Overview
GPT-5.5 is OpenAI's smartest and most intuitive model to date, released on April 23, 2026, with the strongest gains in agentic coding, computer use, knowledge work, and early scientific research. Unlike the incremental GPT-5.4 update earlier this year, GPT-5.5 is a ground-up retraining focused on four domains where progress depends on reasoning across context and taking action over time: agentic coding, computer use, knowledge work, and early scientific research. OpenAI positions it as a step toward what co-founder Greg Brockman described as a future "super app," consolidating gains in long-horizon planning, tool use, and self-verification into a single mainline model available across ChatGPT and Codex.
What's New
First Fully Retrained Base Model Since GPT-4.5
OpenAI describes GPT-5.5 as a new model for complex, real-world work; the official launch does not state that it is the first fully retrained base model since GPT-4.5. The model is designed for agentic behavior out of the box: it "takes a sequence of actions, uses tools like browsing the web, writing code, running scripts, or operating software, checks its own work, and keeps going until the task is finished." This marks a shift from the previous 5.x line toward more agentic, tool-using workflows, but OpenAI does not state that GPT-5.5 is a ground-up architectural retraining.
Agentic Coding and Terminal-Bench 2.0
On Terminal-Bench 2.0—which tests complex command-line workflows requiring planning, iteration, and tool coordination—GPT-5.5 reaches 82.7%, up 7.6 points from GPT-5.4's 75.1%. In OpenAI's published benchmark table, GPT-5.5 leads Claude Opus 4.7 and Gemini 3.1 Pro on Terminal-Bench 2.0: Claude Opus 4.7 scores 69.4% and Gemini 3.1 Pro scores 68.5% on the same benchmark. On SWE-Bench Pro, which evaluates real GitHub issue resolution, GPT-5.5 reaches 58.6% vs GPT-5.4's 57.7%—a narrower gain, and trailing Claude Opus 4.7's 64.3%, though OpenAI's eval table footnote says labs have noted evidence of memorization on this public eval.
Computer Use and OSWorld-Verified
Computer-use capabilities, first introduced as native in GPT-5.4, continue to improve. OSWorld-Verified—which tests operating a real desktop environment via screenshots and keyboard/mouse actions—rises to 78.7% (up 3.7 points from 75.0%). In practice this means fewer stuck sessions when agents are asked to operate browsers, edit documents, or navigate portals without human intervention. Combined with stronger web browsing (the Pro variant reaches 90.1% on BrowseComp vs Gemini 3.1 Pro's 85.9%), GPT-5.5 is OpenAI's strongest autonomous operator to date.
Professional Knowledge Work
On GDPval—which measures agents' ability to produce well-specified knowledge work across 44 occupations spanning the nine industries contributing most to U.S. GDP—GPT-5.5 scores 84.9%, up from GPT-5.4's 83.0%. The gain is modest because the benchmark is approaching saturation, but it still translates to fewer factual slips and cleaner deliverables on spreadsheets, presentations, legal drafts, and analytical documents. For teams already using GPT-5.4 for knowledge-work drafts, 5.5 should be framed as an incremental quality and reliability gain, not a new benchmark-proven parity threshold.
Scientific Research Workflows
OpenAI flags early scientific research as one of 5.5's four focus areas, claiming "meaningful gains on scientific and technical research workflows." Public framing points to drug discovery and multi-step experimental planning as example use cases—domains where agents must maintain long context, coordinate tools, and verify their own outputs. OpenAI publishes dedicated science-related evals, including GeneBench 25.0% vs 19.0% for GPT-5.4 and BixBench 80.5% vs 74.0%; the 1M API context and tool-use improvements may still make GPT-5.5 a more credible partner for research-style workloads than any prior ChatGPT release.
Token Efficiency at Matched Latency
Although API pricing doubles (see Pricing section), OpenAI argues that total cost-to-complete often stays flat or drops because GPT-5.5 "uses significantly fewer tokens to complete the same Codex tasks" than GPT-5.4. Per-token latency is explicitly stated to match GPT-5.4 in real-world serving—so developers running AI agent workloads should not see wall-clock regressions, only per-task token savings that partially offset the price increase.
GPT-5.5 Pro for Research Partners
GPT-5.5 Pro is positioned as a higher-accuracy variant for harder questions and demanding work; early testers used it like a research partner, and it delivers strong results on advanced math, BrowseComp (90.1%), and multi-hop web research. It is aimed at users who can provide rich contextual input and benefit from deeper iteration. Pro is available only on higher-tier plans (Pro, Business, Enterprise) and carries its own premium API pricing tier (see below).
Availability & Access
Launch Availability (Apr 23, 2026)
GPT-5.5 began rolling out to Plus, Pro, Business, and Enterprise users in ChatGPT on April 23, 2026. GPT-5.5 Pro rolled out simultaneously but only to Pro, Business, and Enterprise plans. The model is also available in Codex for the roughly four million developers who use it weekly. Access to GPT-5.5 in the API was announced as "coming very soon" at launch—no specific date had been confirmed for API general availability as of release day.
Legacy Model Transition
OpenAI did not publish a retirement schedule for GPT-5.4 or GPT-5.4 mini and nano alongside the launch. OpenAI did not publish a GPT-5.4 retirement schedule in the GPT-5.5 launch post; verify current ChatGPT and API model availability before stating that GPT-5.4 remains selectable or guaranteed.
