#1 since Oct 23, 2026. No earlier leader is on record.
Leads the hardest expert-level exam across every academic field. These picks are readable without an account, and the details behind each one are on the card.
Pro is $20/month, or $17 if billed yearly. API is $4 per million input tokens and $20 per million output.
Free tier
Not on the free plan. Opus needs Pro or Max.
Privacy default
Anthropic says Free, Pro, and Max chats are used for training only if you allow it, and you choose at signup. The toggle has been reported as on by default, so check it. Settings, then Privacy. Turn off Help improve Claude. Source
When not to use it
Do not trust a proof you have not checked. Ask it to show the steps.
Starter prompt
Solve this step by step, and mark any step you are not sure about.
Humanity's Last Exam not cited. #1 GPQA Diamond (96.1%). Source.
Cost
Needs a paid ChatGPT plan. API is $10 per million input tokens and $50 per million output.
Free tier
Not on the free plan. Plus includes limited Astra use in Work and Codex.
Privacy default
Yes on personal Free, Plus, and Pro. Improve the model for everyone is on by default. Settings, Data controls, turn off Improve the model for everyone. Source
When not to use it
Skip it for a quick arithmetic check. A fast model is enough.
Starter prompt
Work this graduate-level question and cite the principle behind each step.
Weights are free under MIT. API is about $0.44 per million input tokens and $0.87 per million output.
Free tier
The weights are free to download and run yourself.
Privacy default
No, if you download the weights and run them yourself. Hosted Xiaomi apps may differ. Run the open weights on your own machine when chats must stay private. Source
When not to use it
Skip it when you need the single highest exam score. The leader still leads that board.
Starter prompt
Explain this concept with one example, then give me a harder practice problem.
Email
One email when a leader changes. Buttondown is the only service that holds the address, and it uses it only to send these alerts. Open and click tracking stay off. Privacy
Coming soon.
Leader history
Oct 23, 2026. Claude Opus 5.5 is the first leader on record. Humanity's Last Exam 61.4%.
A new number one has to show up on two daily checks before it replaces the current leader. Get alerts.
Source: WhichAI.today, Reasoning leader as of Oct 24, 2026, whichai.today/reasoning
Questions
Which AI is best for reasoning, math & science right now?
Claude Opus 5.5 from Anthropic leads, as of Oct 24, 2026. Leads the hardest expert-level exam across every academic field. The alternative is GPT-6 Astra, and the open-weights pick is MiMo-V2.6-Pro.
Where do I get Claude Opus 5.5?
Claude app, Pro plan ($20/month), choose Opus 5.5. Pro is $20/month, or $17 if billed yearly. API is $4 per million input tokens and $20 per million output.
Has this leader changed?
Tracking for this category began on Oct 23, 2026, with Claude Opus 5.5 as the first recorded leader.