OpenAI opens GPT-5.5 API with 1M context and Responses support

• After spending April 23 in ChatGPT and Codex only, GPT-5.5 and GPT-5.5 Pro flipped live in the API on April 24, with OpenAIDevs' API details specifying Responses and Chat Completions support and OpenAI's update confirming the rollout.
• OpenAIDevs' API details gave GPT-5.5 a 1M context window in the API, while scaling01's pricing screenshot showed the base model at $5 input and $30 output per million tokens.
• OpenAI kept the pitch centered on long-running agent work, with OpenAIDevs' launch thread describing planning, tool use, and ambiguity recovery, while OpenAIDevs' token-efficiency thread claimed GPT-5.5 completes Codex tasks with fewer tokens than GPT-5.4.
• The same-day distribution story was almost as notable as the model itself: OpenRouter's rollout, perplexity_ai's Perplexity rollout, cursor_ai's Cursor rollout, and vercel_dev's AI Gateway rollout all landed within hours.
• Early hands-on reports clustered around end-to-end execution rather than single benchmark wins, with cognition's Devin preview, warpdotdev's speedup claim, and OpenAIDevs' Perplexity case study all emphasizing longer autonomous runs and lower token burn.

You can read OpenAI's launch post, skim the latest model guide, and Simon Willison already pulled out one of the more practical migration notes in his GPT-5.5 prompting guide. There is also a surprisingly useful Vercel changelog, a public CursorBench post, and an active Hacker News thread where the debate quickly moved from launch hype to pricing, access paths, and safety policy.

OpenAI's April 23 messaging still framed GPT-5.5 as a ChatGPT and Codex launch, with the API only "coming soon" in OpenAIDevs' launch post, thsottiaux's announcement, and Sam Altman's launch thread. Roughly 24 hours later, OpenAIDevs' API launch and OpenAIDevs' API details changed that to general API availability, then Sam Altman's API post repeated the update from his own account.

That one-day gap matters because it clarifies what actually shipped first. GPT-5.5 was not an API-first release. It was a Codex and ChatGPT release that only became a full developer platform launch once Responses and Chat Completions support went live the next day.

The most concrete API change in OpenAIDevs' API details is surface area: GPT-5.5 is available through both Responses and Chat Completions, while GPT-5.5 Pro is available through Responses only. The same post also set the headline context limit at 1 million tokens.

That puts the launch in line with OpenAI's current developer stack rather than as a one-off model slug. The latest model guide and OpenAI's launch post both frame GPT-5.5 as the default frontier model for complex coding and professional work, not as a narrow Codex-only checkpoint.

The linked docs also make an important migration point that Simon Willison highlighted in his prompting guide writeup: OpenAI recommends treating GPT-5.5 as a new model family to tune for, not as a drop-in prompt swap for GPT-5.2 or GPT-5.4.

Pricing moved up versus GPT-5.4. scaling01's pricing screenshot showed GPT-5.5 at $5 input and $30 output per million tokens, and scaling01's pricing post put GPT-5.5 Pro at $30 input and $180 output.

OpenAI's counter-argument is token efficiency, not cheaper list price. OpenAIDevs' token-efficiency thread said GPT-5.5 gets better Codex results with fewer tokens than GPT-5.4, while Sam Altman's launch thread said it matches GPT-5.4 on per-token speed but uses significantly fewer tokens per task.

That same framing showed up immediately in partner rollouts. OpenAIDevs' Perplexity case study claimed 56% fewer tokens in Perplexity Computer workflows, and AravSrinivas' orchestration note tied the model switch directly to lower Computer credit burn.

OpenAI's own benchmark card in OpenAIDevs' launch post is very obviously built to sell GPT-5.5 as an agent model first.