Pricing & Plans
ChatGPT subscription tiers remain unchanged at launch; GPT-5.5 access is included where the plan supports the Thinking model tier:
- ChatGPT Free — $0/month, no GPT-5.5 Thinking or GPT-5.5 Pro access; current Free plan lists limited GPT-5.3 Instant and GPT-5 Thinking Mini access.
- ChatGPT Plus — $20/month, includes GPT-5.5 with standard usage limits.
- ChatGPT Pro — from current Pro plan pricing; includes GPT-5.5 Thinking and GPT-5.5 Pro with higher usage tiers.
- ChatGPT Business — Team workspace; GPT-5.5 access included, GPT-5.5 Pro included. Seat pricing varies.
- ChatGPT Enterprise — Custom pricing; GPT-5.5 and GPT-5.5 Pro included with admin controls.
- Codex fast mode — Available in ChatGPT Codex at 1.5× token generation speed and 2.5× cost, with a 400K context window.
API pricing for GPT-5.5 (effective on API general availability):
| Model | Input | Output | Context |
|---|---|---|---|
| gpt-5.5 | $5.00/M tokens | $30.00/M tokens | 1,000,000 |
| gpt-5.5-pro | $30.00/M tokens | $180.00/M tokens | 1,000,000 |
Standard API pricing doubles from GPT-5.4's $2.50 input / $15 output. OpenAI's stated rationale is that GPT-5.5's improved token efficiency means many workflows consume fewer tokens per completed task, so total cost-to-complete can stay flat or improve even at higher per-token rates. Whether that holds in practice depends on workload mix—tool-heavy agentic loops benefit most; simple one-shot completions see the full 2× price increase.
Best For
- Agentic coding teams running long-horizon coding loops that repeatedly plan, execute, and verify—where Terminal-Bench 2.0's 82.7% translates directly to more tasks finished autonomously.
- Computer-use automation builders deploying agents that operate browsers, desktop apps, or internal portals and need near-ceiling reliability on complex multi-screen flows.
- Professional knowledge workers producing spreadsheets, reports, presentations, and analytical documents where 84.9% GDPval parity with professionals reduces cleanup overhead.
- Research analysts and scientific teams synthesizing across many sources or tools in multi-step workflows—particularly where the GPT-5.5 Pro variant's BrowseComp lead pays off.
- Codex power users among the ~4M weekly developers who want the newest base model and are willing to trade higher per-token price for fewer tokens per task.
- Enterprise teams consolidating AI vendors who want OpenAI's newest mainline model rather than specialist variants, and already have admin tooling for Plus/Pro/Business/Enterprise rollouts.
FAQ
Is GPT-5.5 worth the doubled API price over GPT-5.4?
For agentic workloads, often yes. GPT-5.5 reportedly completes equivalent Codex tasks with significantly fewer tokens at the same per-token latency, so total cost-to-complete stays flat or drops despite the 2× per-token rate. For single-shot completions (summaries, translations, classification), the efficiency argument doesn't apply and you pay the full price increase. Evaluate on your workload mix.
How does GPT-5.5 compare to Claude Opus 4.7 and Gemini 3.1 Pro?
GPT-5.5 leads both on Terminal-Bench 2.0 (82.7% vs 69.4% and 68.5%) and the Pro variant leads on BrowseComp (90.1% vs 85.9%). It trails Claude Opus 4.7 on SWE-Bench Pro (58.6% vs 64.3%) for GitHub issue resolution; OpenAI's official table flags reported evidence of memorization on that public eval, so treat the comparison cautiously. Pick by task type: agentic terminal and web work → 5.5; repo-scale code refactors → still competitive with Claude.
When will GPT-5.5 be available in the API?
OpenAI announced API access as "coming very soon" at the April 23, 2026 launch but did not publish a specific date. As of release, GPT-5.5 was available in ChatGPT (Plus, Pro, Business, Enterprise) and Codex. Check OpenAI's API docs for the latest status before migrating production workloads.
Does GPT-5.5 replace GPT-5.4 as the default ChatGPT model?
At launch, GPT-5.5 began rolling out to Plus, Pro, Business, and Enterprise users. OpenAI did not publish a deprecation or default-switch timeline for GPT-5.4 or GPT-5.4 mini and nano. Both remain accessible during the transition, and API workloads pinned to gpt-5.4 continue to run.
What's the difference between GPT-5.5 and GPT-5.5 Pro?
Standard GPT-5.5 is the mainline model available on Plus, Pro, Business, and Enterprise. GPT-5.5 Pro is a research-partner variant restricted to Pro, Business, and Enterprise plans, delivering the strongest results on advanced math, multi-hop web research, and BrowseComp (90.1%). In the API, Pro costs $30/M input and $180/M output vs standard's $5/$30.
Does GPT-5.5 support the 1M token context window?
OpenAI says gpt-5.5 will come to the API with a 1,000,000-token context window; only state the same for gpt-5.5-pro if the current model documentation explicitly confirms it. In Codex, the fast mode variant uses a 400K context window optimized for lower latency. ChatGPT-side context limits depend on plan tier; refer to OpenAI's official documentation for per-plan specifics.
