The smartest AI model right now is Claude Opus 5.5, scoring 58 on the Intelligence Index

Which AI model is the smartest right now?

Every one of the top six models was released in September 2026, so a ranking from the summer is already out of date. This is the board as Artificial Analysis showed it on 11 October 2026, with each model at its strongest reasoning setting. The prices are the list rates for the API, per million tokens. The last cost column is what it cost Artificial Analysis, on average, to run one task of its index.

The dotted line is the Pareto frontier, which is the best score you can get at each cost. Four of the models sit on it. GPT-6.1 Sol reaches 52 for $0.72 a task, and Claude Opus 5.5 adds six points for eight times as much money. Gemini 4 Argon and Claude Sonnet 5.5 fill in the middle. Off the chart to the left, MiMo-V2.6-Pro is the strongest of the open-weights models at 46 for $0.13 a task, and Claude Haiku 5.5 scores 43 for $0.21.

Do other leaderboards agree that Claude Opus 5.5 is the smartest?

Only partly. Claude Opus 5.5 leads the composite index and Humanity's Last Exam, OpenAI's models lead on the ARC-AGI puzzles, and the blind voters on LMArena put Gemini 4 Argon first. Here's how the five strongest models do on four of the independent boards, all of which we read on 11 October 2026.

Each of the boards measures something different, and that's why they don't agree. ARC Prize tests abstract puzzles that are easy for people, and it cost $26,098 to run GPT-6 Astra to its 62.7% on ARC-AGI-3 at max reasoning. LMArena ranks by blind human votes. Argon's 1525 rests on 4,892 votes, against 6,272 for Opus 5.5, so the board still marks it as preliminary.

Pick Claude Opus 5.5 for hard reasoning and for knowledge work. Pick GPT-6 Astra when the task is a new kind of puzzle. Don't pick Argon yet. You can't.

Is the most expensive AI model the smartest?

No. Claude Fable 5.1 and GPT-6 Astra are the priciest models on the board at $50 per million output tokens, and both of them score 53. That's five points under Claude Opus 5.5, which lists at $20. Fable 5.1 was also the most expensive one to test, at $7.63 a task.

The price per token can mislead you in the other direction too. Claude Sonnet 5.5 lists at half the rate of Opus 5.5, yet it wrote 420 million tokens on the index to the 260 million that Opus wrote, so a task cost $5.46 against $5.98. Half the price bought you a 9% saving. GPT-6.1 Sol has the same $2 and $10 list price as Sonnet 5.5, but it wrote 67 million tokens and cost $0.72 a task. How much a model writes sets the bill as much as its rate card does.

Where can you use the smartest AI models, and on which plan?

Claude Opus 5.5 is the easiest of the top models to reach, because Anthropic includes Opus on the $20 a month Claude Pro plan. The two $50 models sit behind $100 plans. Gemini 4 Argon isn't for sale at all. Google says it "is rolling out to a set of trusted cyber defenders through our Fairwind Program," with paid API customers and Google AI Ultra subscribers next and no date given (Google, 30 September 2026).

The plan details come from OpenAI's help centre, Anthropic's Fable article and xAI's Grok 4.7 post.

Whizi carries six of these seven. Its Pro plan at $29.99 a month has Claude Opus 5.5, Claude Sonnet 5.5, GPT-6.1 Sol and Grok 4.7, and its 2,500 monthly credits buy 125 messages with Opus. Powerhouse at $49.99 adds Claude Fable 5.1 and GPT-6 Astra at 100 credits each, so its 10,000 credits cover 100 of those messages for half the entry price of ChatGPT Pro or Claude Max. Whizi runs every model at minimal reasoning effort by default, though, so a turn in Whizi is faster and shallower than the max setting that earned these scores. And if Opus 5.5 is all you want, Claude Pro at $20 is the cheaper way to get it.

How does Artificial Analysis measure intelligence?

The Intelligence Index v4.3.2 combines 10 evaluations in four groups: agents at 30%, general at 30%, coding at 20% and scientific reasoning at 20%. Humanity's Last Exam is one of the ten, weighted at 10%. Artificial Analysis describes the goal as to "assess language model capabilities across reasoning, knowledge, maths and programming" in its methodology.

Scores only compare within one version of the index, so a figure that you copy from an older roundup may sit on a different scale. Each model is also scored at its highest setting, which costs more and runs slower than its lower ones do. If you want a pick by task, for writing or for code or for long documents, see the best AI to use for each job.