OpenAI brings GPT-5 to developers

Coding models usually fall apart the moment they touch messy, real-world codebases. They break existing logic, get stuck in infinite tool loops, or burn thousands of tokens overthinking simple queries.

OpenAI just launched GPT-5 across its API platform, positioning it as its strongest model yet for production coding and agentic workflows.

Benchmark records and early feedback

On SWE-bench Verified, GPT-5 posted a 74.9% score, topping o3's 69.1%. Even better, it used 22% fewer output tokens and 45% fewer tool calls compared to o3 at high reasoning effort.

On Aider polyglot, the model hit 88%, slashing o3's error rate by a third. In OpenAI's internal blind tests for front-end web development, testers picked GPT-5 over o3 70% of the time.

Early testers backed up the claims. Cursor called it the smartest and most steerable model they have used. Windsurf reported that GPT-5 halved tool-calling error rates compared to competing frontier models. Vercel singled out its rare mix of aesthetic taste and clean code, while Manus reported record-breaking numbers on their internal agent benchmarks.

Tool-use reliability also saw massive gains. On Sierra.ai's τ2-bench telecom—where no previous model broke 49%—GPT-5 scored 96.7%. Its total context window spans 400,000 tokens, supporting up to 272,000 input tokens and 128,000 reasoning and output tokens.

Smarter API dials and custom tools

Alongside the raw intelligence upgrades, OpenAI added several practical developer controls.

A new minimal option under the reasoning_effort parameter lets developers skip heavy reasoning when speed matters. Notion highlighted that this low-reasoning mode makes GPT-5 ideal for quick, one-shot completions.

A new verbosity parameter (low, medium, high) allows direct control over response length without endless prompt engineering.

OpenAI also introduced custom tools, enabling the model to interact via plain text rather than rigid JSON payloads, enforced by regex or context-free grammars. The API offers three model tiers: gpt-5, gpt-5-mini, and gpt-5-nano.

Why it matters

For developers, GPT-5 fundamentally changes the unit economics and reliability of coding agents. Cutting tool-calling errors in half while dropping token usage by over 20% means agents break less often and cost significantly less to run on everyday tasks.

For the developer tooling landscape, this sets a brutal new standard. Companies building AI IDEs like Cursor and Windsurf now have a base foundation capable of executing dozens of sequential tool calls without losing context. It moves coding agents from flashy demos to dependable background workers.

We think OpenAI's decision to support non-JSON custom tools and granular reasoning controls is particularly clever. Instead of forcing developers into a one-size-fits-all reasoning loop, they are giving engineers the levers to balance latency, cost, and accuracy based on specific tasks.