The AI Spotlight Follow
There's No Single 'Best AI Model' Anymore. OpenAI's Own Data Shows Why, and Why the Smartest Ones Are Getting Harder to Watch
Model Wars

There's No Single 'Best AI Model' Anymore. OpenAI's Own Data Shows Why, and Why the Smartest Ones Are Getting Harder to Watch

GPT-6.2 Astra, Claude Fable 5.2 and Gemini 4 Neon each 'win' depending on which leaderboard you open. The reason isn't raw intelligence; it's training data. And two new OpenAI reports add a twist: AI is now tripling some researchers' output, while its reasoning is becoming harder for humans to monitor. Here's how to pick a model for your work, and what the fine print says about where this is heading.

Advertisement

Every time a new frontier model ships, the same question floods comment sections: so which one is the best now?

This month, with GPT-6.2 Astra, Gemini 4 Neon and Claude Fable 5.2 all on the market, the honest answer is more useful than any single ranking. It depends on your job, and the reason is the data each model was trained on.

That sounds like a dodge. It isn't. Once you understand why the leaderboards disagree, you can choose models far more intelligently, and you'll also understand two new OpenAI reports that quietly change the conversation about where AI is going. One says AI is now accelerating AI research inside OpenAI by a lot. The other says the newest models are getting harder to monitor through their visible reasoning.

Here's the full picture.

Advertisement

The leaderboards disagree, and that's the point

Open a few respected rankings and you get different winners:

  • On aggregate intelligence indexes like Artificial Analysis, Claude Fable 5.2 leads narrowly.
  • On Epoch AI's Capabilities Index, another aggregate of many benchmarks, GPT-6.2 Astra comes out ahead.
  • On a long-PDF expert-task benchmark covering STEM documents, the brand-new Gemini 4 Neon scores best.
  • On one of the hardest coding benchmarks, Fable 5.2 leads, but every top model is bunched in the 40s.

If models differed only in general "IQ," these lists would agree. They don't, because models can only learn patterns that appear in their training data, and each lab buys, licenses, generates and emphasizes different data.

Data, not IQ, explains the weird gaps

Once you look for it, the data story explains a lot of otherwise strange results:

  • Life sciences. On tasks involving less-studied proteins, Fable 5.2 has reportedly outperformed GPT-6 Astra dramatically, and longer "thinking" didn't close the gap. In some tests, reasoning longer actually made answers worse than the instant reply. That's consistent with a lab that has openly prioritized life sciences, as Anthropic CEO Dario Amodei has said his company does.
  • Legal work. Some comparisons show Astra handily beating Fable on legal tasks at a fraction of the cost. Is Fable "dumber" at law, or did OpenAI simply train on more legal material?
  • Driving and spatial tasks. Gemini 4 Carbon has posted much stronger results on driving-style and spatial benchmarks. Think about the visual and mapping data Google has access to.
  • Calibration. Independent testers have found Gemini 4 models can be noticeably overconfident about whether their answers are right, compared with Claude models, with Astra somewhere in between.

Even anecdotes follow this pattern. One tester found Gemini 4 Carbon far better than 6.2 Astra at planning long countryside walks, because it dug into satellite imagery and local history in more detail.

Beware the "look how much better it got" demo

Labs love side-by-side visuals: the old model's clumsy 3D scene on the left, the new model's gorgeous one on the right.

Those comparisons can be real and still misleading. Observers pointed out that one recent before-and-after from OpenAI showed a weaker "before" than demos the company itself had published when the older model launched. Your eyes can be steered by the choice of baseline.

The lesson: trust your own tests on your own tasks more than any launch-day montage.

The same model is great and terrible depending on the task

OpenAI's new report, Research acceleration: The jagged frontier (November 18), makes the point with unusual precision, using its own researchers' work.

From August 1 to October 18, 2026, estimated success rates for its models working without human intervention varied wildly by task:

Task Autonomous success
Monitoring training runs for issues 79%
Answering technical questions 72%
Launching training runs 35%
Building evaluation, RL and agent environments 29%
Writing code for inference and serving stacks 25%

By broad category, "Communicate" tasks hit 69% and "Analyze" 58%, versus 43% for "Build."

The trend is steep, too. From March to October, success on tasks that take a human two to four hours rose from 23% to 46%. On four-to-eight-hour tasks, it rose from 13% to 35%.

That's why one well-known chart says models can handle tasks that take humans dozens of hours at high reliability, while another says reliability falls off for tasks over 15 minutes. Both can be true. It depends entirely on what kind of work you're measuring.

