# What is GPT-5.5, and what changes for people who use ChatGPT and Codex?

Date: 2026-04-24 · Author: Robert Scales · https://guidetoartificialintelligence.com/en/learn/what-is-gpt-5-5/

**Answer:** GPT-5.5, codenamed "Spud," is OpenAI's newest and most capable model, released on April 23, 2026 to paying ChatGPT and Codex users, with a stronger GPT-5.5 Pro for the top plans. It is built to take on messy, multi-step work such as coding, research and office documents with less hand-holding, while answering as quickly as GPT-5.4 and using fewer tokens.

Yesterday, Thursday, April 23, 2026, OpenAI released **GPT-5.5**, known inside the company by the codename "Spud." OpenAI calls it its smartest model so far and the most intuitive to use. The pitch is simple: instead of guiding the AI one small step at a time, you can describe a tangled, many-part job and let the model plan it, use tools, check its own work and keep going until it is done.

The launch came one week after Anthropic released its own latest model, and it shows how quickly the leading labs are now shipping. Greg Brockman, an OpenAI co-founder, told reporters in a briefing that the model is a "faster, sharper thinker for fewer tokens" than GPT-5.4, the version it replaces at the top of the lineup.

## What OpenAI released, and who gets it

There are two versions. **GPT-5.5** (shown as GPT-5.5 Thinking in ChatGPT) is the main model. **GPT-5.5 Pro** spends more effort per question and is aimed at the hardest, highest-accuracy work, such as legal, business or data-science analysis.

| Where | Who has access on launch day | Details |
|---|---|---|
| ChatGPT, GPT-5.5 Thinking | Plus, Pro, Business, Enterprise | Rolling out now |
| ChatGPT, GPT-5.5 Pro | Pro, Business, Enterprise | Rolling out now |
| Codex (OpenAI's coding agent) | Plus, Pro, Business, Enterprise, Edu, Go | 400,000-token context window; optional Fast mode, 1.5 times faster for 2.5 times the cost |
| API (for developers) | Not yet | Promised "very soon": $5 per million input tokens, $30 per million output, 1-million-token context window |

A **token** is a small chunk of text, often part of a word. The **context window** is how much text the model can hold in view at once: your instructions, the files you give it and its own work so far.

The developer version is waiting for one reason: safety. OpenAI says serving the model through its API, where any company can build it into products at large scale, calls for different safeguards, and it is still working those out with partners. Axios reports that API access will follow once extra cybersecurity guardrails are in place. GPT-5.5 Pro will cost $30 per million input tokens and $180 per million output tokens when it reaches the API.

## How it is meant to work

The word OpenAI leans on is **agentic**. An agentic model does not just reply; it takes actions toward a goal over many steps, such as running code, searching the web, opening files or clicking through software. OpenAI says GPT-5.5's biggest gains are in four areas: coding through agents, operating a computer, everyday office work, and early scientific research.

Three ideas help explain the improvement:

- **Better reading of intent.** The model is meant to grasp what you are really after sooner, so it needs fewer clarifying rounds.
- **Persistence.** It stays on long tasks without stopping early. Cursor, a popular coding editor, says it keeps working noticeably longer than GPT-5.4.
- **Efficiency.** Bigger models are usually slower, but OpenAI says GPT-5.5 answers at the same speed per token as GPT-5.4 and needs fewer tokens to finish the same Codex tasks. OpenAI built and runs it on NVIDIA's GB200 and GB300 NVL72 systems, large racks of linked graphics chips, and says Codex itself helped rewrite how work is split across those chips, raising generation speed by more than 20%.

OpenAI's own test results back up the claims on paper. These are scores the company reported, not independent measurements:

| Test (what it measures) | GPT-5.5 | GPT-5.4 |
|---|---|---|
| Terminal-Bench 2.0 (multi-step command-line work) | 82.7% | 75.1% |
| OSWorld-Verified (using a real computer on its own) | 78.7% | 75.0% |
| GDPval (professional tasks across 44 jobs, wins or ties against experts) | 84.9% | 83.0% |

OpenAI also says an internal version of the model, given a custom setup, helped find a new proof about Ramsey numbers, a classic topic in mathematics about when order must appear in a large network. The proof was then checked with Lean, a program that verifies mathematical proofs line by line.

## Safety: why the API is held back

OpenAI rates GPT-5.5 as **High** in both biology and chemistry and in cybersecurity under its Preparedness Framework, the internal rulebook it uses to judge serious risks. That is one step below its top "Critical" level. In practice it means more filters: OpenAI is adding stricter **classifiers**, small automatic checkers that flag risky requests, and warns that some people may find them annoying at first while they are tuned.

For security professionals who need fewer refusals, OpenAI points to its **Trusted Access for Cyber** program, where verified defenders can apply for a less restricted version. Nearly 200 early-access partners tested GPT-5.5 before the release.

## Why it matters

For everyday users, the main change is how you can work with ChatGPT and Codex: handing over a whole task instead of steering every step. For businesses, Axios points to cost. As AI agents run longer jobs, the price of each token starts to matter as much as raw intelligence, which is why OpenAI stresses doing more with fewer tokens. GPT-5.5 is priced higher per token than GPT-5.4, so whether it ends up cheaper depends on how many fewer tokens it really needs for your work.

The release also says something about the race. OpenAI ships a new top model a week after Anthropic did, and it tests GPT-5.5 directly against Claude Opus 4.7 and Gemini 3.1 Pro in its own tables. If that rhythm holds, new top models may keep arriving only weeks apart.

> **Try it:** Take a task you usually break into several chats, such as turning meeting notes into a summary, a task list and a draft email. Give it to GPT-5.5 in one message, as in the example below, and compare the result with doing it step by step.

## How should you prompt GPT-5.5 to get the most from it?

Describe the outcome, not every step. Because GPT-5.5 is designed to plan on its own, a good prompt reads like a clear work order: the goal, the material, the limits and what "finished" looks like.

```prompt
Here are my notes from today's project meeting (pasted below).
Goal: send the team a clear follow-up by end of day.
1. Write a five-line summary of what we decided.
2. Make a table of action items with owner and due date.
3. Draft a short, friendly email to the team using both.
Flag anything in the notes that is unclear instead of guessing.
```

The last line matters most. A model that keeps going on its own can also fill gaps with confident guesses, so ask it to point out uncertainty. For long jobs in Codex, ask it to run tests or check its results before reporting back, and review the final work yourself before you rely on it.

## Sources

- Introducing GPT-5.5: https://openai.com/index/introducing-gpt-5-5/
- OpenAI releases "Spud" GPT-5.5 model: https://www.axios.com/2026/04/23/openai-releases-spud-gpt-model