GPT-5.5: Benchmarks, Rollout, API Pricing, and Cyber Safeguards

Same latency, higher scores, fewer tokens

OpenAI's stated goal for GPT-5.5 was intelligence without a speed cost: the company says it matches GPT-5.4's per-token latency in production serving while scoring higher across coding, computer-use and knowledge-work evaluations, and using fewer tokens to finish the same Codex tasks. On Terminal-Bench 2.0, a command-line planning and tool-use benchmark, GPT-5.5 scored 82.7 percent against GPT-5.4's 75.1 percent. On SWE-Bench Pro, a real-world GitHub issue benchmark, it reached 58.6 percent versus 57.7 percent, though Anthropic reported signs of memorization on a subset of SWE-Bench Pro problems. On the GDPval knowledge-work benchmark it scored 84.9 percent versus 83.0 percent, and on OSWorld-Verified, which tests operating a real computer environment unassisted, 78.7 percent versus 75.0 percent. [1]

OpenAI attributes the flat latency to work done jointly by Codex and GPT-5.5 on the serving stack itself: Codex analyzed weeks of production traffic to rewrite the heuristics that partition and load-balance requests across a GPU, replacing a fixed number of static chunks with something tuned to actual traffic shapes, which OpenAI says raised token-generation speed by more than 20 percent. The model was co-designed for and served on NVIDIA GB200 and GB300 NVL72 systems. [1][1]

Available three ways, on three different days

GPT-5.5 and the higher-effort GPT-5.5 Pro variant began rolling out on April 23, 2026 to Plus, Pro, Business and Enterprise users in ChatGPT and Codex. OpenAI's announcement said the API would follow separately, writing that it would bring GPT-5.5 and GPT-5.5 Pro to the API "very soon" rather than that same day, citing different safeguard requirements for API deployments. Microsoft's own announcement, published the same day, said GPT-5.5 would reach general availability in Microsoft Foundry "tomorrow," meaning April 24, a day after the ChatGPT and Codex rollout. [1][2]

Atlas interpretation: OpenAI's shipped API pricing figures (5 dollars per million input tokens and 30 dollars per million output tokens for GPT-5.5, a 1 million token context window, and 30 dollars per million input and 180 dollars per million output for GPT-5.5 Pro) appeared in the same announcement that said direct API access was still pending, so the numbers describe a launch OpenAI had already priced but had not yet turned on. [1]

Tighter classifiers, and a wider door for vetted defenders

OpenAI classified GPT-5.5's cybersecurity and biological/chemical capabilities as High, one level below Critical, under its Preparedness Framework, and said testing showed a step up in cyber capability compared with GPT-5.4. Alongside the model, OpenAI said it was deploying stricter classifiers on cyber-risk requests, building on cyber-specific safeguards it first introduced with GPT-5.2. At the same time, it expanded Trusted Access for Cyber, starting with Codex, to give verified users meeting certain trust signals fewer restrictions when using GPT-5.5 for cybersecurity work, and said organizations defending critical infrastructure could apply for further access to cyber-permissive models. [1]

Atlas interpretation: The two moves point the same direction as OpenAI's later, more permissive GPT-5.6-Cyber release in August: restricting a general-purpose model's cyber behavior while building a separate, gated path for vetted users to get more out of it, rather than resolving the underlying tension between capability and misuse inside one model available to everyone. [1]

A codename, and a company already living on the product

Big Technology reported that OpenAI's internal codename for the release was "Spud," and that OpenAI president Greg Brockman described it as "a beginning point" for a planned series of models built around taking action on a user's behalf, rather than an endpoint in itself. OpenAI's own announcement said more than 85 percent of the company was using Codex weekly by the time of release, across engineering, finance, communications, marketing, data science and product management, and gave specific internal examples: reviewing 24,771 K-1 tax forms in Finance, and building a request-scoring and Slack-routing tool in Communications. [3][1]