#OpenAIReleasesGPT-5.5

OpenAI GPT-5.5 Released The First AI That Actually Does Computer Work
On April 23 2026 OpenAI officially released GPT-5.5 In the company's description this is a new class of intelligence for real work We are talking about a model that no longer just chats but plans uses tools controls its own work and carries out multi-step tasks from start to finish
What is GPT-5.5 and What Can It Do
According to OpenAI GPT-5.5 can work hands-free If you give it a vague project it plans uses tools verifies its output and continues until finished
Key Capabilities
1 Agent Coding Writes code debugs opens PRs runs tests notifies Slack on status It became SOTA in Terminal-Bench 2.0 with 82.7 percent Claude Opus 4.7 is 69.4 percent Gemini 3.1 Pro is 68.5 percent
2 Computer Use 78.7 score in OSWorld-Verified It can work like a human on real interfaces
3 Knowledge Work and Research Analyzes data extracts PowerPoint presentations from Excel creates documents and spreadsheets It produces human-level work in 44 professions with 84.9 percent in GDPval
4 Scientific Research 35.4 percent in FrontierMath Tier 4 far ahead of its competitors OpenAI CRO Mark Chen says it can really help expert scientists make progress in areas like drug discovery
Versions and Access
GPT-5.5 Thinking Opened to Plus Pro Business Enterprise users Fast help for harder problems
GPT-5.5 Pro Exclusive to Pro Business Enterprise Research partner for difficult questions where accuracy is more important than speed
Codex Came to Plus Pro Business Enterprise Edu Go plans Has 400K context window plus Fast mode
Price and Efficiency
API 5 dollars per 1M input token 30 dollars per 1M output token Double that of GPT-5.4
Pro API 30 input 180 output
Context 922K tokens equals approximately 1 383 A4 pages Some documents state 1M tokens
Advantage Smarter at the same token latency as GPT-5.4 and finishes the same job with fewer tokens in Codex
Demo What Does It Actually Do
In the video released by OpenAI GPT-5.5 does the following autonomously
Solves a Rubik's cube
Reads and fixes a bug on GitHub opens a PR checks if it is merged
Writes PR merged to Slack
Produces a financial PowerPoint presentation from Excel Revenue EBITDA ARR cash tables
Comparison with Competitors
GPT-5.5 surpassed Anthropic's best coding model Claude Opus 4.7 and Gemini 3.1 Pro in Terminal-Bench 2.0 58.6 percent in SWE-Bench Pro According to Artificial Analysis it is number 2 among 141 models
Reactions Excitement and Criticism
The community is split One part is cheering Let's mother f ing gooooooo The other part criticizes the price increase and rapid release cycle Capability climbs Costs double Tradeoffs deepen Free users say This update is not for us
Towards a Super App
Greg Brockman GPT-5.5 is a big step towards agentic and intuitive computer usage Goal Combine ChatGPT Codex and the AI browser into one super app
Security Note
OpenAI highlights GPT-5.5 for cybersecurity While Anthropic's Mythos creates debate OpenAI says our strategy of directing our model to digital defense is strong
Summary What is the Difference
Gigi B's comment is clear GPT-5.5 is not smarter but more mature It does not let go of the complex goal does not go off the rails does not ask for help 15 times Now the bottleneck is not the AI but how clear your instructions are
GPT-5.4 was released 6 weeks ago This speed puts OpenAI's training-deployment pipeline ahead of competitors