ChatGPT vs Claude vs Gemini, compared clearly.
Independent benchmarks, practical rankings, and side-by-side results.
Every figure below was read off a named public leaderboard on a named date, in that board's own metric, next to the exact model it measured. The boards measure different things, so there is no combined score - pick a board to order the list by.
| Model | Overall Arena Text | Coding Arena WebDev | Research Arena Search | Image Arena T2I | Coding SWE-bench | Research GPQA | Agents tau2-bench | Cost | Details |
|---|---|---|---|---|---|---|---|---|---|
| 1,506 Eloclaude-fable-5#1 of 387 · source for claude-fable-5 on Arena (LMArena) Text Leaderboard | 1,692 Eloclaude-opus-5-max#1 of 112 · source for claude-opus-5-max on Arena WebDev Leaderboard | Not listed | Not listed | 96%Claude Opus 5#1 of 61 · source for Claude Opus 5 on SWE-bench Verified | 93.6%Claude Opus 4.8#6 of 75 · source for Claude Opus 4.8 on GPQA Diamond | 79%Claude Opus 4.5#1 of 8 · source for Claude Opus 4.5 on tau2-bench (pass rate, averaged across domains) | $20 per month As published: Free / $20 mo | View details | |
| 1,486 Elogemini-3.1-pro-preview#13 of 387 · source for gemini-3.1-pro-preview on Arena (LMArena) Text Leaderboard | 1,536 Elogemini-3.6-flash#18 of 112 · source for gemini-3.6-flash on Arena WebDev Leaderboard | Not listed | Not listed | 63.8%Gemini 2.5 Pro#55 of 61 · source for Gemini 2.5 Pro on SWE-bench Verified | 92.2%Gemini 3.5 Flash#15 of 75 · source for Gemini 3.5 Flash on GPQA Diamond | 69%Gemini 3 Pro#3 of 8 · source for Gemini 3 Pro on tau2-bench (pass rate, averaged across domains) | $20 per month As published: Free / $20 mo | View details | |
| 1,481 Elogpt-5.5-high#17 of 387 · source for gpt-5.5-high on Arena (LMArena) Text Leaderboard | 1,623 Elogpt-5.6-sol-xhigh#6 of 112 · source for gpt-5.6-sol-xhigh on Arena WebDev Leaderboard | Not listed | Not listed | 85%GPT-5.3 Codex#7 of 61 · source for GPT-5.3 Codex on SWE-bench Verified | 94.6%GPT-5.6 Sol#3 of 75 · source for GPT-5.6 Sol on GPQA Diamond | 73%GPT-5.2#2 of 8 · source for GPT-5.2 on tau2-bench (pass rate, averaged across domains) | $20 per month As published: Free / $20 mo | View details | |
| 1,469 Elogrok-4.5#34 of 387 · source for grok-4.5 on Arena (LMArena) Text Leaderboard | 1,554 Elogrok-4.5#12 of 112 · source for grok-4.5 on Arena WebDev Leaderboard | Not listed | Not listed | 76.7%Grok 4.20#32 of 61 · source for Grok 4.20 on SWE-bench Verified | 90.1%Grok 4.3#22 of 75 · source for Grok 4.3 on GPQA Diamond | Not listed | $0 As published: Free with X / $8 mo with X Premium | View details | |
| 1,427 Elomistral-medium-3.5#94 of 387 · source for mistral-medium-3.5 on Arena (LMArena) Text Leaderboard | Not listed | Not listed | Not listed | 77.6%Mistral Medium 3.5 128B#26 of 61 · source for Mistral Medium 3.5 128B on SWE-bench Verified | Not listed | Not listed | No monthly entry price · $0.25 per M tokens As published: Free (open) / API from $0.25/M tokens | View details | |
| Not listed | Not listed | 1,138 Eloppl-sonar-reasoning-pro-high#27 of 32 · source for ppl-sonar-reasoning-pro-high on Arena Search Leaderboard | Not listed | Not listed | Not listed | Not listed | $20 per month As published: Free / $20 mo | View details | |
| Not listed | Not listed | Not listed | 938 Elostable-diffusion-v35-large#76 of 77 · source for stable-diffusion-v35-large on Arena Text-to-Image Leaderboard | Not listed | Not listed | Not listed | from $10 per month As published: Free (open-source) / hosted from $10 mo | View details |
No models match.
