OpenAI unveils GPT-5.5, vying for corporate-level payments

OpenAI has launched its most powerful model to date, GPT-5.5, achieving a quantum leap in code writing, scientific research, and knowledge work, while maintaining the same level of reasoning latency as previous models and reaching higher levels of intelligence, signifying AI's comprehensive evolution from a question-and-answer tool to an autonomous system capable of completing complex computer tasks.

In terms of model capabilities, GPT-5.5 leads Claude Opus 4.7 and Gemini 3.1 Pro in multiple tests including the Terminal-Bench 2.0 programming test, the OSWorld-Verified benchmark that measures agent control over real computer environments, and the GDPval benchmark that tests cross-44 professional knowledge work capabilities.

GPT-5.5 is now open to ChatGPT's Plus, Pro, Business, and Enterprise users, with the Codex platform launching in sync. Regarding API pricing, OpenAI states that although GPT-5.5 is priced higher than its predecessor, its higher token efficiency makes the overall cost of use competitive.

OpenAI co-founder and President Greg Brockman states that the model can autonomously handle tasks with limited instructions and can call upon email, spreadsheet, calendar, and other applications to execute user commands. "It will find ways to solve problems and deal with ambiguous situations," he said, "It's a more intuitive experience."

MagicPath CEO Pietro Schirano noted that GPT-5.5 completed a branch merge task involving hundreds of front-end modifications and refactoring changes in about 20 minutes. He stated: "It feels like working with a higher-order intelligent life form, even evoking a sense of awe."

Performance Leap: Higher Intelligence, Same Latency

The core technological breakthrough of GPT-5.5 lies in breaking the model rule that "stronger means slower." OpenAI states that GPT-5.5 has achieved latency comparable to GPT-5.4 in actual service environments while significantly outperforming its predecessor in multiple benchmarks.

In the field of agent programming:

GPT-5.5 scored 82.7% on Terminal-Bench 2.0, a significant improvement over GPT-5.4's 75.1%; it reached 58.6% on the SWE-Bench Pro, which tests real GitHub issue resolution capabilities; and also surpassed GPT-5.4 on the Expert-SWE internal long-term programming task benchmark (with a median task completion time of about 20 hours).

It is noteworthy that GPT-5.5 achieved higher scores in the above three tests with fewer token consumptions.

In terms of computer use, GPT-5.5 led with a score of 78.7% in the OSWorld-Verified test over Claude Opus 4.7's 78.0%. In the tool invocation test Tau2-bench Telecom, GPT-5.5 reached 98.0% without any prompt tuning, while GPT-5.4 only achieved 92.8%.

According to the programming comprehensive index from the external evaluation agency Artificial Analysis, GPT-5.5 achieved the same level of intelligent performance at approximately half the cost of competitive cutting-edge programming models.

Engineering Testing: A Qualitative Change in Developer Experience

Feedback from several enterprise technology leaders' real-world testing shows that GPT-5.5 brings improvements to actual engineering scenarios beyond what benchmark numbers can reflect.

Dan Shipper, founder