What is the smartest AI in 2026? The models that lead, and where
Claude Fable 5.2 tops the main independent ranking, but GPT-6.2 Astra and Gemini 4 Neon win key tests. Here's which AI is smartest, and for what.

The smartest AI you can use right now is probably Claude Fable 5.2 from Anthropic, which tops the independent Artificial Analysis Intelligence Index with a score of 65 out of 100. It isn't a clean win, though. OpenAI's GPT-6.2 Astra and Google's Gemini 4 Neon beat it on several individual tests, and for most everyday questions you won't feel much difference between them.
So the answer to "what is the smartest AI" depends on what you mean by smart. I think that's the most useful thing to take away. It's also why three companies can each claim a crown in the same month without any of them lying outright.
Smartest depends on which exam you set
There's no IQ test for software. What we have instead are benchmarks: fixed sets of tasks with known answers, run on each model under the same conditions. Early ones were mostly multiple-choice trivia, and the top models now score close to full marks on many of them. The newer ones look more like jobs. Terminal-Bench-Science asks a model to carry out scientific work in a command line; Harvey's legal benchmark checks whether a contract review is correct and free of made-up facts; FrontierCode 2.0 checks whether code changes would actually get merged by a project's maintainers.
Composite scores try to boil all of that down to one number. The Artificial Analysis Intelligence Index, the one you'll see quoted most, averages nine evaluations across four equally weighted areas: knowledge work, coding, science and general knowledge. It's run by a company that doesn't build models, which is why I lean on it more than any chart a lab publishes about itself.
Still, a single number hides a lot.
The same group publishes a breakdown by field, and there the leader changes. Gemini 4 Neon ranks first for finance and accounting and for legal work, while Fable 5.2 leads in strategy, healthcare, engineering and economics. A school report card is a fair comparison here: the student with the best average isn't always top in chemistry, and if you're hiring a chemist, you probably shouldn't care much about the average.
The leading models side by side
Four models from three labs sit at the top of most rankings. Here's how they compare, using the scores the companies published plus the independent index where a score exists.
| Model | Maker | Where it stands out | API price per million tokens (input / output) |
|---|---|---|---|
| Claude Fable 5.2 | Anthropic | Highest Intelligence Index score (65); 74.2% on Humanity's Last Exam with tools | $8 / $40 |
| GPT-6.2 Astra | OpenAI | Leads Agents' Last Exam and Harvey's legal benchmark, by OpenAI's own figures | $8 / $40 |
| Gemini 4 Neon | 61 on the Intelligence Index; first on Zapier's AutomationBench (56.3%) | $3 / $15 | |
| Gemini 4 Carbon | Google's strongest model, aimed at long, multi-step agent work | $2 / $10 (introductory) |
Claude Fable 5.2
Anthropic's launch post for Fable 5.2 makes some big claims. The company says the model helped solve two open problems in the physics of black holes and formalized the proof of Fermat's Last Theorem in 30 hours. On Humanity's Last Exam, a set of very hard questions across dozens of subjects, it scored 74.2% with tools, against 67.7% for Anthropic's own Claude Opus 5.5. One odd detail: Fable and Claude Mythos 5.2 are the same underlying model. Mythos drops some cybersecurity and biology safeguards, so Anthropic only gives it to vetted organizations.
Anthropic doesn't even push Fable for daily use. It tells people to start with Opus 5.5 and move up to Fable only for the most ambitious technical work. That surprised me a little, and honestly I think it's good advice.
GPT-6.2 Astra
OpenAI pitches GPT-6.2 Astra as its frontier model for professional work like slides, spreadsheets, legal briefs and code. On Agents' Last Exam, which tests long multistep workflows, OpenAI says Astra scored 5.5 points ahead of Fable 5.2 at 5% of the cost. On Harvey's legal test, it claims more than double Fable's pass rate. Those are OpenAI's numbers, run OpenAI's way, and a footnote admits the Fable result leaves out nine tasks that used a fallback model. Small print matters.
Gemini 4 Neon and Carbon
Google is making a different case: near-top intelligence for less money. Gemini 4 Neon scored 61 on the Intelligence Index, third overall at launch, and it costs $3 per million input tokens against $8 for Fable and Astra. Gemini 4 Carbon is the one Google calls its strongest, built for "tasks that test the limits of a model's intelligence like complex, multi-step coding and debugging".
Here's a small example of why I don't trust any single lab's chart. On Terminal-Bench-Science, Anthropic reports 74.6% for Fable 5.2, Google calls Neon "best in class" at 70.5%, and OpenAI says Astra "sets a new high". They can't all be first. The likely explanation is different settings, versions and test dates; nobody's inventing numbers, but each company picks the comparison that flatters it.
