Key findings
- Anthropic models hold the top four places: Claude Opus 5.5 and Claude Fable 5.1, each on Claude Code and on OpenCode.
- The best OpenAI pairing is GPT-6 Astra on Codex in fifth place at 1235, at about half the cost per task of the leader.
- GPT-6.1 Sol on OpenCode places seventh at 1196 for $0.51 a task, about one fifth of what the leader costs.
- Kimi K3 on OpenCode is the highest-placed Chinese model, in 13th place at 1147.
The overall top ten
The overall board compares every pairing on every task it ran. Quality is a score fitted from wins and losses on the same task, on an Elo-like scale where 1000 is the reference. Time is the average length of one run and cost is the average model spend per task. The three are ranked separately and never combined.
| Place | Model | Harness | Quality | Time per run | Cost per task |
|---|---|---|---|---|---|
| 1 | Claude Opus 5.5 | Claude Code | 1335 | 746.5s | $2.60 |
| 2 | Claude Fable 5.1 | Claude Code | 1320 | 874.7s | $3.66 |
| 3 | Claude Fable 5.1 | OpenCode | 1318 | 808.5s | $2.90 |
| 4 | Claude Opus 5.5 | OpenCode | 1298 | 1065.0s | $3.17 |
| 5 | GPT-6 Astra | Codex | 1235 | 722.8s | $1.28 |
| 6 | GPT-6 Astra | OpenCode | 1208 | 965.1s | $1.78 |
| 7 | GPT-6.1 Sol | OpenCode | 1196 | 582.2s | $0.51 |
| 8 | GPT-6 Sol | OpenCode | 1190 | 544.0s | $0.50 |
| 9 | GPT-6.1 Sol | Codex | 1187 | 500.7s | $0.47 |
| 10 | GPT-5.6 Sol | Codex | 1174 | 585.5s | $1.02 |
The first two places are 15 points apart. On this scale a 15-point gap means the higher pairing is expected to win about 52% of head-to-head comparisons, so the top two are close to even. The methodology explains how the score is fitted.
How far apart the pairings are
The 49 pairings cover a wide range on all three measures:
- Quality runs from 1335 down to 641, a gap of 694 points. The median pairing scores 999.
- Cost per task runs from $0.049 (Qwen3.8 Flash on OpenCode) to $3.66 (Claude Fable 5.1 on Claude Code), a factor of about 74. The median is $0.51.
- Time per run runs from 387.6 seconds (Claude Haiku 4.5 on Claude Code) to 1800.6 seconds (Qwen3.8 Max on Codex), a factor of about 4.6.
Higher quality usually costs more, but not always. Claude Opus 5.5 on Claude Code is both cheaper and faster than Claude Fable 5.1 on Claude Code while scoring higher. The quality and cost analysis lists the pairings no other pairing beats on both measures.
Leaders by field
Each field has its own board, worked out only from that field's tasks. Claude Fable 5.1 leads eight of the twelve fields and Claude Opus 5.5 leads the other four.
| Field | Tasks | First place | Quality |
|---|---|---|---|
| Presentations | 54 | Claude Fable 5.1 · OpenCode | 1446 |
| Documents | 28 | Claude Fable 5.1 · Claude Code | 1336 |
| Visual design | 6 | Claude Fable 5.1 · OpenCode | 1413 |
| Data analysis | 20 | Claude Fable 5.1 · Claude Code | 1413 |
| Website | 24 | Claude Opus 5.5 · Claude Code | 1307 |
| 3D | 18 | Claude Opus 5.5 · Claude Code | 1327 |
| Finance | 26 | Claude Fable 5.1 · OpenCode | 1303 |
| Medical | 54 | Claude Opus 5.5 · Claude Code | 1356 |
| Legal | 93 | Claude Fable 5.1 · Claude Code | 1402 |
| Consulting | 56 | Claude Fable 5.1 · Claude Code | 1370 |
| Education | 21 | Claude Opus 5.5 · Claude Code | 1353 |
| Audio & video | 6 | Claude Fable 5.1 · OpenCode | 1467 |
Claude Fable 5.1 leads more fields while Claude Opus 5.5 leads overall. Two things explain this. Every board is fitted on its own, so field places do not add up to the overall place. Claude Fable 5.1 on Claude Code has also run 231 of the 406 tasks so far, against 402 for Claude Opus 5.5 on Claude Code, so it has fewer comparisons behind its score.
OpenAI and Chinese models
OpenAI's best pairing is GPT-6 Astra on Codex at 1235, 100 points behind the leader. That gap means the leader is expected to win about 64% of comparisons with it. GPT-6 Astra costs $1.28 a task on Codex, about half of Claude Opus 5.5 on Claude Code. The head-to-head comparison goes through the two field by field.
The GPT-6 Sol family offers most of that quality for much less money. GPT-6.1 Sol on OpenCode scores 1196 for $0.51 a task, and on Codex it scores 1187 for $0.47 with the fastest runs in the top ten at 500.7 seconds.
Kimi K3 on OpenCode is the highest-placed Chinese model, 13th at 1147. DeepSeek Flash follows in 19th place at 1039 for $0.18 a task. The article on Chinese models covers the rest.
How to read these numbers
- Rankings are for pairings. The same model on another harness is a separate entry and can place quite differently.
- Quality scores mean something only within one board. To compare fields, compare places.
- Cost is the model spend at each vendor's official prices. It leaves out subscriptions, grading and any infrastructure.
- Pairings have run different numbers of tasks so far. Fewer tasks means the score leans more on its reference and can still move.
The live numbers are on the leaderboard, and every model and harness has its own page under models and harnesses.
Questions
Which AI agent is best for real work right now
On the November 2026 Work Arena leaderboard, Claude Opus 5.5 running in Claude Code has the highest overall quality score, 1335 across 402 tasks. Claude Fable 5.1 on Claude Code is close behind at 1320 and leads eight of the twelve fields.
What is the best value AI agent
GPT-6.1 Sol is the strongest choice at around $0.50 a task: 1196 on OpenCode and 1187 on Codex. DeepSeek Flash on Codex scores 1031 for $0.11, and is the cheapest pairing above the median.
How often is the Work Arena leaderboard updated
The leaderboard is recalculated as new runs finish and are graded. The figures in this article come from the update of 2 November 2026; the leaderboard pages always show the latest data.
Figures in this article come from the leaderboard data of 2 November 2026. The leaderboard shows any changes since.