Compare / Anthropic vs OpenAI
Anthropic vs OpenAI: which AI lab is ahead?
Each lab's best model on sale goes head to head on every kind of work we measure, from our own benchmarks and licensed sources. 43 models in all; a lab is only as good as the model you would actually pick from it.
Kinds of work led by each lab's best model there, of the 10 both are measured on.
Results as of 19 November 2026 (catalogue release 2026-11-19-3608133e8331). Models on sale means listed on OpenRouter today.
Anthropic or OpenAI? The short answer
Anthropic is ahead: its best models beat OpenAI's in 6 of the 10 kinds of work both are measured on; OpenAI leads in 3, and 1 is a tie.
- Pick Anthropic for user surveys, marketing planning, agents and tool use, reasoning and knowledge, writing and sticking to the facts.
- Pick OpenAI for product listings, decks from an analysis and speed under pressure.
- On our overall leaderboard, Claude Fable 5.2 is #1 (88 of 100) and GPT-6.2 Astra is #2 (86).
- Prices start at $0.20 per million tokens for Anthropic (Claude Haiku 5.5) and $0.058 for OpenAI (gpt-oss-20b).
Head to head
Each line names the model behind the number. Prices are blended, three input tokens to one output, at the developer's list price where we have read it; see every model's price. Intelligence Index from Artificial Analysis; overall score from our leaderboard.
Which lab is better for what
For each kind of work, each lab's best model there and its rank among every model on sale. Further right is better; the better rank is in green.
Every model, by intelligence and price
Each dot is a model on sale with an Intelligence Index from Artificial Analysis. Up is smarter; left is cheaper. Hover or tap a dot for its name.
AnthropicOpenAI
On our overall leaderboard
The 13 of their models we rank overall: one score out of 100 across everything we publish (how the score works). Only models with enough of our own results are ranked.
- #1Claude Fable 5.288
- #2GPT-6.2 Astra86
- #3GPT-6.2 Sol84
- #4Claude Opus 5.580
- #5GPT-6.1 Sol78
- #6GPT-6.1 Terra77
- #7GPT-6 Astra76
- #10GPT-6 Sol66
- #12Claude Fable 5.164
- #14Claude Sonnet 5.560
- #27Claude Haiku 5.538
- #27GPT-6 Luna38
- #39Claude Haiku 4.57
Benchmark by benchmark
The headline result of every benchmark both labs have models on: each lab's best model and its rank among the models on sale. Green marks the better rank.
| Measure | Anthropic | OpenAI |
|---|---|---|
| Agents and tool use | ||
| τ²-bench · Task success (pass^1) · tau2 v1.0.1, retail% of tasks, higher is better | 79.6%Claude Opus 4.5 · 3 of 6 | 81.6%GPT-5.2 · 2 of 6 |
| τ²-bench · Task success (pass^1) · tau2 v1.0.1, airline% of tasks, higher is better | 84.0%Claude Opus 4.5 · 1 of 6 | 83.0%GPT-5.2 · 2 of 6 |
| τ²-bench · Task success (pass^1) · tau2 v1.0.1, telecom% of tasks, higher is better | 92.3%Claude Opus 4.5 · 2 of 6 | 89.7%GPT-5.2 · 4 of 6 |
| Berkeley Function Calling Leaderboard (BFCL) V4 · Overall accuracy · BFCL% correct, higher is better | 77.5%Claude Opus 4.5 · 1 of 36 | 63.1%o3 · 5 of 36 |
| OpenHands Index · Average% resolved, higher is better | 81.0%Claude Fable 5 · 1 of 31 | 65.9%GPT-5.5 · 5 of 31 |
| Microsoft STATE-Bench · Task success (pass@1) · STATE-Bench v0.7% of tasks, higher is better | 53.4%Claude Opus 4.7 · 2 of 4 | 55.7%GPT-5.4 · 1 of 4 |
| Artificial Analysis · Terminal-Bench 4.0% of tasks, higher is better | 63.6%Claude Sonnet 5.5 · 2 of 107 | 59.6%GPT-6 Astra · 3 of 107 |
| Artificial Analysis · Terminal-Bench 2.1% of tasks, higher is better | 91.4%Claude Fable 5.1 · 1 of 107 | 89.9%GPT-6 Astra · 2 of 107 |
| Artificial Analysis · τ-bench banking% of tasks, higher is better | 47.2%Claude Fable 5.1 · 7 of 104 | 44.3%GPT-5.6 Sol · 12 of 104 |
| Vending-Bench 2 · Money after a yearUS dollars, higher is better | $11,182Claude Opus 5 · 3 of 53 | $15,515GPT-6 Astra · 1 of 53 |
| Decks from an analysis | ||
| DeckBench · Deck ratingrating, higher is better | 1,403Claude Fable 5.2 · 4 of 35 | 1,551GPT-6 Astra · 2 of 35 |
| Marketing planning | ||
| ROASBench · Overall scorescore out of 100, higher is better | 61.3Claude Fable 5.2 · 1 of 35 | 58.4GPT-6.1 Terra · 2 of 35 |
| Product listings | ||
| CatalogBench · Reliably publish-ready% of products, higher is better | 71.4%Claude Fable 5.2 · 6 of 35 | 75.0%GPT-6.2 Sol · 2 of 35 |
| CatalogBench · Reliably publish-ready · sales brief% of products, higher is better | 44.6%Claude Fable 5.2 · 9 of 35 | 75.0%GPT-6.2 Sol · 1 of 35 |
