# OpenAI's new flagship GPT-5.5 got smarter without getting slower!

Hi everyone, it's me, Shii-chan! I've got a really big one for you today: OpenAI just announced its new flagship model, GPT-5.5. I couldn't stop grinning while reading it, so let's dig in together!

## What was announced?

This one comes from OpenAI's News. OpenAI released GPT-5.5, which it calls its smartest and most intuitive model to use yet. Starting April 23, 2026, it began rolling out to Plus, Pro, Business, and Enterprise users in ChatGPT and Codex, with the even smarter GPT-5.5 Pro going to Pro, Business, and Enterprise users. The next day, April 24, both GPT-5.5 and GPT-5.5 Pro also became available in the API.

GPT-5.5 got especially strong in four areas: agentic coding, computer use, knowledge work, and early scientific research. You can hand it a messy, multi-part task and trust it to plan, use tools, check its own work, and keep going until the job is done.

## Why it matters

For a long time, smarter usually meant slower. Bigger, more capable models tended to take longer to serve.

GPT-5.5 breaks that pattern. It matches GPT-5.4's per-token latency in real-world serving while operating at a much higher level of intelligence. On top of that, it uses significantly fewer tokens to finish the same Codex tasks, so it's efficient as well as capable. On Artificial Analysis's Coding Index, it delivers state-of-the-art intelligence at half the cost of competitive frontier coding models.

## What changes

Coding is the clearest place to see it. On Terminal-Bench 2.0, which tests complex command-line workflows, GPT-5.5 hits 82.7% for state-of-the-art accuracy, and on SWE-Bench Pro, which measures real GitHub issue resolution, it reaches 58.6% — all while using fewer tokens than GPT-5.4.

It's dependable beyond code too. It's better at making documents, spreadsheets, and slides, scoring 84.9% on [GDPval](https://openai.com/index/gdpval/), a benchmark covering knowledge work across 44 occupations, and 78.7% on OSWorld-Verified, which checks whether a model can operate real computer environments. Inside OpenAI, more than 85% of the company now uses Codex every week.

Science moved forward as well. GPT-5.5 clearly beats GPT-5.4 on GeneBench, an eval for genetics and quantitative-biology data analysis, and an internal version with a custom harness found a new proof about Ramsey numbers, a central topic in combinatorics, which was later verified in Lean. Contributing an actual mathematical argument, not only code or explanation, is a fresh kind of result.

Early testers were lively about it. One engineer at NVIDIA even said:

> Losing access to GPT-5.5 feels like I've had a limb amputated.

Dan Shipper, founder of Every, described it this way:

> The first coding model I've used that has serious conceptual clarity.

## Dive Deep

The speed comes from rethinking inference as one integrated system rather than a pile of separate optimizations. GPT-5.5 was co-designed for, trained with, and served on NVIDIA's GB200 and GB300 NVL72 systems. The fun part: Codex and GPT-5.5 itself helped with the speedups. After analyzing weeks of production traffic, Codex wrote custom heuristic algorithms for load balancing and partitioning, and token generation speeds went up by over 20%. The model helped improve the very infrastructure that serves it.

Cybersecurity gets careful handling. OpenAI treats GPT-5.5's biology/chemistry and cyber capabilities as High under its Preparedness Framework (though it did not reach Critical), and it added stricter classifiers around higher-risk activity. At the same time, to keep power in defenders' hands, it's opening up advanced capabilities with fewer restrictions to verified users through [Trusted Access for Cyber](https://openai.com/index/scaling-trusted-access-for-cyber-defense/), starting with Codex. The full safety details live in the [system card](https://deploymentsafety.openai.com/gpt-5-5).

Let's look at pricing too. In the API, gpt-5.5 is $5 per 1M input tokens and $30 per 1M output tokens, with a 1M-token context window. The higher-accuracy gpt-5.5-pro is $30 input and $180 output per 1M tokens. Batch and Flex run at half the standard rate, and Priority at 2.5x. In Codex, it's available with a 400K context window, plus a Fast mode that generates tokens 1.5x faster for 2.5x the cost. Check the [pricing page](https://openai.com/api/pricing/) for the full details.

## Wrap-up

- GPT-5.5 is OpenAI's new flagship, with big gains in agentic coding, computer use, knowledge work, and scientific research.
- It keeps GPT-5.4's real-world latency while getting smarter and more token-efficient, scoring 82.7% on Terminal-Bench 2.0 and 84.9% on GDPval.
- It's rolling out in ChatGPT and Codex, with the API opening the next day, starting at $5 input / $30 output per 1M tokens.
- Its cyber capability is rated High with tighter safeguards, but defensive work gets opened up through Trusted Access.

If you're an engineer who wants to hand a long, messy task to an AI, or someone who moves across code, docs, and research all day, this one is really for you!