Composite-Bench
Agents solve generated browser tasks in planning, rule extraction, deduplication, routing, and information gathering. Success requires matching a certified optimal solution. A separate number-partitioning task measures solution quality rather than contributing to the success rate.
The public package has 30 development instances and 120 hidden test instances across six task types. The success aggregate covers 100 tasks and excludes number partitioning. Instance version 2.0.0 and compute protocol v1.1 use separate version numbers. All stopped or budget-exhausted attempts count.
- Made by
- Composite
- Released
- July 2026
- Kind
- Web browsing
- Level
- Expert
- Scored
- Checked automatically
- Answers
- Agent task
- Tasks
- Public, plus a private test set
- Measures
- Capability
- Subject
- Mathematics
- Language
- English
- Interaction
- Agentic
- Environment
- Browser
- Items from
- Generated
- Refresh
- New versions replace it
- Run by
- A host
- In views
- Computer use, Agents
- Strongest model tested
- Claude Fable 5 #12 on the capability index
- Last model added
- July 2026
- Scores refreshed
- November 20th, 2026
Claude Fable 5 leads with 74% among models released by July 2026
Best score per model from the benchmark's own leaderboard, read on November 20th, 2026. Where a model ran with several settings or harnesses, we keep its best. Every run is in the next chart.
The newest model with a score here came out in July 2026, and nobody has tested a newer one. Among the 20 highest-ranked models on the capability index, these have no score yet:
- GPT-6.2 Astra
- Claude Fable 5.2
-
Gemini 4 Carbon
- GPT-6 Astra
- and 14 more
- 1 Claude Fable 5 Maximum reasoning effort 74%
- 2 Claude Opus 4.8 Maximum reasoning effort 73%
- 3 GLM-5.2 Maximum reasoning effort 63%
- 4 Kimi K3 Maximum reasoning effort 31%
- 5
Gemini 3.5 Flash Maximum reasoning effort 27%
- 6
MiniMax-M3 Maximum reasoning effort 26%
- 7 GPT-5.6 Sol Maximum reasoning effort 18%
- 8 Grok 4.5 Maximum reasoning effort 9.00%
Source: Official Composite-Bench leaderboard
| # | Model | Score ± | Setting | Date | Source |
|---|---|---|---|---|---|
| 1 | Claude Fable 5 | 74% 65 to 82 | Maximum reasoning effort | Jun 2026 | Official leaderboard |
| 2 | Claude Opus 4.8 | 73% 64 to 81 | Maximum reasoning effort | May 2026 | Official leaderboard |
| 3 | GLM-5.2 | 63% 53 to 72 | Maximum reasoning effort | Jun 2026 | Official leaderboard |
| 4 | Kimi K3 | 31% 23 to 41 | Maximum reasoning effort | Jul 2026 | Official leaderboard |
| 5 | | 27% 19 to 36 | Maximum reasoning effort | May 2026 | Official leaderboard |
| 6 | | 26% 18 to 35 | Maximum reasoning effort | Jun 2026 | Official leaderboard |
| 7 | GPT-5.6 Sol | 18% 12 to 27 | Maximum reasoning effort | Jul 2026 | Official leaderboard |
| 8 | Grok 4.5 | 9.00% 5 to 16 | Maximum reasoning effort | Jul 2026 | Official leaderboard |
Rankings by part
We read 5 parts from the benchmark’s own leaderboard, each ranking models on one part of the test, such as steward_7d, rulebook_7d and dedup. Claude Fable 5 leads 2 of them, and 2 other models lead the rest. Arrows show how far a model moved from its overall rank among the models we track.
steward_7d
8 models we track have a score here.
- 1 Claude Fable 5 65%
- 2 Claude Opus 4.8 55%
- 3 GLM-5.2 20%
- 4
Gemini 3.5 Flash ↑1 0%
- 5
MiniMax-M3 ↑1 0%
- 6 Grok 4.5 ↑2 0%
- 7 GPT-5.6 Sol 0%
- 8 Kimi K3 ↓4 0%
rulebook_7d
8 models we track have a score here.
- 1 Claude Opus 4.8 ↑1 65%
- 2 GLM-5.2 ↑1 55%
- 3 Claude Fable 5 ↓2 50%
- 4
Gemini 3.5 Flash ↑1 5.00%
- 5
MiniMax-M3 ↑1 5.00%
- 6 Kimi K3 ↓2 5.00%
- 7 Grok 4.5 ↑1 0%
- 8 GPT-5.6 Sol ↓1 0%
dedup
8 models we track have a score here.