| Professional work | ||
| APEX-Agents · Tasks passed% of tasks, higher is better | 75.5%Claude Sonnet 5.5 · 1 of 37 | 64.7%GPT-6 Astra · 7 of 37 |
| GDP.pdf · Professional document tasks% of rubric, higher is better | 30.6%Claude Opus 5.5 · 2 of 31 | 30.7%GPT-5.6 Sol · 1 of 31 |
| Remote Labor Index · Projects done to client standard% of projects, higher is better | 17.9%Claude Fable 5.1 · 2 of 12 | 20.8%GPT-6 Astra · 1 of 12 |
| Reasoning and knowledge | ||
| Artificial Analysis · Artificial Analysis Intelligence Indexindex score, higher is better | 57.6Claude Opus 5.5 · 1 of 185 | 52.7GPT-6 Astra · 4 of 185 |
| Artificial Analysis · Humanity's Last Exam% of questions, higher is better | 61.4%Claude Opus 5.5 · 1 of 186 | 54.7%GPT-6 Astra · 7 of 186 |
| Artificial Analysis · GPQA Diamond% of questions, higher is better | 93.7%Claude Fable 5.1 · 8 of 168 | 96.3%GPT-6 Astra · 1 of 168 |
| Speed under pressure | ||
| BulletBench · Ladder Elo · Lightning 10+1ladder Elo, higher is better | 33Claude Sonnet 5.5 · 16 of 18 | 584GPT-6.1 Sol · 4 of 18 |
| BulletBench · Ladder Elo · Bullet 60sladder Elo, higher is better | 491Claude Fable 5.2 · 13 of 36 | 928GPT-6.1 Sol · 2 of 36 |
| Sticking to the facts | ||
| Vectara Hallucination Leaderboard · Hallucination rate · Vectara% of summaries, lower is better | 9.8%Claude Haiku 4.5 · 40 of 71 | 3.1%GPT-5.4 nano · 1 of 71 |
| Arena (formerly LMArena) · Overall · Text factualityArena rating, higher is better | 1,507Claude Opus 5.5 · 2 of 109 | 1,486GPT-5.5 · 6 of 109 |
| SimpleQA Verified (Epoch AI) · Correct answers% of questions, higher is better | 72.2%Claude Opus 5.5 · 4 of 68 | 75.6%GPT-6 Astra · 1 of 68 |
| User surveys | ||
| SurveyBench · Analysis ratingrating, higher is better | 1,670Claude Fable 5.2 · 1 of 26 | 1,476GPT-6.2 Astra · 2 of 26 |
| Writing | ||
| Arena (formerly LMArena) · Overall · TextArena rating, higher is better | 1,505Claude Opus 4.6 · 2 of 153 | 1,484GPT-5.6 Sol · 15 of 153 |
| UGI Leaderboard · Writing score · UGIscore out of 100, higher is better | 74.2Claude Fable 5 · 3 of 124 | 69.7GPT-5.5 · 11 of 124 |
Their models, one against one
- Claude Fable 5.2 vs GPT-6.2 AstraBenchmarks, price and which to pick
- Claude Fable 5.2 vs GPT-6.2 SolBenchmarks, price and which to pick
- Claude Fable 5.2 vs GPT-6.1 TerraBenchmarks, price and which to pick
- Claude Opus 5.5 vs GPT-6.2 SolBenchmarks, price and which to pick
- Claude Opus 5.5 vs GPT-6.1 SolBenchmarks, price and which to pick
- Claude Opus 5.5 vs GPT-6 AstraBenchmarks, price and which to pick
- Claude Sonnet 5.5 vs GPT-6.1 SolBenchmarks, price and which to pick
- Claude Haiku 4.5 vs GPT-6.1 TerraBenchmarks, price and which to pick
Quick answers
Is Anthropic or OpenAI better?
Anthropic is ahead: its best models beat OpenAI's in 6 of the 10 kinds of work both are measured on; OpenAI leads in 3, and 1 is a tie.
What is Anthropic's best model?
Claude Fable 5.2: #1 on the Spring Prompt overall leaderboard.
What is OpenAI's best model?
GPT-6.2 Astra: #2 on the Spring Prompt overall leaderboard.
Which is better for agents and tool use, Anthropic or OpenAI?
Anthropic: its best model there, Claude Sonnet 5.5, ranks 2 of 159 models on sale, against 4 for OpenAI's GPT-6 Astra.
Which is better for reasoning and knowledge, Anthropic or OpenAI?
Anthropic: its best model there, Claude Opus 5.5, ranks 1 of 187 models on sale, against 3 for OpenAI's GPT-6 Astra.
Which is better for writing, Anthropic or OpenAI?
Anthropic: its best model there, Claude Fable 5, ranks 2 of 169 models on sale, against 15 for OpenAI's GPT-5.5.
Which is cheaper, Anthropic or OpenAI?
OpenAI's cheapest model with results, gpt-oss-20b, costs $0.058 per million tokens (three input to one output), against $0.20 for Claude Haiku 5.5 from Anthropic.
Other labs compared
- Google vs OpenAIBest model against best model
- Anthropic vs GoogleBest model against best model
- OpenAI vs xAIBest model against best model
- Anthropic vs xAIBest model against best model
- Meta vs OpenAIBest model against best model
- Anthropic vs MetaBest model against best model
- DeepSeek vs OpenAIBest model against best model
- Anthropic vs DeepSeekBest model against best model
Which model is best for your product?
Benchmarks aren't your task. Run models from Anthropic, OpenAI and the rest on your own prompt, and see the cheapest one that passes.