GPT-5.5, launched on April 23, 2026, marks the transition from conversational assistant to autonomous agent. OpenAI wants its model to execute complex tasks without constant supervision.
From assistant to agent: a paradigm shift
Just six weeks after the release of GPT-5.4, OpenAI unveiled GPT-5.5 on April 23, 2026. The pace is fast, deliberately so. The company isn't trying to hide it: Greg Brockman, co-founder and president of OpenAI, himself acknowledged during a press briefing that "model releases are now so frequent that it's probably getting hard to distinguish one from another." ("there are enough model releases that it's probably getting hard to distinguish one from another.") But according to him, GPT-5.5 stands out precisely because it can do more with fewer directives.
The central promise of this model is simple: the user no longer needs to detail every step. Brockman states it directly during this same briefing: GPT-5.5 is "way more intuitive to use – it can look at an unclear problem and figure out what needs to happen next." ("way more intuitive to use – it can look at an unclear problem and figure out what needs to happen next.") OpenAI specifies in its official documentation that the user can now entrust the model with "a messy, multi-part task and trust it to plan, use tools, check its work, navigate through ambiguity, and keep going." ("a messy, multi-part task and trust it to plan, use tools, check its work, navigate through ambiguity, and keep going.")
This positioning marks a clear break from the chatbot era. OpenAI is no longer selling an input-response tool: official discourse now speaks of "results" rather than raw quality metrics.
Benchmarks that legitimize ambition
The figures published by OpenAI support this claim. On Terminal-Bench 2.0, a test that evaluates complex command-line tasks requiring planning, iteration, and tool coordination, GPT-5.5 reaches 82.7%, compared to 69.4% for Anthropic's Claude Opus 4.7. On GDPval, which measures an agent's performance across 44 office work professions, the model achieves 84.9%. On OSWorld-Verified, a benchmark that tests autonomous use of real computer environments, it shows 78.7%.
These results place GPT-5.5 above its direct competitors, notably Google's Gemini 3.1 Pro and Anthropic's Claude Opus 4.5, according to data released by OpenAI. The company also claims increased efficiency : GPT-5.5 maintains the same latency per token as GPT-5.4 in real-world conditions, while using significantly fewer tokens to accomplish the same tasks in Codex, according to OpenAI's official press release published on April 23, 2026. The model is therefore faster to conclude without consuming more resources.
A notable technical element complements this picture: the context window reaches 1 million tokens for all API accesses, a capability that opens the door to long workflows, such as analyzing complete data corpora or managing extensive software projects.
Access, pricing, and deployment
The launch on April 23, 2026, opened access to GPT-5.5 for paying subscribers of ChatGPT Plus, Pro, Business, and Enterprise, as well as within the Codex tool. The following day, April 24, OpenAI extended availability to the API with GPT-5.5 and GPT-5.5 Pro now accessible to developers.
| Version | Access available | Use case |
|---|---|---|
| GPT-5.5 | Plus, Business, Enterprise, API | Standard agent tasks, coding |
| GPT-5.5 Pro | Pro, Business, Enterprise | Long tasks, high precision |
The API price has doubled compared to GPT-5.4, increasing to 5 dollars for input and 30 dollars for output per million tokens. A pricing strategy that clearly targets companies ready to pay for autonomous execution rather than simple text generation. For API deployments, OpenAI specifies that "different protections" will be necessary, implicitly acknowledging that the model's increased autonomy calls for adapted safeguards.
OpenAI faces competitive pressure
The deployment speed, six weeks between GPT-5.4 and GPT-5.5, is not insignificant. It reflects an intense race between major AI labs to capture enterprise contracts. Google with Gemini, Anthropic with Claude, and newer players maintain constant pressure on development cycles.
Greg Brockman presented GPT-5.5 as "a big step towards more agentic and intuitive computing" during the press briefing on April 23, 2026. TechCrunch reports that he also discussed this model as a step towards what he calls OpenAI's "super app," a unified application capable of handling all types of digital work continuously. The stakes go beyond individual models : with GPT-5.5, OpenAI is building the argument that AI can now manage a complete workflow, not just an isolated step.
The question is no longer about what a model can do punctually, but about the amount of work it can absorb without human intervention. The next step will be to verify if enterprise adoption actually follows at this pricing level, and how Google and Anthropic will respond in the coming weeks.
No Comment! Be the first one.