Rank 186
Claude Opus 5.5
$1/task
AI Model Evaluation
LLM Model Benchmarks
Human preference arena leaderboard
Rank 283.5
GPT-6.1 Sol
$0.5/task
Rank 383.5
GPT-6 Astra
$2.5/task
A work-first guide to professional, agentic, coding, regulated-domain, STEM, preference, safety, and cost benchmarks. Current score cards use the BenchLM public feed; Artificial Analysis is shown separately as an independent evaluation suite.
Updated 2026-10-13
Benchmark map
All benchmark types
A compact map of the evals worth knowing before comparing Claude, ChatGPT, Gemini, open-weight models, or any new frontier release.
Professional work
Business, legal, finance, healthcare, research, documents, and human-facing output.
Current best
88.8Claude Opus 5.5
$1/task
#2 79.3 · $2.5/task · Claude Fable 5.1
#3 77.6 · $1.25/task · Claude Opus 5
Design
Design fidelity, human preference, working web applications, and presentation quality.
Current best
38.3%Claude Opus 4.6
FigmaBench
#2 23.3% · Mistral Large 3
#3 19.9% · Kimi K2.5
Security
Offensive and defensive capability, vulnerability work, secure coding, and guarded deployment.
Current best
93.3GPT-6 Astra
$2.5/task
#2 90.5 · $0.5/task · GPT-6.1 Sol
#3 86.1 · $1/task · Claude Opus 5.5
Coding
Repository repair, fresh code generation, and terminal execution.
Current best
83.6Claude Sonnet 5.5
$0.5/task
#2 82.8 · $1/task · Claude Opus 5.5
#3 77.1 · $2.5/task · Claude Fable 5.1
Connectivity
Tools, browsers, interfaces, retrieval systems, and sustained context.
Current best
88.8Claude Opus 5.5
$1/task
#2 79.3 · $2.5/task · Claude Fable 5.1
#3 77.6 · $1.25/task · Claude Opus 5
STEM
Broad intelligence, science, mathematics, and difficult held-out reasoning tasks.
Current best
93.3GPT-6 Astra
$2.5/task
#2 90.5 · $0.5/task · GPT-6.1 Sol
#3 86.1 · $1/task · Claude Opus 5.5
Alignment
Whether a model follows intent or cheats, deceives, schemes, sandbags, or games an evaluation.
Current best
86Claude Opus 5.5
$1/task
#2 83.5 · $0.5/task · GPT-6.1 Sol
#3 83.5 · $2.5/task · GPT-6 Astra
Cost
Cost per completed task, speed, retries, reliability, and operational efficiency.
Current best
100DeepSeek V4.1 Flash
$0.07/task
#2 94.8 · $0.07/task · MiMo-V2.6-Pro
#3 36.5 · $0.18/task · Step 5 Preview
Selector
Pick the work, then pick the benchmarks
Choose the outcomes that matter. The selector returns the benchmark families that should carry the most weight.
Selected work winner
Claude Opus 5.5
$1/task estimated
Fit score
86.5
Based on the selected work. Cost changes the ranking only when Cost Per Task is selected.
Choose priorities
Relevance to task
Terminal and developer-environment agents
Function calling and tool selection
Fresh coding and contest programming
Repository issue repair
Long-context retrieval and memory
Professional workflow execution
Human preference arenas
Model snapshot
Current BenchLM scores
BenchLM snapshot fetched 2026-10-13. Restricted-access and unknown-availability models are excluded; the comparison keeps overall leaders plus top public Anthropic, OpenAI, Google/Gemini, and DeepSeek coverage.
| Rank | Model | Overall | Coding | Agentic | Reasoning | Context | $/M In/Out |
|---|---|---|---|---|---|---|---|
| 1 | Claude Opus 5.5Anthropic1M / $4/$20 | 86 | 82.8 | 88.8 | 86.1 | 1M | $4 / $20 |
| 2 | GPT-6.1 SolOpenAI1.1M / $2/$10 | 83.5 | 71.1 | 74.4 | 90.5 | 1.1M | $2 / $10 |
| 3 | GPT-6 AstraOpenAI1.1M / $10/$50 | 83.5 | 75.8 | 70.7 | 93.3 | 1.1M | $10 / $50 |
| 4 | Claude Fable 5.1Anthropic1M / $10/$50 | 82.6 | 77.1 | 79.3 | 85.2 | 1M | $10 / $50 |
| 5 | Claude Sonnet 5.5Anthropic1M / $2/$10 | 82.4 | 83.6 | 68.3 | 82.7 | 1M | $2 / $10 |
| 6 | Gemini 4 ArgonGoogle1M / $2/$10 | 80.9 | 71.6 | 74.1 | 80.6 | 1M | $2 / $10 |
| 7 | Claude Fable 5Anthropic1M / $10/$50 | 80.9 | 71.4 | 74.9 | 84.3 | 1M | $10 / $50 |
| 8 | Claude Opus 5Anthropic1M / $5/$25 | 79.3 | 70.9 | 77.6 | 81 | 1M | $5 / $25 |
| 9 | GPT-6 SolOpenAI1.1M / $2/$10 | 78.1 | 66.5 | 60.2 | 73.2 | 1.1M | $2 / $10 |
| 10 | GPT-5.6 SolOpenAI1.1M / $4/$20 | 74.8 | 67.7 | 67.4 | 76.4 | 1.1M | $4 / $20 |
| 11 | Muse Wonder FullMetaNA / NA/NA | 74.1 | 73.8 | NA | 83.4 | NA | NA / NA |
| 12 | MiMo-V2.6-ProXiaomi1M / $0.44/$0.87 | 72.9 | 63.6 | 66 | 83.4 | 1M | $0.44 / $0.87 |
| 13 | Gemini 3.8 FlashGoogle1M / $0.75/$3.75 | 69.2 | 61 | 65.3 | 74.3 | 1M | $0.75 / $3.75 |
| 14 | Step 5 PreviewStepFun1M / $1/$2.7 | 68.9 | 60.6 | 63.7 | 83.4 | 1M | $1 / $2.7 |
| 15 | Gemini 3.7 FlashGoogle1M / $0.75/$3.75 | 66.6 | 57.9 | 57.9 | 81.6 | 1M | $0.75 / $3.75 |