Anthropic API and OpenAI API expose the same models as Claude and ChatGPT, so their benchmark rows are identical and they are listed once here. Their pricing differs - see each tool page.
View the full Arena (LMArena) Text Leaderboard →
Why the other 38 tools are not on this board
Not publicly benchmarked
In the comparison, out of the figures. No public capability benchmark covers these, and the reason is not the same in every case.
- Meta Llama Harness - runs on benchmarked models; the harness changes the result
Meta ended the Llama line. Its current models ship as Muse - Muse Spark (frontier, closed weights, April 2026) and Muse Glimmer (30B, Apache 2.0, 2026-08-10). Confirmed by absence: no Llama entry anywhere on the Arena text board (387 models), the SWE-bench Verified board (61 models), or a 381-model aggregate leaderboard. Muse scores exist but belong to Muse, and attributing them to Llama would break the entity rule. The reason class is the nearest fit of the four available, not a precise description - the precise statement is that this model family has been superseded and no current board ranks it. - You.com Harness - runs on benchmarked models; the harness changes the result
- Elicit Harness - runs on benchmarked models; the harness changes the result
- Cursor Harness - runs on benchmarked models; the harness changes the result
- GitHub Copilot Harness - runs on benchmarked models; the harness changes the result
- Devin Desktop (formerly Windsurf) Harness - runs on benchmarked models; the harness changes the result
- Replit Harness - runs on benchmarked models; the harness changes the result
- v0 by Vercel Harness - runs on benchmarked models; the harness changes the result
Expected on Design Arena as a product-level entry. Two readings on 2026-08-11 say otherwise: designarena.ai renders its tables client-side and served only Loading placeholders, and the readable mirror of its Website board lists 89 MODELS (Kimi K3, Claude Opus 5, GLM-5.2) with no v0, Lovable or Bolt entry at all. So there is no product leaderboard to cite, and v0 is classed with the other coding harnesses. - Make (formerly Integromat) SaaS application - no public capability benchmark exists
- Zapier SaaS application - no public capability benchmark exists
- n8n SaaS application - no public capability benchmark exists
- Activepieces SaaS application - no public capability benchmark exists
- Midjourney SaaS application - no public capability benchmark exists
Expected on the image arena and genuinely is not there. Checked the Arena text-to-image board on 2026-08-11: 77 models, zero Midjourney entries. It is a closed app with no public API, and API access is the entry condition for that board. Its absence is a distribution fact, not a quality signal - do not render it as a low score. - Figma AI Harness - runs on benchmarked models; the harness changes the result
- Jasper Harness - runs on benchmarked models; the harness changes the result
- Copy.ai Harness - runs on benchmarked models; the harness changes the result
- Writesonic Harness - runs on benchmarked models; the harness changes the result
- Notion AI Harness - runs on benchmarked models; the harness changes the result
- LangChain Framework - benchmarks measure the models it orchestrates
- LangGraph Framework - benchmarks measure the models it orchestrates
- CrewAI Framework - benchmarks measure the models it orchestrates
- AutoGen Framework - benchmarks measure the models it orchestrates
- Agno Framework - benchmarks measure the models it orchestrates
- Pinecone Infrastructure - capability benchmarks do not apply
- Weaviate Infrastructure - capability benchmarks do not apply
Same as qdrant: present in ann-benchmarks, results published as plots only, no readable numeric table on the overview or dataset pages as of 2026-08-11. Nothing recorded rather than something estimated off a chart. - Qdrant Infrastructure - capability benchmarks do not apply
