OpenAI ships GPT-5.1: faster thinking, cheaper tokens, and tools that actually ship code

OpenAI is rolling out GPT-5.1 in the API platform today, and the headline isn’t just about a bigger model—it’s about a smarter allocation of compute. The company has fundamentally retrained how the model thinks, allowing it to dynamically scale its reasoning. On a trivial prompt like asking for an npm command, GPT-5.1 spits out an answer in about 2 seconds using just 50 tokens. The previous GPT-5 burned roughly 250 tokens and took 10 seconds for the same job. That’s not an incremental speed bump; it’s a fivefold reduction in token spend for everyday tasks.

Beyond raw speed, the model introduces a “no reasoning” mode (setting reasoning_effort to ‘none’) that kills the thinking phase entirely for latency-sensitive workloads. Customer Sierra says that mode alone showed a 20% improvement in low-latency tool-calling performance compared to GPT-5’s minimal reasoning. The model defaults to this fast mode, but developers can crank the dial up to ‘low,’ ‘medium,’ or ‘high’ when they need more deliberate chain-of-thought. Balyasny Asset Management threw it into their full evaluation suite and reported that GPT-5.1 not only outperformed both GPT-4.1 and GPT-5, but ran 2-3x faster while consuming roughly half the tokens of leading competitors at comparable quality.

OpenAI isn’t just tweaking the engine; they’re adding new wrenches to the toolbox. GPT-5.1 ships with two new tools: an apply_patch model designed to edit code more reliably, and a shell tool that lets the model execute shell commands directly. They’ve also been working closely with developer-focused startups like Cursor, Cognition, and Augment Code to sand down the model’s coding personality. The goal—and early reports suggest they’ve achieved it—is a model that feels more intuitive, less prone to overthinking, and better at communicating what it’s doing during multi-step tasks.

The pricing infrastructure got a quiet but critical upgrade too. Extended prompt caching now retains context for a full 24 hours instead of a few minutes. For long-running coding sessions or multi-turn chats, this means follow-up requests will hit the cache far more often, driving latency and costs down without any increase in storage fees. Cached tokens remain 90% cheaper than uncached ones. It’s a logistics play that makes building agentic workflows significantly more economical. OpenAI is clearly betting that developer stickiness in the agent race won’t just come from the smartest model, but from the one that’s the least annoying and most affordable to run at scale.