Deck appendix · page 12 of the deck, in full

Which AI is best at what, the whole board.

Page 12 of the deck says no AI is best at everything and the leader keeps changing. This is the table behind that, with nothing trimmed for space, plus how it is scored and where it comes from.

Engine	Tier	Crypto	Biohack	Current events	Investing	Cybersec	Medical	Science	Legal
Claude Fable	Pro	78	85	70	88	90	92	95	91
Claude Opus	Pro	63	82	68	86	88	90	93	89
OpenAI o3	Pro	82	84	72	90	94	93	96	90
GLM	Pro	70	72	78	75	76	78	82	79
GPT-5.6 Sol	Pro	80	83	82	88	91	91	94	88
DeepSeek R1	Pro	80	77	60	88	88	81	94	79
Claude Sonnet	Plus	75	80	68	82	86	86	90	85
Claude Haiku	Plus	62	65	60	68	70	72	76	72
GPT-4o	Plus	72	75	80	80	82	84	86	82
Gemini Pro	Plus	44	78	90	0*	84	88	92	84
Kimi	Plus	66	68	78	74	72	75	80	80
Qwen Max	Plus	72	74	80	78	80	82	88	81
DeepSeek	Plus	76	75	65	82	84	80	91	78
GPT-OSS	Free	50	52	45	55	58	56	62	54
Gemini Flash Lite	Free	48	60	82	45	62	68	72	60
Llama 3.1 8B	Free	54	55	48	58	62	60	68	58
Grok	Staged	75	74	92	80	80	78	86	76
Mistral	Staged	70	72	65	76	82	78	85	84
Command R+	Staged	60	62	75	78	68	70	72	82

91-100 expert 76-90 strong 51-75 competent 21-50 weak 0-20 unreliable

What the tier column means

Pro & Plus: connected and available to customers on those plans.
Free: available to everyone, including people who have never signed up.
Staged: built and ready, just not switched on for customers yet.

How to read a score

A score is that engine's strength on that subject, out of 100. It is not a rating of the company and it is not an overall ranking. The whole point of the table is that the same engine can be excellent at one subject and poor at the one next to it, so an average across the row would throw away the only useful information in it.

*Gemini's 0 on investing is a real result, not a data-entry error. That model is built to refuse investment questions, so it scores 0 on a subject it will not answer. We left it in rather than tidy it away, because it is a good example of the thing we are measuring: a refusal is a wrong answer to the person who asked.

Where these numbers come from, and what they are not

This is our best current read of the landscape from public tests, July 2026. It is not our own lab measurement yet, and we would rather say that plainly than let a coloured grid imply more than it should.

What is measured by us, and published: the accuracy and cost result on page 1 of the deck. Held-out questions the system had never seen, counted, not estimated.
What this table is today: the starting position, assembled from public benchmark results, so that on day 1 we already know roughly who to ask about what.
What it becomes: the live version is built from real question outcomes on our own platform. Every time a person corrects an answer, the engine that gave it moves on the subject it gave it about. That is the version no AI lab can publish, because a lab grading itself against its competitors is not a thing anyone would believe.

Why this is the moat and not just a feature. The scoreboard is 1 asset doing 2 jobs. It decides where to send the next question, which is what makes the answers good and the cost low. And the corrections that keep it current are the same corrections the AI labs buy. Nobody has to be persuaded to maintain it: it is a byproduct of people using the product.

Nothing on this page is a claim about any company. It is a working record of which model we would send a given question to today, and it changes. Full sourcing available on request.

← Back to the deck Who else is doing this The forecast, line by line GTM notes & the stack