OpenAI Launches GPT-5.5: 82.7% Terminal-Bench, 58.6% SWE-Bench Pro, Cheaper Per Codex Task
OpenAI released GPT-5.5 — codenamed Spud — on April 23, rolling it out to Plus, Pro, Business, and Enterprise tiers in ChatGPT and Codex. GPT-5.5 Pro, the extended-capability variant, goes to Pro, Business, and Enterprise users only. API availability is listed as coming “very soon,” pending safety requirements for scale deployment.
Benchmark Numbers
| Eval | GPT-5.5 | GPT-5.4 | Claude Opus 4.7 | Gemini 3.1 Pro |
|---|---|---|---|---|
| Terminal-Bench 2.0 | 82.7% | 75.1% | 69.4% | 68.5% |
| Expert-SWE (internal) | 73.1% | 68.5% | — | — |
| GDPval (wins/ties) | 84.9% | 83.0% | 80.3% | 67.3% |
| OSWorld-Verified | 78.7% | 75.0% | 78.0% | — |
| FrontierMath Tier 1–3 | 51.7% | 47.6% | 43.8% | 36.9% |
| FrontierMath Tier 4 | 35.4% | 27.1% | 22.9% | 16.7% |
| CyberGym | 81.8% | 79.0% | 73.1% | — |
| BrowseComp | 84.4% | 82.7% | 79.3% | 85.9% |
| SWE-Bench Pro | 58.6% | — | — | — |
The Terminal-Bench 2.0 score is corroborated by the independent leaderboard at tbench.ai, where Codex running GPT-5.5 entered at 82.0%, clearing ForgeCode+GPT-5.4’s prior top score of 81.8%. The gap between OpenAI’s 82.7% figure and the leaderboard’s 82.0% reflects different agent configurations and k-values used in each run.
Efficiency Angle
OpenAI describes GPT-5.5 as “more efficient as well as more capable”: it completes identical Codex tasks with fewer tokens than GPT-5.4, and matches GPT-5.4’s per-token latency in real-world serving despite the intelligence step-up. On Artificial Analysis’s Coding Index, it posts state-of-the-art scores at roughly half the cost of competitive frontier coding models, according to OpenAI’s release notes — a direct response to Anthropic’s pricing pressure from Opus 4.7.
Where the Gains Are
Strongest improvements over GPT-5.4 land in agentic coding (FrontierMath Tier 4: +30.6% relative), OSWorld computer use (+4.9%), and Expert-SWE long-horizon tasks. BrowseComp web research improves slightly; CyberGym shows a 3.5-point gain.
Early testers cited the model’s ability to hold context across large systems and reason through ambiguous failures without explicit prompting. Cursor CEO Michael Truell noted GPT-5.5 “stays on task significantly longer without stopping early” compared to GPT-5.4. One NVIDIA engineer with early access reportedly described losing access as “having a limb amputated.”
What Is Not Changed
Pricing has not been announced for the API tier. GPT-5.5 is not yet available to free-tier ChatGPT users. No context window or multimodal changes were published alongside the release. Safety and preparedness testing involved “nearly 200 trusted early-access partners” before launch, with targeted testing for cybersecurity and biology capability.
The Codex-first rollout and the framing around “getting work done on a computer” positions GPT-5.5 as OpenAI’s answer to Claude Code — directly competing on the agentic coding workflows Anthropic has been optimizing for since late 2025.