GPT-5.5: OpenAI rolled out the smartest model – and it already works for you

On April 23, 2026, OpenAI showed GPT-5.5 — and this is not just another “plus 2% on benchmarks”. According to the company, this is “a new class of intelligence for real work,” and judging by the numbers, this is not a marketing slogan, but a real generational change.

The model is already rolling out in ChatGPT and Codex for Plus, Pro, Business and Enterprise subscribers. In addition, the GPT-5.5 Pro is a heavier version for Pro/Business/Enterprise, sharpened for the most difficult tasks. The API promises "very soon.".

## What is GPT-5.5 and why is it talked about

The key idea of GPT-5.5 is **agency. Instead of explaining each step to her, you throw the model a dirty, multi-component task: “fix this bug”, “analyze 6 months of data”, “make an application on the screen”, and she herself:

• planning a solution,
• uses tool-use (calls tools, browser, terminal),
• check your own conclusion,
• it deals with ambiguities,
• ** and don't give up until the task is closed.*.

OpenAI saw a particularly strong jump in four areas:

• Agentic coding – writing and debugging code for long-term tasks.
• Computer use - the model actually "sees the screen", clicks, types, switches between windows.
• Knowledge work - documents, tables, presentations, discussion.
• Scientific research - scientific problems in biology, mathematics, bioinformatics.

At the same time, which is especially nice: speed remained at the level of GPT-5.4 (latency per token), despite the fact that the model is noticeably smarter. And tokens on the same task in Codex, it spends less, not more.

## Benchmarks: figures that are really impressive

OpenAI compared GPT-5.4, GPT-5.5 Pro, Claude Opus 4.7 and Gemini 3.1 Pro. Here are the key metrics:

• Anthropic itself reported signs of memorization on parts of SWE-Bench Pro tasks.

Particularly significant are Terminal-Bench 2.0 (+7.6 p.p. relative to GPT-5.4), FrontierMath Tier 4 (+8.3 p.p.) and Tau2-bench Telecom - 98% without prompt tuning, that is, the model understands the problem "from the first word".

## Agentic coding: the main feature of the release

This is the part where GPT-5.5 really breaks off. The model holds the context of large systems, talks about ambiguous bugs, tests hypotheses with tools and carefully drags changes through the entire codebase.

A few testimonials from early testers:

• Dan Shipper (Every): "The first coding model to have real conceptual clarity." He had a real-life post-launch bug that took a few days to dismantle one of the best engineers and eventually rewrote part of the system. GPT-5.4 didn't pull that off. GPT-5.5 – *issued the same refactoring as the engineer.
• Pietro Schirano (MagicPath): GPT-5.5 stinked a branch with hundreds of front-end and refactoring changes to the main (which also changed a lot) ** in one sitting in ~20 minutes**.
• NVIDIA Engineer: Losing access to GPT-5.5 is like having a limb amputated.
• Fabian Hedin (CTO Lovable): auth flow, real-time sync, multi-file edits - everything started to get ** the first time, without endless iterations**.

In Codex, the model actually plays the role of an engineer: it writes code, refactors, debagrams, tests, validates – and predicts what will need to be checked/tested, without explicit requests from the user.

## Knowledge work: a model that can use a computer

The same abilities that make GPT-5.5 strong in code make it powerful in everyday computing. The model understands the point better, so the whole cycle of knowledge work itself goes through: find → understand → do → check → give out a ready-made artifact.

In Codex, it is better than GPT-5.4 at generating documents, tables and presentations. In conjunction with computer use, this is the feeling that “the model is really using the computer with me”: sees the screen, clicks, types, switches between applications.

What's telling is that OpenAI is already using it internally:

• 85%+ employees learn Codex every week (engineering, finance, communications, marketing, data science, product).
• Comms team for 6 months of data on speaking-requests built a scoring framework and a Slack-agent, which handles low-risk queries.
• Finance ran 24,771 Form K-1 (71,637 pages) and accelerated work by 2 weeks compared to last year.
• A Go-to-Market employee automated weekly business reports — saving 5-10 hours a week.