- 1 Claude Fable 5 100%
- 2 Claude Opus 4.8 85%
- 3 GLM-5.2 85%
- 4
MiniMax-M3 ↑2 65%
- 5
Gemini 3.5 Flash 60%
- 6 Kimi K3 ↓2 40%
- 7 Grok 4.5 ↑1 35%
- 8 GPT-5.6 Sol ↓1 15%
orienteering
8 models we track have a score here.
- 1 GLM-5.2 ↑2 70%
- 2 Claude Opus 4.8 65%
- 3 Claude Fable 5 ↓2 65%
- 4 Kimi K3 45%
- 5
MiniMax-M3 ↑1 35%
- 6 GPT-5.6 Sol ↑1 15%
- 7
Gemini 3.5 Flash ↓2 10%
- 8 Grok 4.5 5.00%
gauntlet_7g
8 models we track have a score here.
- 1 Claude Opus 4.8 ↑1 95%
- 2 Claude Fable 5 ↓1 90%
- 3 GLM-5.2 85%
- 4 Kimi K3 65%
- 5
Gemini 3.5 Flash 60%
- 6 GPT-5.6 Sol ↑1 60%
- 7
MiniMax-M3 ↓1 25%
- 8 Grok 4.5 5.00%
The best score, 74%, costs $1,166 for the full run
Runs on the cost frontier score higher than any run that's cheaper. The costs are the leaderboard's own estimates, so compare them only within this benchmark.
Source: Official Composite-Bench leaderboard
| Model | Effort | Harness | Score | Cost of the full run | Output tokens |
|---|---|---|---|---|---|
| Claude Fable 5 | Maximum reasoning effort | 74% | $1,166 | ||
| Claude Opus 4.8 | Maximum reasoning effort | 73% | $605 | ||
| GLM-5.2 | Maximum reasoning effort | 63% | $107 | ||
| Kimi K3 | Maximum reasoning effort | 31% | $187 | ||
| Maximum reasoning effort | 27% | $286 | |||
| Maximum reasoning effort | 26% | $40 | |||
| GPT-5.6 Sol | Maximum reasoning effort | 18% | $671 | ||
| Grok 4.5 | Maximum reasoning effort | 9.00% | $196 |
Number one over time
Lab claims count from the day the card came out, independent runs from the model’s release.
Source: Official Composite-Bench leaderboard
Scores over time
All the scores we have, one dot each, so a model tested at several settings appears more than once.
Source: Official Composite-Bench leaderboard
No lab model card reports it yet
We read the headline results of 540 launch posts from 57 labs, and none of them list this benchmark.
More in Computer and browser use
- BrowseComp In use Hard-to-find facts that need persistent web browsing. Short verifiable answers graded for exact correctness. Claude Fable 5 93.3%
- ScreenSpot-Pro In use GUI grounding on high-resolution professional-software screenshots. The model must locate the target UI element. GPT-6 Astra 92.7%
- AndroidWorld In use Agents complete 116 tasks across 20 real Android apps in an emulator. Each task checks the device's system state to judge success. GPT-6 Astra 100%
- OSWorld 2.1 In use Agents complete 108 long-horizon workflows across 31 self-hosted websites and desktop apps, averaging 318 tool calls. Scored by binary completion and checkpoint partial credit. Claude Opus 5 44.3%
- WebArena-Verified In use Agents complete 812 audited WebArena tasks on self-hosted websites. Deterministic evaluators check the agent's answer and captured network traces.
Muse Spark 1.1 69%
- OSWorld 2.1 (offline subset) In use Agents complete the 82 OSWorld 2.1 tasks on the official offline list, with no internet access. Scored by checkpoint partial credit and strict pass rate. Claude Fable 5.2 89.3%
Data audit
All 10 checks pass. Every benchmark
Catalog record 2 pass
- Pass Task access Public tasks, plus a private test set.
- Pass Confidence Checked against a primary source.
Scores 4 pass 1 note
- Pass Independent runs Scores from the official leaderboard (8 entries).
- Pass Models scored 8 models we track.
- Pass Newest score The newest is from July 2026, for Kimi K3.
- Note Frontier models None of the top 10 has a score. The strongest tested is Claude Fable 5 (#12).
- Pass Headroom The best is 74%, under 85%, so it still separates models.
Leaderboard import 3 pass
- Pass Importer We read the board on research.composite.com: 8 rows for models we track at the last read.
- Pass Last read November 20th, 2026.
- Pass Read worked The last read worked.
Consistency 1 pass 1 note
- Note Lab claims No model has both a lab number and an independent run.
- Pass Known issues No published audit found problems.