And xAI's Grok? Grok 4.7 sits around 12th to 14th of the 15 models in the Artificial Analysis field-by-field comparison, so it's not in the conversation for "most advanced AI" at the moment (that could change with one release).
Strengths and limits of benchmark rankings
Benchmarks do some things well. They're repeatable, they let you compare models from rival companies on the same tasks, and independent ones like Artificial Analysis give you a check on vendor marketing. They also show cost, which matters more than people think: Neon gets within four points of Fable on the index at well under half the token price.
The limits are just as real. Labs choose which tests to show you, and they tend to show the ones they win. Scores are usually reported at the highest "effort" setting, where a model thinks for much longer and burns far more compute than it would in a normal chat. Anthropic, for instance, sets Fable 5.2 to medium effort by default in its own apps, so the 65 on the index isn't quite what you'd get typing a question on your phone. Test questions can also leak into training data over time, which slowly inflates scores on older benchmarks.
Then there's the bigger gap: benchmarks measure tasks with right answers. They say little about judgement, taste, or whether a model tells you when it doesn't know. I'd guess that's where most of the frustration with chatbots actually comes from, and it's hard to put on a leaderboard.
Picking a model for your own work
If you're asking which is the smartest AI chatbot for you, the first filter is your plan. These top models mostly sit behind paywalls. On Claude's plans, Fable isn't available on the free tier; Pro users pay for it with usage credits, and Max and Ultra subscribers can spend up to half their weekly limit on it. ChatGPT's free tier runs GPT-6 Luna, with the Astra-powered Pro reasoning reserved for the top ChatGPT plan. Google gives Neon to AI Pro and Ultra subscribers in the Gemini app, while Carbon goes to Ultra subscribers and to AI Mode in Search.
For emails, summaries, homework help and most coding, a mid-tier model is probably enough. OpenAI itself calls GPT-6.2 Sol its "go-to model for everyday work", and Anthropic says the same about Opus 5.5. I'd save the frontier models for long, hard jobs: a big refactor, a messy data analysis, a research question with no obvious answer.
The most telling test is your own work. Run the same real task through two models and compare the results; that'll tell you more than any index. If you build software, an AI gateway lets you send different requests to different models without rewriting your app. Our Claude vs ChatGPT comparison covers the everyday differences between the two apps.
The next few months
The top spot changes hands fast. Fable 5.1 has already given way to 5.2, and GPT-6 Astra to 6.2 Astra, and each release reshuffled the board. I'd expect whatever leads today to be overtaken before spring, probably by a model that doesn't exist yet.
Benchmarks are moving too. The ARC Prize has shifted to ARC-AGI-3, which asks agents to adapt on the fly to new interactive environments instead of solving static puzzles. Safety is becoming part of the ranking story as well: the fact that Anthropic keeps Mythos 5.2 behind a verification program says something about how capable these models now are in cybersecurity, and why guardrails are getting more attention.
Is any of this artificial general intelligence? In my view, no. These models are brilliant at tasks with clear goals and still uneven at plenty of things a person does without thinking. Our explainer on what AGI means goes into that gap.
FAQ
What is the smartest AI in the world?
By the independent Artificial Analysis Intelligence Index, Claude Fable 5.2 from Anthropic is currently the smartest AI, with a score of 65 out of 100. GPT-6.2 Astra and Gemini 4 Neon are close behind and win some individual tests, so the lead is narrow and could change with the next release from any of the three labs.
Who is the smartest AI?
There isn't one AI that wins at everything. Claude Fable 5.2 leads the main composite ranking, GPT-6.2 Astra leads OpenAI's legal and long-workflow tests, and Gemini 4 Neon ranks first for finance and legal work in one field-by-field comparison. The smartest choice for you depends on the kind of work you do.
What is the smartest AI chat bot?
Among chatbots you can sign up for today, Claude with Fable 5.2 is probably the smartest, but it needs a paid plan. ChatGPT and Gemini offer close rivals in GPT-6.2 Astra and Gemini 4 Neon. For everyday questions, I think the free versions of all three apps are good enough for most people.
What is the most intelligent AI?
Claude Fable 5.2 has the highest score on the Artificial Analysis Intelligence Index, which averages nine tests across coding, science and knowledge work. That's intelligence on hard tasks with checkable answers. It doesn't capture judgement or common sense.
Who is the smartest AI in the world?
No single company owns the title. Anthropic's Claude Fable 5.2 tops the leading independent index, while OpenAI and Google each publish benchmarks where their own models come first. Leadership has been changing every few weeks, so treat any ranking as a snapshot and test the models on your own real tasks before paying.


