LLM Benchmark Leaderboard
Benchmark scores for 11 major language models, compiled July 2026 from public leaderboards and vendor announcements. Blank cells mean no published score — scores are never estimated. Sortable by provider and evaluation metric.
| Model | Params | SWE-bench V | SWE-bench Pro | GPQA-D | OSWorld | Terminal-Bench | ARC-AGI-2 | Context | Notes |
|---|---|---|---|---|---|---|---|---|---|
Claude Fable 5 Anthropic · 2026 | Unknown | 95.0 | — | — | — | — | — | 1M | Anthropic premium tier — leads SWE-bench Verified among generally available models |
GPT-5.5 OpenAI · 2026 | Unknown | 88.7 | — | 93.6 | — | — | — | 1M | SWE-bench Verified score is vendor-reported |
Claude Opus 4.8 Anthropic · 2026 | Unknown | 88.6 | 69.2 | 93.6 | — | — | — | 1M | Leads SWE-bench Pro among active models |
Claude Sonnet 5 Anthropic · 2026-06 | Unknown | 82.1 | 63.2 | 96.2 | 88.3 | — | 84.7 | 1M | GPQA Diamond record holder (96.2%) at mid-tier pricing |
Gemini 3.1 Pro Google · 2026 | Unknown | 80.6 | — | 94.3 | — | — | 77.1 | 1M | Strong multimodal reasoning; video input |
DeepSeek V4 Flash DeepSeek · 2026-04 | 284B MoE (13B active) | 73.7 | — | 86.0 | — | — | — | 1M | Open weights (MIT); 79.0% SWE-bench Verified at max reasoning effort |
DeepSeek V4 Pro DeepSeek · 2026-04 | 1.6T MoE (49B active) | 73.6 | — | — | — | — | — | 1M | Open weights (MIT); 80.6% SWE-bench Verified at max reasoning effort |
Claude Haiku 4.5 Anthropic · 2025-10 | Unknown | 73.3 | — | — | — | — | — | 200K | Fastest and cheapest Claude |
o3 OpenAI · 2025-04 | Unknown | 69.1 | — | 83.3 | — | — | — | 200K | Scores from OpenAI's April 2025 announcement |
GPT-5.4 OpenAI · 2026 | Unknown | — | — | — | 75.0 | — | — | 1M | Value flagship with tool search |
Gemini 3.5 Flash Google · 2026 | Unknown | — | 55.1 | — | — | 76.2 | — | 1M | Agentic coding at fast-tier pricing |
Data as of May 2025. Scores may vary by source, prompt format, and evaluation methodology. Always verify against primary sources.
About
This leaderboard aggregates published benchmark scores for leading large language models across six current evaluations: SWE-bench Verified (real-world software engineering), SWE-bench Pro (contamination-resistant coding), GPQA Diamond (PhD-level science), OSWorld-Verified (computer use), Terminal-Bench 2.1 (agentic terminal work), and ARC-AGI-2 (abstract reasoning). Scores are sourced from public leaderboards and vendor announcements as of July 2026 — blank cells mean no published score, never an estimate. Filter by provider to narrow the comparison, click column headers to sort.
How to use
- 1 Use the provider filter chips to show only models from specific vendors.
- 2 Click any column header to sort by that benchmark — click again to reverse.
- 3 Score cells are color-coded: green (≥90), amber (75–89), muted (<75).
- 4 Missing scores are shown as "—" (not all models are evaluated on all benchmarks).
- 5 Click "What is X?" links to expand descriptions of each benchmark.
- What is SWE-bench Verified?
- SWE-bench Verified is a set of 500 human-validated software engineering tasks taken from real GitHub issues in popular Python repositories. The model must produce a code patch that passes the repository's own test suite. It has become the standard measure of practical, agentic coding ability — as of July 2026 the top published score is 95% (Claude Fable 5).
- What is GPQA Diamond?
- GPQA Diamond is a set of 198 graduate-level physics, chemistry, and biology questions written by domain experts and designed to be "Google-proof" — skilled humans with full web access score only ~34%. Frontier models now exceed 90%, making it one of the few knowledge benchmarks that still differentiates top models.
- Are these benchmarks still reliable?
- Benchmark saturation is why this leaderboard tracks the current generation of evaluations: classics like MMLU, HumanEval, and GSM8K are effectively solved (top models cluster above 90%), so the industry moved to harder, contamination-resistant tests like SWE-bench Pro and ARC-AGI-2. Even so, vendor-reported scores can use favourable scaffolding — for robust comparisons, combine benchmarks with task-specific evaluations on your own data.