GPT-5.5

Roboflow's vision benchmark of GPT-5.5 reports a shift to a 32x32 patch-based vision encoder that improves handling of high-resolution documents, complex charts, and structured layouts. On the Roboflow Vision Evals suite, GPT-5.5 placed fourth of 63 models at 76.12% overall accuracy, clearing 51 of 67 visual prompt evaluations. Domain strengths included 88.9% document understanding, 86.7% defect detection, and 85.7% object understanding. A clear weak point was precise counting at 30% accuracy, flagged as a limitation for quantitative visual tasks. Roboflow positioned GPT-5.5 as a fit for zero-shot dataset generation, document parsing, and unstructured visual reasoning, with integration available through Roboflow Playground and Workflows. The analysis was dated May 13, 2026, building on OpenAI's April 23, 2026 GPT-5.5 release.

GitHub announced on April 24, 2026 that OpenAI's GPT-5.5 is generally available on GitHub Copilot, with the company reporting that the model shows its strongest performance on complex, multi-step agentic coding tasks and resolves real-world coding challenges that prior GPT models could not. The model launches under pro According to the changelog, GPT-5.5 surfaces in the model picker across Visual Studio Code, Visual Studio, the Copilot CLI, the Copilot cloud agent, github.com, GitHub Mobile for iOS and Android, JetBrains IDEs, Xcode, and Eclipse, with GitHub noting that rollout will be gradual so some users may need to check back. Th

OpenAI's API docs introduce the model as a flagship for complex professional work and coding, positioned as a step above GPT-5.4 and GPT-5.4 Mini. It supports text and image input, text output, and a reasoning.effort parameter ranging from none to xhigh, with medium as the default. The context window is 1,050,000 tokens with a 128,000-token max output. Pricing is set at $5.00 per million input tokens, $0.50 per million cached input tokens, and $30.00 per million output tokens, with a 2x input and 1.5x output surcharge on prompts exceeding 272K tokens across standard, batch, and flex tiers. Regional data-residency endpoints carry a 10% uplift. Available surfaces include Chat Completions, Responses, Realtime, Realtime translation/transcription, and Live sessions.

CodeRabbit's early hands-on testing of GPT-5.5 reports the model is quicker, leaner, and more direct in coding workflows, favoring scoped changes over broad rewrites while preserving existing behavior. In their live code-review agent, GPT-5.5 reached 79.2% expected-issue-found against CodeRabbit's current multi-model review baseline, with the vendor noting that baseline is a combined system, not a single competing model. The reviewers also observed reduced token consumption on long-running agent tasks and faster surfacing of user-facing progress before internal work completes. Practical gains showed up most in debugging-oriented reviews involving access control, error handling, and API behavior, where the model isolated regressions and pointed to preserving fixes. CodeRabbit framed GPT-5.5 as well-suited to teams building agent workflows or asking the model to make real changes in codebases.