GPT-5.5 hits the API: 1M-token context, doubled pricing and High cyber risk
OpenAI published GPT-5.5 on April 23. Instead of commenting on the launch benchmarks, I want to look at what actually matters for anyone integrating the model: the API spec sheet, the pricing table and the risk classification in the system card. That is where the decisions that affect your bill and your architecture live.
What the model is
In the system card, OpenAI describes GPT-5.5 as “a new model designed for complex, real-world work,” aimed at tasks such as writing code, researching online, analyzing information, creating documents and spreadsheets, and moving across tools. It is, in other words, a model built for agents, not just for chat.
On the model page, the key numbers are:
- ID:
gpt-5.5, with the snapshotgpt-5.5-2026-04-23. - Knowledge cutoff: December 1, 2025.
- Context window: 1,050,000 tokens.
- Max output: 128,000 tokens.
- Modalities: text and image in, text out.
- Reasoning effort:
none,low,medium(default),highandxhigh.
Having none and xhigh at the extremes gives plenty of room to tune cost and latency per call. In practice, you can use the same model for a quick classification with no reasoning and for a long agent task with maximum reasoning, changing a single parameter.
Pricing: double GPT-5.4
Per million tokens, GPT-5.5 costs $5 for input, $0.50 for cached input and $30 for output. The page itself lists GPT-5.4 for comparison at $2.50, $0.25 and $15. That is exactly double on every line.
Two rules may catch people off guard when planning a budget:
- Long prompts: once input goes past 272K tokens, pricing rises to 2x on input and 1.5x on output, and that applies to the whole session. The rule covers standard, batch and flex processing.
- Regional endpoints: data residency processing carries a 10% uplift.
In other words, the 1M-token window exists, but using it costs very differently from using the first 272K. For coding agents that keep accumulating context, it is worth watching that threshold closely and investing in prompt caching, which cuts the cost of repeated input to a tenth.
Where and how to use it
The model is available on the Chat Completions, Responses and Batch endpoints. Realtime, Assistants, fine-tuning and the other endpoints are not supported. Among its features, the page lists streaming, structured outputs, function calling, file search, file uploads, image input, web search and prompt caching. In the Responses API, the tool list also includes code interpreter, hosted shell, apply patch, computer use and MCP, which reinforces the agent focus.
Safety: High in bio and in cybersecurity
The system card gives GPT-5.5’s rating under OpenAI’s Preparedness Framework:
- Biological and chemical: High capability.
- Cybersecurity: High capability, but below Critical. OpenAI states the model cannot develop functional zero-day exploits in many hardened critical systems without human intervention, which is its Critical threshold.
- AI self-improvement: does not reach High.
For developers, this has a practical consequence: models rated High usually come with additional safeguards, and some requests in those areas may be refused or monitored. If your product touches offensive security or biology, test those flows before migrating.
My take
GPT-5.5 looks like a clear step toward agents that work for long stretches: a huge context, up to 128K tokens of output, an extra level of reasoning effort and a set of hosted tools covering shell, code editing and computer use.
Pricing, though, is what will decide adoption. Doubling the cost over GPT-5.4 only pays off if the model solves tasks in fewer attempts. My suggestion is not to migrate everything at once: start with the flows where the previous model fails most, measure cost per completed task rather than per token, and keep prompts under 272K tokens whenever possible. The High cybersecurity rating is also a reminder that these tools are getting powerful enough to deserve care in deployment.