OpenAI releases GPT-5.4 Pro: new records on ARC-AGI-2, FrontierMath and logic tasks

OpenAI has unveiled GPT-5.4 Pro, a new flagship version that takes a clear step forward not only on dry metrics but also in how the model behaves on hard tasks. Where the Pro suffix was often seen as just a pricier tier, here it looks like a separate level of quality.

A leap in the tests

The headline figure from the review is 83.3% on ARC-AGI-2, versus 54% for the previous version. For a class of tasks in which models must not guess a pattern but actually infer a rule from examples, that is a very sharp rise. The result matters not in itself but as a signal: OpenAI has strengthened the model's ability to work where superficial heuristics break down and where the structure of the task has to be held all the way to the answer.

Progress on FrontierMath is just as telling. This problem set was long considered almost off-limits for mainstream AI models. Such tests used to serve mostly to demonstrate limits; now they increasingly show how well a model can build a long chain of reasoning without losing a step. Against that backdrop, GPT-5.4 Pro looks not merely faster or more convenient but noticeably deeper in its intellectual profile.

Testing beyond the tables

The review's authors did not stop at benchmarks and ran the model through more applied scenarios. Instead of abstract percentages, they looked at how GPT-5.4 Pro handles tasks that require combining logic, planning and attention to detail. This format is more interesting than an ordinary table because it shows not a single strong skill but the model's behavior under load, when an error in the middle of the chain ruins the whole result at the first failure.

• A logic puzzle with servers and dependencies between nodes
• Tasks that require holding several conditions at once
• Scenarios that involve finding a non-obvious path to a solution
• A full stealth simulator on a canvas, where planning and sequencing of actions matter

According to the description of the tests, the new version's strength is not only the correct final answer but also stability during the solution. The model loses context less often, holds constraints better and slides into random guesses less quickly when a task goes beyond the standard examples in the training corpus. For users this matters more than a record number on a leaderboard: this is how a real quality gain feels in daily work.

What was surprising in its behavior

One of the most telling episodes in the review concerns not math but the model's research behavior. While solving a problem, GPT-5.4 Pro found a forgotten 2011 scientific paper on the internet and used it as a shortcut to the answer. On one hand, this is impressive: the model does not just cycle through memorized patterns but can find outside support where it genuinely helps. On the other, such an episode immediately raises the question of the limits of autonomy and of verifying the sources it finds.

This is an important shift in the very nature of interacting with AI. Users increasingly work not with a talking encyclopedia but with a system that combines reasoning, search and adaptation of strategy to the task. That is why comparing models only by token count or response speed no longer explains their real value well. What becomes key is something else: how reliably the model can think, search and avoid breaking on an unconventional route.

What it means

The bar for top models has risen again, and GPT-5.4 Pro shows that the next stage of competition is no longer about basic text coherence but about depth of reasoning and stability in complex scenarios. For the market, this means a faster move from the "smart chatbot" to a working tool for analysis, math, programming and multistep tasks, where humans previously still had to back up the model at almost every step.