GPT-5.5 beats Opus on cost

The pricing headline is less interesting than the workflow economics. If a model finishes the job in fewer turns, fewer retries, and less prompt bloat, the effective cost of getting real work done can drop even when list prices rise.

• –OpenAI says GPT-5.5 is more token efficient than GPT-5.4 and delivers better Codex results with fewer tokens for most users.
• –The model posts strong benchmark numbers on coding, tool use, and computer-use tasks, including Terminal-Bench 2.0, SWE-Bench Pro, and OSWorld-Verified.
• –That makes the Opus comparison matter: developers care about total task cost, not just per-token API pricing.
• –The bigger signal is category shift toward agentic efficiency, where benchmark wins and lower completion cost are increasingly the same thing.
• –For teams running long-horizon coding and research loops, token efficiency can matter more than raw benchmark bragging rights.