OpenAI releases GPT-5.5

Background

OpenAI released GPT-5.5 in stages rather than through a single switch. On 23 April 2026, GPT-5.5 Thinking and GPT-5.5 Pro became available to paid ChatGPT users ^[1][2]. The same model was also offered in Codex, and it reached the API one day later ^[1]. That order put the chat product first, then the coding product and developer access. For a company shipping a model to different audiences, the sequence matters: consumers see the model in ChatGPT, while developers encounter it through the API, where pricing and context limits apply.

Release and rollout

GPT-5.5 and GPT-5.5 Pro became available in the API on 24 April 2026 ^[3]. The rollout then widened again on 5 May 2026, when a faster GPT-5.5 Instant became the default for all ChatGPT users ^[1]. On 7 May 2026, OpenAI followed with GPT-5.5-Cyber, a cybersecurity variant released in limited preview ^[1]. By that point the release had several distinct surfaces: paid ChatGPT, the API, Codex, a default chat model for all users, and a specialized security preview. Each surface had its own audience and access rules.

Performance and efficiency

OpenAI reported state-of-the-art scores for GPT-5.5, including 82.7% on Terminal-Bench 2.0 ^[1]. It also said the model matched GPT-5.4's per-token latency in real-world serving while performing at a much higher level of intelligence, and that it used significantly fewer tokens on the same Codex tasks ^[3]. Those claims mix capability and operating cost. A benchmark score measures how well a model does on a defined test; latency measures how long a response takes; token count measures how much text the model processes. In production, all three matter. A model that is more capable but much slower or much more expensive can be harder to deploy. OpenAI's statement was that GPT-5.5 did not trade away per-token latency to reach a higher level of intelligence, and that it needed fewer tokens for the same coding work ^[3].

Pricing, access, and safeguards

For API developers, gpt-5.5 was priced at $5 per million input tokens and $30 per million output tokens, with a 1 million token context window ^[3]. gpt-5.5-pro was priced at $30 per million input tokens and $180 per million output tokens ^[3]. A context window is the amount of text a model can take into account at once; a token is a chunk of text, roughly a word piece, and it is also the unit used for API billing. The gap between input and output prices means developers pay different rates for text sent to the model and text generated by it.

In Codex, GPT-5.5 was offered on Plus, Pro, Business, Enterprise, Edu, and Go plans with a 400,000 token context window ^[3]. A Fast mode generated tokens 1.5 times faster at 2.5 times the cost ^[3]. The Codex terms therefore differed from the API terms: a smaller context window and plan-based access, plus an optional speed setting with a clear price multiplier.

OpenAI released the model with what it called its strongest set of safeguards to date, after feedback from nearly 200 trusted early-access partners ^[3]. Early-access partners are outside organizations that use a model before general release and can provide feedback. The number is a concrete measure of how much outside testing OpenAI says preceded the launch.

What the release changed

The most visible change for ChatGPT users was the 5 May 2026 switch that made GPT-5.5 Instant the default for all users ^[1]. For developers, the API release on 24 April 2026 set the model's price and context terms ^[3]. For Codex users, the release added GPT-5.5 on several plans, with a separate context limit and a paid Fast mode ^[3]. The cybersecurity preview added a narrower, security-focused use case ^[1].

The launch also tied performance claims to practical serving claims. OpenAI reported both a benchmark result and a latency/token comparison, and it published prices alongside the API release ^[1][3]. That combination is what a model release looks like when it is aimed at both end users and builders: a chat interface, a coding product, an API price list, and a safety process all arrive together. The later Instant and Cyber releases extended the same model family beyond the first paid ChatGPT launch ^[1].