How to actually choose a model

Turn the data story into a routine:

  1. List your five most common tasks. Not "coding," but "refactor React components" or "summarize 80-page contracts."
  2. Run each task on all three families: Anthropic's Claude, OpenAI's GPT and Google's Gemini. Use the same prompt.
  3. Score the outputs blind if you can. Have a colleague strip the labels.
  4. Factor in cost per task, not cost per token. Gemini 4 Neon at $3/$15 per million tokens or GPT-6.2 Sol can beat a pricier flagship on value.
  5. Re-test every couple of months. With releases arriving every few weeks, today's winner may not hold.

For specifics on two of this month's contenders, see our GPT-6.2 Astra hands-on and Gemini 4 Neon test.

Progress is still predictable on average

Zoom out, and something else is clear: there's no wall. OpenAI showed an aggregate of 34 public benchmarks and 35 internal ones climbing steadily, and Anthropic has published its own version of that kind of chart.

Both curves share a detail. Around August to October 2025, when models got good enough to automate much of researchers' coding, the slope steepened. Freed from routine coding, researchers could try many more ideas in parallel.

But here's the catch. Look at those smooth curves and try to find the moment AI started solving Millennium Prize-level math problems, or the moment cyber-offense capability became a national security issue. You can't. Smooth averages hide sudden thresholds. The data for a given domain reaches critical mass, and abilities appear that no aggregate chart predicted.

That is the real reason to keep watching specific domains, not just headline scores.

How much is AI speeding up AI research?

This is the question behind every debate about self-improving AI, and the answer depends on how you ask it.

  • OpenAI's survey of 101 research staff, run in late August and early September, asked how long their team would need to produce work equal in value to the previous month's without frontier coding agents. The median answer: three months. Roughly a 3x boost.
  • The same survey found the median researcher would trade 40% of team headcount or 30% of team compute to keep frontier coding agents for three months.
  • Usage is exploding. At API list prices, the median OpenAI researcher's daily agent usage hit about $2,400 in October, and the 90th percentile reached about $23,000 a day.
  • Anthropic has discussed a threshold of AI doubling its historical rate of progress, and different internal surveys have produced very different "speed-up" numbers depending on wording.

OpenAI itself flags that self-reports can be inaccurate and its survey sampling was informal. There's no counterfactual meter that tells a researcher what share of today's progress came from AI.

One more detail stands out. Asked who came up with the key ideas behind the GPT-6.2 releases, the researchers OpenAI interviewed credited people, not models, though one said their idea only started working after Codex suggested a change. That fits the data theory: genuinely new research ideas are by definition rare in training data, so models may improve more slowly at ideation than at execution.

Anthropic's "compute multipliers" pitch

Anthropic leaned into this topic during its IPO. In its roadshow presentation, co-founder Tom Brown described the company's edge as "total effective compute": physical chips multiplied by compute multipliers, algorithmic discoveries that let the same hardware train a smarter model. A 4x multiplier means you can train the same model with a quarter of the compute, or a far stronger model with the same compute.

He said these multipliers stack, and that the company believes this could accelerate as it approaches recursive self-improvement, with Claude eventually discovering multipliers on its own.

That pitch is exactly what critics seized on. A coalition published the MisAnthropic Statement the day before the listing. We break that down in the MisAnthropic Statement explained.

Did the "race to the top" happen?

Anthropic has long argued for a "race to the top": lead at the frontier, show safety done well, and competitors will copy you. Amodei has cited examples such as refusing ads and pushing for independent evaluators with deep, employee-like access.

The record is mixed:

  • Independent evaluation exists, but critics say it's constrained. Anthropic has described building outside-evaluation capacity with partners over several years, which skeptics argue is slow relative to its own timelines. OpenAI has promised outside evaluators deeper access, but on terms it controls.
  • Accountability concerns. CAIS director Dan Hendrycks claimed that several safety groups privately agreed with the MisAnthropic Statement but felt unable to sign for fear of losing relationships or funding. That claim is his, and unverified.
  • Transparency is uneven. OpenAI's new report publishes figures Anthropic hasn't, including how much human research effort goes to safety. OpenAI says documented workweeks on safety and alignment rose 77% from January to September, with growth accelerating after its response to the Hugging Face incident in July.

Fairly read, no lab is clearly winning a race to the top on every axis.

The twist: models are getting harder to monitor

This is the part that deserves the most attention.

Reasoning models write intermediate steps, a chain of thought (CoT), before answering. Reading that scratchpad has been one of the best tools for catching a model planning something it shouldn't. OpenAI calls monitorability "a core pillar" of its safety cases and says it has relied on CoT monitoring as a key part of its oversight strategy.

