GPT‑5.5 costs up to 92% more

OpenAI’s GPT‑5.5 story is not really one product launch. It’s three separate moves that landed within about two weeks — a flagship API model on April 23, a cheaper everyday ChatGPT default on May 5, and a cyber-focused access program on May 7. The easy read is “new model, better model.” But the part buyers actually have to wrestle with is pricing. GPT‑5.5 looks materially more capable than GPT‑5.4, yet the token math got steeper at the same time. ### What actually launched? GPT‑5.5 first showed up as OpenAI’s new flagship model for coding, research, data analysis, and tool-using work. OpenAI then said GPT‑5.5 and GPT‑5.5 Pro were available in the API on April 24, after the initial April 23 announcement. Separately, GPT‑5.5 Instant became ChatGPT’s default model on May 5, aimed at faster everyday use rather than maximum depth. Then came the May 7 cyber announcement, which opened GPT‑5.5 and a more tightly controlled GPT‑5.5‑Cyber track to trusted security users. (openai.com) ### Why is the price jump the headline? Because the official pricing page makes the tradeoff blunt. GPT‑5.5 costs $5 per 1 million input tokens and $30 per 1 million output tokens. GPT‑5.4 costs $2.50 input. OpenAI’s GPT‑5.5 model page also frames GPT‑5.4 as the cheaper comparison point, with GPT‑5.5 output landing around 92% higher than GPT‑5.4’s roughly $15.63 benchmark. Input cost is a straight 100% jump. That means “better model” is not the same thing as “easy upgrade.” (openai.com) ### So is GPT‑5.5 just more expensive? Not quite. OpenAI is making a different argument — that GPT‑5.5 finishes hard work with fewer wasted steps. The company says GPT‑5.5 matches GPT‑5.4 on per-token latency in real-world serving, uses fewer tokens on the same Codex tasks, and is better at outcome-first prompting, tool use, and long multi-step execution. Basically, OpenAI is asking customers to judge price per completed task, not price per token. (openai.com) ### What is GPT‑5.5 Instant doing here? It shows OpenAI knows not every user wants flagship-model economics. GPT‑5.5 Instant is the mass-market version — the default ChatGPT model, tuned for clearer answers, lower hallucination rates, and personalization controls. But even that launch came with a notable safety signal: OpenAI said this is the first Instant model it classifies as “High” capability in both cybersecurity and biological and chemical preparedness, which triggered stronger safeguards. (openai.com) ### Why the special cyber access program? Because stronger cyber capability is useful and risky at the same time. OpenAI’s May 7 post says most legitimate defensive teams should use GPT‑5.5 with Trusted Access for Cyber, while GPT‑5.5‑Cyber is reserved for narrower research and testing settings with more restrictions. That split tells you something important — OpenAI thinks the capability jump is real enough that access itself now needs to be tiered, not just the price sheet. (openai.com) ### What does this mean for enterprise buyers? The old buying question was “Is the new model smarter?” The new one is “Where does the extra intelligence pay back?” If a team is doing high-value coding, research, or agentic workflows, GPT‑5.5 may justify the premium. If the workload is mostly routine summarization or chat, the catch is obvious — the bill can rise faster than the benefit. That pushes companies toward a model mix, not a one-model standard. (openai.com) ### Why does benchmarking get messier now? Because OpenAI is no longer shipping one clean ladder from small to big. It’s shipping a family — flagship, Pro, Instant, and cyber-restricted variants — each with different safeguards, access rules, and economics. That makes raw benchmark wins less decisive. The practical question becomes which version you can actually buy, deploy, and govern. (openai.com) ### Bottom line? GPT‑5.5 looks like a real capability step. But the market story on May 11 is not just performance — it’s segmentation. OpenAI is charging more, fencing off the riskiest uses more tightly, and telling customers to pay for outcomes instead of tokens. Buyers now have to decide whether that math works in their own stack. (openai.com 1) (openai.com 2)