On April 24, 2026, OpenAI officially launched its most powerful AI model to date — GPT-5.5, which not only achieves a qualitative leap in intelligence but also redefines the economic boundaries of large models with its astonishing energy efficiency. Designed for the "Agent" era, this model can autonomously plan, call tools, and perform multi-step complex tasks even with ambiguous instructions, excelling in code generation, scientific exploration, and enterprise automation.
🧠 Key Breakthrough: From "Passive Response" to "Active Execution"
The biggest advancement of GPT-5.5 lies in its strong autonomy. Users no longer need to give step-by-step instructions; they can simply issue vague multi-objective commands like "analyze this financial report and generate a visual report, pointing out potential risks," and the model will automatically break down the task, call data interfaces, write scripts, validate results, and iterate for optimization until completion. OpenAI CEO Greg Brockman emphasized, "It accomplishes more with less guidance."
In authoritative tests, GPT-5.5 demonstrated dominant power:
- Terminal-Bench2.0: accuracy of 82.7%;
- SWE-Bench Pro (GitHub problem solving): 58.6%;
- GeneBench & BixBench: leading in gene analysis and bioinformatics tasks, and helped mathematicians discover a new proof of Ramsey numbers.
⚡ Energy Efficiency Revolution: Token Cost Drops 97%, Output per Megawatt Increases 50 Times
Despite being larger and more capable, GPT-5.5 achieved an unprecedented efficiency leap:
- Token consumption required to complete the same task has significantly decreased;
- Measured on NVIDIA GB200NVL72 system, the cost of processing one million tokens is reduced to 1/35 of the previous generation;
- Token output per megawatt of power increased 50 times, completely rewriting the economics of AI computing power.
This breakthrough was made possible by the deep collaboration between OpenAI and NVIDIA. Huang Renxun and Altman personally promoted Codex (a programming agent based on GPT-5.5) across the entire NVIDIA company, over 10,000 employees have already used it in engineering, legal, marketing, and other scenarios. Engineers reported that what used to take days of debugging cycles is now shortened to hours, and complex codebase experiments are compressed from weeks to overnight.
💼 Enterprise Implementation: From Tax Audits to System Architecture Understanding
Over 85% of OpenAI employees use Codex weekly, with the finance team using it to audit over 70,000 pages of tax documents. External customers such as NVIDIA confirmed that GPT-5.5 has significantly enhanced its ability to understand system architecture, predict code review needs, and locate deep faults, truly achieving "natural language as delivery."
💰 Pricing and Availability
- GPT-5.5 is now available to Plus and Pro users;
- API will be launched soon, with pricing as follows:
- Standard version: $5 per million tokens (approximately 34.2 RMB);
- Pro version: $30 per million tokens.