The flagship deltas versus GPT-5.4 from OpenAIDevs' launch post and OpenAI's launch post were:

• Terminal-Bench 2.0: 75.1% to 82.7%
• OSWorld-Verified: 75.0% to 78.7%
• Toolathlon: 54.6% to 55.6%
• FrontierMath Tier 4: 27.1% to 35.4%
• CyberGym: 79.0% to 81.8%
• GDPval: 83.0% to 84.9%

Two details stand out in the evidence pool. First, OpenAI kept repeating that GPT-5.5 is strongest on longer workflows, command-line work, and persistence across research loops in OpenAIDevs' extended thread. Second, some third-party chatter immediately pushed on benchmark variance, with ValsAI's Terminal Bench correction revising one external Terminal-Bench configuration and nrehiew_'s variance post asking why official and external runs were so far apart.

That leaves the headline chart intact, but it also means the API story is arriving with an unusually live debate over harness setup and reasoning-effort settings.

The rollout spread far beyond OpenAI's own surfaces within hours.

A quick inventory from the evidence:

• OpenRouter shipped both GPT-5.5 and GPT-5.5 Pro on day one, per OpenRouter's rollout.
• Perplexity added GPT-5.5 for Max subscribers and made it the default orchestrator for Computer, per perplexity_ai's rollout and AravSrinivas' default-switch post.
• Cursor added GPT-5.5 and immediately advertised it as the top model on CursorBench at 72.8%, per cursor_ai's Cursor rollout.
• Vercel AI Gateway exposed both slugs, openai/gpt-5.5 and openai/gpt-5.5-pro, per vercel_dev's AI Gateway rollout and the linked Vercel changelog.
• GitHub Copilot started rolling it out the same day, according to gdb's Copilot post, pierceboggan's rollout note, and code's VS Code post.
• Devin exposed GPT-5.5 as an Agent Preview, per cognition's Devin preview.
• Warp's agent added support and reported a roughly 40% speedup over GPT-5.4 on one migration task, per warpdotdev's speedup claim.
• Zed, Windsurf, Venice, Capy, Tembo, Cline, and AI Gateway all posted same-day availability as well, via zeddotdev's rollout, Windsurf's rollout, AskVenice's rollout, capydotai's rollout, tembo's rollout, and cline's Terminal-Bench post.

This is the part of the launch that felt like Christmas come early for coding agent nerds. GPT-5.5 was not trapped inside one vendor shell for long.

The early user reports were much more consistent on workflow feel than on abstract frontier bragging. The common claim was that GPT-5.5 carries tasks further without extra steering.

Examples from the first day:

• cognition's Devin preview said GPT-5.5 runs longer and more autonomously than prior GPT models they tested, including production issue investigation end to end.
• haider1's backend test said GPT-5.5 followed a tangled payment-flow bug across multiple files and fixed the side effects without extra hand-holding.
• bridgemindai's workflow report and bridgemindai's side-by-side bug test both described one-shot debugging wins over Claude Opus 4.7 on specific repo tasks.
• omarsar0's Codex report said the switch from Claude Code to Codex was smoother than expected, partly because GPT-5.5 was sharper, terser, and better with Chrome CDP style workflows.
• willdepue's Codex post said the model looked less impressive on evals than in actual Codex use on complex technical projects.
• nummanali's autonomy post claimed a 10-plus hour uninterrupted work loop on MLX optimization and an inference engine build.

The personality shift also came up repeatedly. haider1's tone post described it as more composed and more like GPT-4o in tone, while onusoz's schema-design note and BlancheMinerva's complaint show that terser style did not land uniformly.

One of the least-advertised launch details came from the agent ecosystem, not OpenAI's main posts. sqs' Amp disclosure said GPT-5.5 use in Amp's deep mode currently falls under OpenAI's Safety Retention Policy, with non-zero data retention in about 0.05% of cases that OpenAI's classifier flags as severe cybersecurity abuse risk.

That disclosure helps explain why API access lagged the ChatGPT and Codex launch by a day, and why the Hacker News thread quickly filled with questions about cyber policy, malware-analysis workflows, and access constraints. It also gives the ecosystem rollout a sharper edge: GPT-5.5 arrived everywhere fast, but not every agent wrapper is equally comfortable with the safety terms attached to it yet.