OpenAI launches GPT‑5 for developers: code and agents

What if your AI assistant didn't just “respond”, but rolled up its sleeves and helped you build software from start to finish? That's the wink OpenAI gives with GPT‑5 for developers: less friction, more work actually done.

What OpenAI announces today

On August 7, 2025, OpenAI released GPT‑5 on its API platform and presents it as its best model for code and agent tasks. It comes in three sizes — gpt-5, gpt-5-mini and gpt-5-nano — so you can balance performance, cost and latency depending on the use case. There's also a non-reasoning variant in ChatGPT exposed as gpt-5-chat-latest, aimed at more direct experiences. (openai.com)

Key idea: GPT‑5 isn't just “smarter”; it's more collaborative. It can explain what it will do before and between tool calls, and follow very detailed instructions without getting lost. (openai.com)

What's new that really changes your workflow

• Thought control with reasoning_effort, now including a minimal level for ultra‑fast replies when you don't need long chains of reasoning.
• Length control with verbosity (low, medium, high) so you decide whether you want terse answers or more developed explanations.
• New “custom tools”: the model can call tools using plain text instead of JSON, and you can constrain formats with context‑free grammars.
• Better handling of tool errors, information retrieval in long contexts, and configurable preambles for more stable agents. (openai.com)

Practical examples?

• A support bot that chains actions (search, validate, record) without getting stuck.
• A coding assistant that plans, edits, compiles, summarizes and suggests next steps in a real repo.
• A web prototyper that spins up an app in minutes, detects build errors and documents what it did. (openai.com)

Performance (in plain terms)

If you care about evidence, there are numbers. On real software engineering tasks (SWE‑bench Verified), GPT‑5 reaches 74.9% and outperforms o3; on Aider polyglot it scores 88% on code editing. Translation to day‑to‑day? Less time wrestling with details and more useful commits. (openai.com)

For agentic tasks (following instructions and orchestrating tools), GPT‑5 also sets new highs, which you feel when work needs chained steps without constant supervision. (openai.com)

Factuality and safety

OpenAI reports ~80% fewer factual errors vs o3 on benchmarks like LongFact and FactScore. They also improve performance on health-related prompts, while recommending verification when risk is high. It's an important step if you use agents for sensitive decisions, production data, or automations that affect live systems. (openai.com)

Pricing and availability

GPT‑5 is already available in the Responses API and Chat Completions API (and is the default in Codex CLI). Reference pricing: gpt-5 at $1.25 per 1M input tokens and $10 per 1M output; gpt-5-mini at $0.25 and $2; gpt-5-nano at $0.05 and $0.40, respectively. It is also rolling out in Microsoft 365 Copilot, GitHub Copilot, Copilot and Azure AI Foundry. (openai.com)

Budget tip: use prompt caching, batching, and adjust reasoning_effort to control costs without sacrificing quality where it matters. (openai.com)

How to adopt it without drama

• Start with your “quick” cases: use reasoning_effort: minimal on low‑complexity endpoints and raise the level for critical tasks.
• Set verbosity: low for backend APIs (compact responses) and high in flows that need context and explanations.
• Migrate agents to “custom tools” when rigid JSON gets in the way; grammars help keep format under control.
• Ask the model to “explain before acting” in tool chains: it improves debugging and trust.
• If you're coming from ChatGPT, remember: the non‑reasoning gpt-5-chat-latest is different from the API gpt-5. Adjust prompts and expectations. (openai.com)

Who does it make more sense for now?

• Product teams that prototype on a weekly cadence and need to go from idea to demo without blocking engineering.
• Devs maintaining large codebases who require reliable answers about “how everything fits.”
• Ops and support teams automating tickets with multiple steps and validations.

The takeaway

GPT‑5 brings the promise of a “collaborative AI” down to earth for code and agents: more accuracy, more control and fewer back‑and‑forths. The key? Start small, measure costs and increase reasoning only where it adds value. Let the machine handle the rest. (openai.com)