ann-benchmarks does benchmark Qdrant, but publishes recall/QPS only as plotted images - the overview and the glove-100-angular dataset page carry no numeric table. No figure could be read, so none is recorded. Any number would also be tied to one machine and one index config, which is why the class says capability benchmarks do not apply. - Railway Infrastructure - capability benchmarks do not apply
- Vercel Infrastructure - capability benchmarks do not apply
- Supabase Infrastructure - capability benchmarks do not apply
- Bubble SaaS application - no public capability benchmark exists
- Retool SaaS application - no public capability benchmark exists
- FireCrawl Infrastructure - capability benchmarks do not apply
- Hugging Face Infrastructure - capability benchmarks do not apply
- Resend Infrastructure - capability benchmarks do not apply
- Slack AI Harness - runs on benchmarked models; the harness changes the result
- Airtable SaaS application - no public capability benchmark exists
- Linear SaaS application - no public capability benchmark exists
- Stripe Infrastructure - capability benchmarks do not apply
How these figures were gathered
Benchmarks used:
- Arena (LMArena) Text Leaderboard · Measures general capability, scored in Elo points.
- Arena WebDev Leaderboard · Measures coding tasks, scored in Elo points.
- SWE-bench Verified · Measures coding tasks, scored as a percent.
- GPQA Diamond · Measures research and retrieval, scored as a percent.
- tau2-bench (pass rate, averaged across domains) · Measures multi-step agent tasks, scored as a percent.
- Arena Search Leaderboard · Measures research and retrieval, scored in Elo points.
- Arena Text-to-Image Leaderboard · Measures image generation, scored in Elo points.
Every figure is that leaderboard's own native metric, read on the date shown and reproduced as published. Nothing is normalised to a common scale, and there is no overall score, no overall rank and no weighted average anywhere on this page: these benchmarks measure different things, so a combined number would be one nobody published.
Caveats, in the source's own words:
- Arena (LMArena) Text Leaderboard Human preference votes, not correctness. A model can win here by being pleasant to read while being wrong. Scores carry a confidence interval of roughly plus or minus 3 to 10 points, so neighbouring ranks are usually ties.
- Arena WebDev Leaderboard Preference votes on generated web front-ends. Rewards code that looks right in a preview pane; it does not compile, test or run the result.
- SWE-bench Verified Treat as soft. UC Berkeley RDI (April 2026) scored 100% on all 500 tasks with an agent that wrote no solution code, by dropping a pytest hook that forces every test to pass. OpenAI separately found flawed tests in 59.4% of audited problems. Vendor-reported figures on this board are not independently replicated.
- GPQA Diamond Multiple-choice graduate science questions, so roughly 25% is the floor from guessing alone. The top of this board is compressed inside two points, which is narrower than the gap between reruns of the same model.
- tau2-bench (pass rate, averaged across domains) The softest numbers on this page, and they should be read as indicative only. In April 2026 UC Berkeley RDI pointed one scanning agent at the eight most-cited agent benchmarks and reached near-perfect scores on seven of them without doing the tasks - 100% on SWE-bench Verified, Terminal-Bench and FieldWorkArena, about 100% on WebArena, about 98% on GAIA - by exploiting the graders rather than the problems. GAIA is additionally contaminated: its validation answers are public on HuggingFace and its answer normalisation strips all whitespace and punctuation, so wrong answers collide with right ones. Independent audits have found reward hacking in 30%+ of evaluation runs. Agentic scores measure the harness and the grader as much as the model.
- Arena Search Leaderboard Board was last refreshed 2026-07-21, three weeks before this reading, and ranks only 32 models. Preference votes, not citation accuracy.
- Arena Text-to-Image Leaderboard Entry requires a public API, so several widely used image products are absent entirely rather than ranked low. Absence here is not a score.