| 16 | DeepSeek V4.1 FlashDeepSeek1M / $0.3/$1.2 | 64.6 | 61.1 | 60.5 | 83.4 | 1M | $0.3 / $1.2 |
| 17 | dots3-note PreviewDots Studio512K / $0/$0 | 63.9 | NA | 58.4 | 76.6 | 512K | $0 / $0 |
| 18 | Hy4 previewTencent1M / $0/$0 | 60.9 | 57.1 | 54.5 | NA | 1M | $0 / $0 |
| 19 | Qwen 3.6 Max (preview)Alibaba246K / $1.04/$6.24 | 54.6 | 43.3 | NA | NA | 246K | $1.04 / $6.24 |
| 20 | Hy3 PreviewTencent262K / $0/$0 | 52 | 31 | NA | NA | 262K | $0 / $0 |
FAQ
Quick answers
Claude Sonnet 5.5 vs Claude Opus 5.5 vs Claude Fable 5.1
Claude Sonnet 5.5 currently has the highest live software-engineering score shown here: 83.6. Restricted-access and unknown-availability models are excluded. Use it as the first model to inspect for coding work, then validate with repo repair, terminal-agent, and local codebase tasks.
What is the most trustworthy LLM benchmark?
There is no single most trustworthy benchmark. The best benchmark is the one that closely matches the task, uses a transparent and consistent harness, is fresh enough to reduce contamination, and reports enough metadata to compare models fairly.
Should I trust vendor-published benchmark results?
Trust them as useful claims, not final evidence. Provider system cards and launch posts often have the freshest model data, but the result should be checked against benchmark-owner leaderboards, third-party runs, and local evaluations before it drives a production decision.
Why do two pages show different scores for the same model and benchmark?
Usually because the benchmark version, subset, scaffold, tools, effort level, context budget, number of attempts, or scoring rule changed. Same benchmark name does not guarantee same experiment.
What benchmark matters most for coding agents?
Use several: audited repository-repair tasks, Terminal-Bench for shell and tool persistence, LiveCodeBench for fresh coding ability, and local repo tasks for the work that matters to your team. Treat SWE-Bench Pro cautiously while its reported task-quality problems remain unresolved.
How should token cost be evaluated?
Use cost per completed task, not price per million tokens alone. Count input, cached input, output, reasoning tokens when billed, tool costs, retries, long-context surcharges, latency, and the cost of human correction.
Inspect first
Sources
How to read the evidence
Benchmark owners
Trust the exact task they run; never treat one task as a universal rank.
AA + Vals
Use for comparable agent systems and finished professional work.
METR
Use for long-horizon autonomy, monitorability, and frontier risk.
LMArena
Use for human preference and feel, not correctness.
BenchLM
Use as broad, fresh radar; inspect the underlying evidence before deciding.
- FigmaBench structured Figma design benchmark
- Design2Code benchmark paper
- Design Arena methodology
- WebDev Arena methodology and results
- Vibe Code Bench paper and leaderboard
- PPTBench presentation benchmark
- SlidesBench and AutoPresent paper
- Artificial Analysis model leaderboard
- AA-Briefcase professional work leaderboard
- Artificial Analysis intelligence methodology
- Artificial Analysis API Reference
- BenchLM public leaderboard endpoint
- BenchLM public pricing endpoint
- Models.dev model database
- LMArena Leaderboard
- Chatbot Arena human preference paper
- MT-Bench and Chatbot Arena judge paper
- AlpacaEval length-controlled correlation notes
- LiveBench paper
- LiveCodeBench repository
- SWE-bench repository
- SWE-bench Pro by Scale AI
- OpenAI audit of SWE-Bench Pro task quality
- Terminal-Bench 2.1
- Saving SWE-Bench realistic agent evaluation
- Berkeley Function-Calling Leaderboard
- OSWorld computer-use agent paper
- GDPval real-world work benchmark
- Vals real-world professional benchmarks
- METR time-horizon and autonomous-capability research
- Personalized benchmark preference divergence
- Leaderboard Illusion benchmark audit
- LegalBench
- LegalBench-RAG
- FinanceBench
- Open FinLLM Leaderboard paper
- HealthBench by OpenAI
- CVE-Bench
- Cybench by Epoch AI
- OpenAI note on SWE-bench Verified saturation
Third-party data note: live rows come from public benchmark and pricing feeds, not internal Dreamers testing. Restricted-access and unknown-availability rows are excluded from public leaders; public previews remain public. Cost per task estimates use 100K input and 30K output tokens; tools, retries, cache, and failed attempts are excluded.