OpenAI releases GPT-5.5: agentic model stronger than GPT-5.4, but API pricing doubled

OpenAI on April 23 introduced GPT-5.5 and positioned the release not as a routine model update, but as the foundation for working AI agents. The company claims the new version plans better, uses tools, and completes complex multi-step tasks without constant human prompts.

What's the Bet

The main idea of the release is that GPT-5.5 should work not as a chatbot for single answers, but as an executor of long tasks on a computer. According to OpenAI's description, the model is stronger in agent programming, interface work, data analysis, document preparation, and research tasks.

Instead of a scenario where users manually spell out each step, companies are offered a way to give the model "messy" multi-part tasks and let it build the plan itself, check intermediate results, and move toward the final answer. OpenAI makes a special point that the growth in capabilities didn't slow down the model in production. GPT-5.

5, according to the company, maintains GPT-5.4's latency per token but spends fewer tokens on the same tasks in Codex. The model was also developed and deployed alongside NVIDIA GB200 and GB300 NVL72 systems, showing that frontier model releases are now tied not only to algorithms but also to infrastructure.

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

Greg Brockman.

Where the Growth Is Visible

OpenAI's strongest argument is benchmarks related not to academic questions but to practical agent work. GPT-5.5 showed notable gains where models need to not just answer but plan a sequence of actions, invoke tools, maintain long context, and complete a task. At the same time, the release doesn't look like an unconditional victory in all categories: in some external assessments, competitors maintain strong positions, especially where tool orchestration via MCP is important.

• 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% for solving real GitHub issues in a single pass.
• MRCR v2 on 1M tokens: 74.0% versus 36.6% for GPT-5.4, almost a twofold jump in long context.
• BrowseComp in Pro version: 90.1%, but in MCP Atlas the lead remains with Claude Opus 4.7, and GPT-5.5 had no published result there.

Price and Access

The most controversial part of the release is the economics. For API, OpenAI set the price at $5 per million input tokens and $30 per million output tokens, while GPT-5.4 rates were half that: $2.

50 and $15 respectively. GPT-5.5 Pro is even pricier—$30 for input and $180 for output.

OpenAI argues that comparing only token prices is no longer sufficient: the model solves the same tasks with fewer tokens, so the actual cost of a working scenario, according to the company, grows not twofold but roughly 20%. Access to GPT-5.5 first went to ChatGPT and Codex for paid tiers Plus, Pro, Business, and Enterprise, with API access following on April 24.

Within OpenAI itself, the release is presented as already a working tool: the company says Codex is used weekly by more than 85% of employees. Examples include processing six months of public speaking requests, analyzing 24,771 tax forms K-1 totaling 71,637 pages, and automating weekly business reports, saving employees 5–10 hours per week.

What This Means

GPT-5.5 shows where the market is heading: models are now sold not as "another smart chat" but as a layer for autonomous digital work. For business, the question is no longer just which model is smarter, but how much a completed task costs, factoring in retries, checks, and tool integration. This is the field where OpenAI intends to justify its higher price tag.