OpenAI Ships GPT-5 API: 74.9% on SWE-bench, Half the Tool Errors

OpenAI just dropped GPT-5 into its API platform, and the early numbers from partners suggest it isn’t just another benchmark bump—it’s a meaningful shift in reliability for coding and agentic workloads. The model scores 74.9% on SWE-bench Verified, climbing from o3’s 69.1%, but the more interesting stat is how it gets there. GPT-5 uses 22% fewer output tokens and makes 45% fewer tool calls than o3 did at high reasoning effort. That efficiency matters when you’re paying by the token and agents are looping through dozens of steps.

On Aider polyglot, a code editing benchmark, it hits 88%—which represents a one-third reduction in error rate compared to o3. The front-end results are equally stark. Internal testing shows GPT-5 beats o3 on frontend web development 70% of the time. Vercel went as far as saying it’s in “a category of its own” across both aesthetic quality and code output. Those aren’t marginal gains. They’re the kind of jumps that change which model you reach for first thing in the morning.

The agentic performance might be the real headline here. GPT-5 scored 96.7% on τ2-bench telecom, a tool-calling benchmark released just two months ago. Windsurf reported that the model “has half the tool calling error rate over other frontier models.” When you’re chaining dozens of API calls in sequence and parallel, cutting error rates in half isn’t a nice-to-have—it’s the difference between a task completing and a task dying silently three steps in. Manus called it the best single-model performance they’ve ever recorded on their internal benchmarks.

OpenAI is shipping three sizes—gpt-5, gpt-5-mini, and gpt-5-nano—alongside new API controls that developers have wanted for a while. A verbosity parameter (low, medium, high) lets you dial response length up or down. The reasoning_effort parameter now accepts a “minimal” value for faster responses without the preamble. There’s also a new custom tools option that accepts plaintext instead of JSON, constrained by developer-supplied context-free grammars. That’s a quiet but significant quality-of-life improvement for anyone who’s wrestled with an agent that keeps malforming its tool calls. One note: the GPT-5 in the API is the reasoning model powering maximum performance in ChatGPT, not the full system of router and non-reasoning models. The non-reasoning variant is available separately as gpt-5-chat-latest. Cursor’s team summed up the sentiment: “It’s become our daily driver for everything from scoping and planning PRs to completing end-to-end builds.”