There are two modes available in ChatGPT: GPT-5.5 Thinking (for complex tasks, but quickly) and GPT-5.5 Pro (for very hard work – law, business, data science, research).

## GPT-5.5 as a co-scientist

It's interesting here too. The model doesn’t just “answer complex questions,” it **holds the entire research cycle: hypothesis → data → test → interpretation → next step.

Examples:

• GeneBench (multi-stage scientific analysis in genetics and quantitative biology): GPT-5.5 is markedly ahead of GPT-5.4. The tasks that the experts occupy **multi-day projects **, the model solves itself.
• *BixBench (Real Bioinformatics): Leadership among published models.
• Ramsey numbers: The internal custom-bound version of GPT-5.5 found new evidence of an asymptotic fact about off-diagonal Ramsey numbers. The proof was then verified in Lean. It’s no longer “coding” — it’s a real mathematical contribution.
• Derya Unutmaz (Jackson Laboratory): analyzed the gene expression dataset (62 samples, ~28,000 genes) - obtained a detailed research report with key questions and insights. Work that would take the team months.
• Bartosz Naskręcki (mathematician): in 11 minutes one prompt received an application on algebraic geometry with visualization of the intersection of square surfaces and conversion to the Weierstrass model.

## Efficiency of inferencing of the new generation

In order for GPT-5.5 to operate at the latency level of GPT-5.4, OpenAI redefined the inference as a single system rather than a set of isolated optimizations.

Co-designed, trained and served on the NVIDIA GB200 and GB300 NVL72. And what's particularly funny is that Codex and GPT-5.5 itself helped write the infrastructure that serves it. One specific case: Codex analyzed weeks of food traffic and wrote custom heuristics for partitioning and load balancing. The result is +20% to the token generation rate.

## Prices and availability

There is a Fast mode in Codex - ** Tokens are generated 1.5 times faster at 2.5x prices**. For the API, there is Batch and Flex at half the price of the standard, and Priority is 2.5x of the standard.

Yes, GPT-5.5 is more expensive than GPT-5.4 per token. But it’s also smarter and spends fewer tokens on the same task—in Codex, the difference is often offset.

## Security: OpenAI's toughest set of safeguards to date

OpenAI classified GPT-5.5’s biological/chemical and cyber capabilities as High by their Preparedness Framework. The model did not reach the Critical cyber level, but the growth relative to GPT-5.4 is noticeable.

What they added:

• More stringent classifiers of potentially dangerous cyber queries (users may at first consider them “boring” – OpenAI promises to muddy over time).
• Separate protections against repeated abuse.
• **Trusted Access for Cyber: Verified defenders (e.g., organizations protecting critical infrastructure) can access cyber-permissive models like GPT-5.4-Cyber with less restrictions.
• Working with government partners to protect critical infrastructure such as power grids, water supply, and tax data.

Before the release, the model passed a full cycle of evaluations, redteaming (internal and external), targeted tests in biology and cyber, as well as feedback from ~200 trusted early-access partners.

## Bottom line: what does it mean in practice

GPT-5.5 is not just another step. This is the moment when agentic coding and computer use are transformed from demos into a working tool to which you can delegate real multi-hour tasks.

Key takeaways:

1. Agentic coding level state-of-the-art - Terminal-Bench 82.7%, Expert-SWE 73.1%, and real reviews about "limb amputated".
2. GPT-5.5 works at the level of GPT-5.4 over latency, but is noticeably smarter.
3. Token efficiency higher* – Codex often costs less than GPT-5.4.
4. Real computer use - OSWorld-Verified 78.7%, the model actually clicks on the interfaces.
5. **Scientific Contribution — new evidence about Ramsey numbers, BixBench, GeneBench — is a co-scientist, not a toy.
6. Price Adequate to Opportunities - $5/$30 for M tokens for the base version, $30/$180 for Pro.

If you live at Codex or write agents, chances are you’ve already felt it. If not, it’s time to see what GPT-5.5 can take you out of your routine.

Official announcement: openai.com/index/introducing-gpt-5-5