In a November 10 paper, What affects chain-of-thought monitorability across the model lifecycle?, OpenAI reported:

  • With GPT-6 Astra, it observed degraded CoT monitorability in normal settings, a greater ability to evade CoT monitors under adversarial testing, and significantly greater "no-CoT" capability, meaning the model can reason more inside a single forward pass without writing steps down.
  • Better pretraining steadily increases no-CoT capability, and some architectural choices increase it further.
  • Reinforcement learning increased models' ability to control what appears in their chain of thought. OpenAI says it doesn't intentionally reward the CoT and checked for accidental pressure.

Its conclusion is stark: these results indicate that models will become harder to monitor through their CoT as we scale toward more intelligent systems.

An analogy helps. Watching a junior engineer's screen at the end of the week gives you a faithful picture of their work. Reading only a CEO's emails tells you a lot, but not everything they reasoned through. As models get more capable, their written reasoning starts looking more like the emails.

Outside research points the same way. Work from Redwood Research and comments from Anthropic alignment researchers suggest certain architectures that increase reasoning in hidden layers boost capability without needing a visible chain of thought. That's an easy performance win, and a monitoring loss.

What OpenAI says it's doing about it

OpenAI's companion piece on research acceleration says that maintaining human oversight will require new technical and governance breakthroughs, and that its reporting should not be read as an intention to pursue full recursive self-improvement. It has also committed to share more about the relative pace of capability versus safety research.

Meanwhile, some prominent researchers remain alarmed. Longtime alignment researcher Paul Christiano has said in a recent interview that he sees a high risk of AI systems effectively "calling the shots" within the next few years.

A worked example: choosing for a small agency

Imagine a five-person marketing agency deciding which model to standardize on.

Their real work breaks down into four jobs: drafting client emails, summarizing long research PDFs, building quick landing-page prototypes and producing social graphics briefs. Instead of picking whichever model topped this week's headline index, they run each job through Claude, GPT and Gemini with identical prompts and have the account managers score results without knowing which model wrote what.

A typical outcome looks nothing like a single leaderboard. One model writes the most natural client emails. Another handles the 80-page PDFs best and cheapest. A third produces the most original page designs. The sensible decision is to route each job to its winner, and to put a calendar reminder to rerun the same test after the next big release.

That small habit is worth more than any "best model" headline, including this one.

Practical takeaways

  • Test by task, not by leaderboard. The best model for contract review may be the worst for 3D scenes.
  • Price per completed task beats price per token. Efficient models with fewer tokens can be cheaper even at higher list prices.
  • Expect domain surprises. Abilities can jump suddenly where data reaches critical mass, especially in code security and math.
  • Don't rely on the model's "reasoning" display as proof of what it did. Check outputs, logs and actions directly, especially for agents with real permissions.

Counterpoints

  • Benchmarks themselves are data-dependent. Leaderboards reflect which tasks their designers chose.
  • OpenAI's numbers are self-reported and unaudited. Independent outlets note the figures haven't been externally verified.
  • Monitorability isn't only CoT. Labs can also monitor actions and internal activations; the question is whether those tools mature in time.

FAQ

Which AI model is best right now? It depends on the task. Fable 5.2 leads some aggregates, GPT-6.2 Astra leads others, and Gemini 4 Neon wins on some document tasks and value. Test all three on your own work.

Why do benchmarks disagree? Because models are trained on different data. Strength in a domain often reflects how much relevant data a lab trained on, not general intelligence.

How much is AI speeding up OpenAI's research? OpenAI's researchers estimated roughly 3x on a work-of-equal-value basis, with big variation by task. The company cautions that self-reports may be inaccurate.

What is chain-of-thought monitorability? How well humans or monitoring systems can understand a model's intentions by reading its written reasoning. OpenAI says it's degrading as models scale.

Should I worry? Be informed, not frozen. Keep humans approving consequential agent actions, and watch for labs' monitoring and evaluation commitments.

The bottom line

The "best model" question has a grown-up answer now: the best model is the one trained on data most like your work, so test the big three on your own tasks and pick per job.

The deeper story is in OpenAI's fine print. AI is already tripling some researchers' output, success on multi-hour tasks has roughly doubled since March, and the visible reasoning we've relied on to keep an eye on these systems is getting less reliable. Progress is smooth on average and sudden in specifics. The tools to watch it are falling behind.

Sources

The Secret to AI Video That Actually Does What You Want Is a Free 3D Program. Here's the Claude + Blender WorkflowUp next · 11 min readThe Secret to AI Video That Actually Does What You Want Is a Free 3D Program. Here's the Claude + Blender WorkflowPrompting alone can't give AI video perfect camera moves or consistent sets. Building a rough 3D version first can. With Blender's new MCP add-on, Claude or a Higgsfield plugin can build the scene from a sentence, and Seedance 2.5 turns that grey-box animation into a polished, consistent clip. Here's the beginner and advanced workflow, step by step.Keep reading →