OpenAI launches GPT-5 for AI developers

OpenAI now offers GPT-5 through its API, allowing developers to build applications that can code, use tools and complete long tasks with less supervision. The company presents it as its most capable model for coding and AI agents.

The launch includes three versions: gpt-5, gpt-5-mini and gpt-5-nano. Each aims for a different balance of quality, speed and cost.

A model built for coding

GPT-5 can fix errors, edit projects, explain codebases and build web interfaces. It is also designed to follow detailed instructions and tell users what it plans to do before calling tools.

In tests published by OpenAI, it achieved:

• 74.9% on SWE-bench Verified, compared with 69.1% for o3. This test measures the resolution of real-world software engineering problems. OpenAI says 23 of the 500 tasks were excluded because of infrastructure issues.
• 88% on Aider polyglot, a code-editing evaluation, compared with 79.6% for o3.
• Evaluator preference in 70% of comparisons with o3 when generating code for web interfaces.

According to OpenAI, GPT-5 achieved these results using 22% fewer output tokens and 45% fewer tool calls than o3 with a high reasoning level. That could translate into faster responses and lower costs for complex tasks, although the actual result will depend on how the model is integrated into each application.

More autonomy for agents

The main difference is not just its ability to write code. GPT-5 can chain many actions, sequentially or in parallel, without losing sight of the goal as easily. An agent could query a database, modify files, run tests and fix errors within the same process.

In the τ2-bench telecom test, which focuses on tool use and managing states that can change during a task, GPT-5 scored 96.7%. OpenAI also says it improves at following tool instructions, recovering from errors and searching for information across large volumes of context.

The model supports up to 272,000 input tokens and can generate up to 128,000 tokens of reasoning and response. Combined, it handles a maximum context of 400,000 tokens. In practice, this makes it possible to work with much larger software projects, documents or search results without splitting them into as many parts.

More control for developers

The API adds several new options for deciding how the model responds:

• reasoning_effort adds the minimal level, designed for fast responses to simple tasks. It also keeps the low, medium and high levels.
• verbosity lets you request low, medium or high response lengths.
• custom tools allow developers to send plain text to a tool instead of requiring the model to always build a JSON structure. They also accept developer-defined rules to limit the response format.

GPT-5 can show progress messages before and between tool calls. This helps you understand whether the agent is searching for information, taking an action or fixing an error, without confusing those messages with its internal reasoning.

How much it costs

The prices published by OpenAI are:

• gpt-5: $1.25 per million input tokens and $10 per million output tokens.
• gpt-5-mini: $0.25 for input and $2 for output per million tokens.
• gpt-5-nano: $0.05 for input and $0.40 for output per million tokens.

It is also coming to the Responses API, Chat Completions API and Codex CLI, where it becomes the default model. The reasoning-free version used in ChatGPT is available in the API as gpt-5-chat-latest, at the same price as gpt-5.

For you, the most visible change will come through the tools you use, not necessarily as a new OpenAI application. Code editors, enterprise assistants and services capable of taking actions will be able to complete longer tasks with less intervention. Even so, OpenAI acknowledges that its results need to be verified, especially when code, data or decisions have important consequences. The next thing to watch is how much of this improvement in testing holds up in real projects and how much ultimately affects the price of these services.