
Artificial Analysis: Kimi K3.1 Pro launch benchmarks
Independent source · snapshot · source
Terminal-Bench 4.0 (agentic terminal use, by the Laude Institute) as run by Artificial Analysis. Terminal-Bench 4.0 pass rate, %
| Model | Score |
|---|---|
| Kimi K3.1 Pro | 67% |
| Claude Sonnet 5.5 | 64% |
| Kimi K3 | 13% |
Artificial Analysis says 67% is the highest score of any model it had run on Terminal-Bench 4.0 at publication. It lists no other models' Terminal-Bench scores in the text we could read, so this is not a full ranking.
Moonshot AI: Kimi K3.1 Pro launch post
Vendor-reported, not independently reproduced · snapshot · source
Terminal-Bench 4.0, reported by the vendor. Moonshot ran Kimi K3.1 Pro with the Claude-Code harness; for the other models it took the maximum value reported in their blog posts or on the official Terminal-Bench 4.0 leaderboard. Terminal-Bench 4.0 pass rate, %
| Model | Score |
|---|---|
| Claude Opus 5.5 | 66.4% |
| Kimi K3.1 Pro | 65.7% |
| GPT-6 Astra | 58.2% |
Vendor-reported and not independently reproduced by us. It disagrees with Artificial Analysis's run, which is why we show both.
DeepSeek-V4.1-Flash-1017 Terminal-Bench 4.0, as reported by DeepSeek
Vendor-reported, not independently reproduced · snapshot · source
Terminal-Bench 4.0 pass rate from DeepSeek's Hugging Face model card, run by DeepSeek with the Claude Code harness. Terminal-Bench 4.0 pass rate, %
| Model | Score |
|---|---|
| DeepSeek-V4.1-Flash-1017 | 42.3% |
| GLM-5.3 | 37.9% |
| DeepSeek-V4.1-Flash (Sep 10) | 31.2% |
Vendor-reported and not independently reproduced by us. Different harness and run from the official board above, so do not compare these directly with its rows.
How we picked, and what we left out
- Best overall is the top row of the LMArena WebDev leaderboard, an independent crowd-vote ranking of web and agentic coding work with a stated sample size. Best value is the highest-ranked row priced at no more than half the top row’s output price. Best open-weight is the highest-ranked row with an MIT or Apache licence. Best for terminal agents is the top row of the official Terminal-Bench 4.0 board.
- Terminal-Bench 4.0 appears in four tables because the runs differ: the official board, Artificial Analysis’s own run, Moonshot’s report and DeepSeek’s report. Harnesses and settings differ, so we do not merge them. On Kimi K3.1 Pro, Artificial Analysis (67%) and Moonshot (65.7%) disagree, so we show both.
- SWE-bench Pro numbers are vendor-reported and shown only for models with a figure on the registry we read. We left out third-party estimates, numbers we could not read from a primary source, and vendor internal benchmarks.
- October 2026 data changes daily. Check the linked leaderboards before relying on the order.
FAQ
What is the best LLM for coding in October 2026?
Claude Opus 5.5 is the best LLM for coding on the LMArena WebDev leaderboard (1807 vs 1774 for GPT-6 Astra, snapshot Oct 22, 2026) and tops the official Terminal-Bench 4.0 board at 64.8%, at $4 / $20 per 1M input / output tokens. Best value: Claude Sonnet 5.5 ($2 / $10). Best open-weight: DeepSeek-V4.1-Flash-20261017 ($0.15 / $0.60). All reported from the sources below, not run by us.
Which cheap or open-weights model is best for coding?
Best value: Claude Sonnet 5.5 at $2 / $10 per 1M input / output tokens. 98% of the top LMArena WebDev score (1769 vs 1807) at 50% of its output price; 61.8% on Terminal-Bench 4.0; 81.3% on SWE-bench Pro (vendor-reported). Best open-weight: DeepSeek-V4.1-Flash-20261017 at $0.15 / $0.60. Highest-ranked open-weights row on LMArena WebDev (rank 14, 1665), MIT licence. DeepSeek reports 42.3% on Terminal-Bench 4.0 for it (vendor-run). Kimi K3.1 Pro ranks higher but its weights were unpublished at the snapshot.
Which model is best for terminal and command-line agents?
Claude Opus 5.5. #1 on the official Terminal-Bench 4.0 board at 64.8% (Claude Code, max), ahead of Claude Sonnet 5.5 at 61.8%. Artificial Analysis measured Kimi K3.1 Pro at 67% in its own run, a different harness and not on the official board.
Are these benchmark numbers tested by you?
No. Every number is reported from the source named beside it, with that source's snapshot date. We did not run these models. Vendor-reported numbers are labelled as such and sit apart from the neutral ones.
How far ahead is Claude Opus 5.5?
Claude Opus 5.5 scores 1807 (± 13) on the 2026-10-22 LMArena WebDev snapshot; the next model scores 1774 (± 10). LMArena shows a rank spread of 1 for Claude Opus 5.5 and 2 to 3 for the runner-up.