GPT‑5.5 costs 49–92% more

API pricing stories usually sound boring — until they hit a company budget. That is basically what happened with GPT‑5.5. OpenAI launched the model in late April with a simple pitch: yes, the token price is higher, but the model should often answer more efficiently. Then OpenRouter looked at what happened after customers actually switched from GPT‑5.4 to GPT‑5.5, and the real bill still came out much higher. ### What actually changed? OpenAI made GPT‑5.5 available in the API on April 24, 2026. The posted price for the standard model is $5 per 1 million input tokens and $30 per 1 million output tokens. GPT‑5.4, by comparison, is listed at $2.50 input and $15 output. So on paper, the sticker price doubled. ### Wasn’t the whole pitch about efficiency? Yes — and that is the important wrinkle. (openrouter.ai) OpenAI positioned GPT‑5.5 as a stronger model for coding and professional work, and the implied tradeoff was that better reasoning and shorter completions could offset some of the higher token rates. OpenRouter’s analysis says that offset is real, but not big enough. In its switcher cohort, longer prompts led GPT‑5.5 to generate 19% to 34% fewer completion tokens, yet total costs still rose. (openai.com) ### Where does the 49% to 92% number come from? OpenRouter did not just compare list prices. It looked at users who had been sending traffic to GPT‑5.4 and then switched that traffic to GPT‑5.5. That matters because it tries to capture production behavior instead of lab-style assumptions. The result was a real-world cost increase of 49% to 92%, depending on prompt length. Shorter prompts got hit harder because there was less room for shorter completions to make up the price jump. (openrouter.ai) ### Why do prompt lengths matter so much? Because pricing is a mix of what you send in and what the model sends back. If a workload has huge prompts — say long code files, contracts, or retrieval-heavy agent context — then shaving output tokens can help a bit. But if the prompt is short, the doubled token rates dominate the bill. Think of it like buying a car with better fuel economy after gas prices doubled — efficiency helps, but the math can still get worse overall. (openrouter.ai) ### Does this mean GPT‑5.5 is overpriced? Not automatically. Buyers are not paying for tokens in the abstract — they are paying for work completed. If GPT‑5.5 solves more tasks per call, reduces retries, or lowers human review time, the higher model bill can still be worth it. But the catch is that procurement teams now need proof. “It uses fewer tokens” is no longer a strong enough argument by itself. (openrouter.ai) ### Why is OpenRouter’s data getting attention? Because it is one of the few public looks at migration behavior after launch. Most model vendors publish benchmark wins and list prices. Very few publish what happens to a mixed production workload once customers actually reroute traffic. That makes this less a story about one percentage point and more a story about how AI buyers should evaluate upgrades. (openrouter.ai) ### What should teams do with this? They should test by workload, not by headline. Coding copilots, document agents, customer support flows, and internal search all have different prompt-output shapes. A model that is “only” 49% more expensive in one lane can be 92% more expensive in another. Cached-input discounts, batch processing, and routing only the hardest tasks to GPT‑5.5 can soften the blow, but none of that changes the core point — the new model is not a free efficiency upgrade. (openrouter.ai) ### Bottom line? GPT‑5.5 may be better. But better and cheaper are not the same thing. Right now, the public evidence says most teams should treat it as a premium model with a real budget premium attached. (openrouter.ai) (openai.com)