OpenAI releases GPT-5.5: an agentic model stronger than GPT-5.4, but with API prices twice as high

On April 23, OpenAI introduced GPT-5.5, presenting the release not as a routine model update but as a foundation for working AI agents. The company claims the new version plans better, uses tools more effectively and carries multipart tasks through to a result without constant prompting from a human.

What OpenAI is betting on

The central idea of the release is that GPT-5.5 should work not as a chatbot for one-off answers, but as an executor of long tasks on a computer. According to OpenAI, the model is stronger in agentic coding, working with interfaces, data analysis, document preparation and research tasks. Instead of a scenario where the user spells out every step by hand, the company suggests handing the model "messy" multipart tasks and letting it build the plan, check intermediate results and move toward the final answer on its own.

OpenAI places separate emphasis on the claim that the gain in capability did not slow the model down in production. GPT-5.5, the company says, keeps GPT-5.4's per-token latency but uses fewer tokens for the same tasks in Codex. The model was also developed and deployed together with NVIDIA GB200 and GB300 NVL72 systems, which shows that frontier model releases now depend not only on algorithms but increasingly on infrastructure.

"This is a real step forward toward the kind of computing we expect in the future," said

Greg Brockman.

Where the gains show

OpenAI's strongest argument is its benchmarks, which test practical agentic work rather than academic questions. GPT-5.5 shows a notable gain where the model has to do more than answer: plan a sequence of actions, call tools, hold a long context and finish the task. The release is not an unconditional win in every category, though: competitors keep strong positions in some external evaluations, especially where tool orchestration via MCP matters.

• Terminal-Bench 2.0: 82.7% versus 75.1% for GPT-5.4 and 69.4% for Claude Opus 4.7.
• SWE-Bench Pro: 58.6% on tasks that involve resolving real GitHub issues in a single pass.
• MRCR v2 at 1 million tokens: 74.0% versus 36.6% for GPT-5.4, nearly a twofold jump in long context.
• BrowseComp for the Pro version: 90.1%, but in MCP Atlas the lead stayed with Claude Opus 4.7, and GPT-5.5 had no published result there.

Price and availability

The most contentious part of the release is the economics. For the API, OpenAI set the price at $5 per million input tokens and $30 per million output tokens, while GPT-5.4's rates were half that: $2.5 and $15 respectively. GPT-5.5 Pro costs even more: $30 for input and $180 for output.

OpenAI objects that comparing price per token alone is no longer enough: the model solves the same tasks with fewer tokens, so the actual cost of a working scenario, by the company's estimate, rises not twofold but by roughly 20%.

Access to GPT-5.5 first arrived in ChatGPT and Codex for the paid Plus, Pro, Business and Enterprise plans, and API access followed on April 24. Inside OpenAI itself, the release is presented as an already working tool: the company says more than 85% of employees use Codex every week. Its examples include processing six months of public speaking requests, analyzing 24,771 K-1 tax forms totaling 71,637 pages, and automating weekly business reports, which saves employees 5-10 hours a week.

What this means

GPT-5.5 shows where the market is heading: models are no longer sold as "yet another smart chat" but as a layer for autonomous digital work. For businesses, the question is no longer only which model is smarter, but how much a completed task costs once retries, checks and tool integration are counted. That is the ground on which OpenAI intends to justify the higher price.