GPT-5 for Developers Release Notes

OpenAI has released GPT-5 via its API platform, introducing a new family of models optimized for coding and complex agentic tasks. GPT-5 establishes new state-of-the-art (SOTA) benchmarks in software engineering and tool-calling, while introducing developer-centric controls for reasoning effort and response verbosity.

State-of-the-Art Coding and Frontend Capabilities

GPT-5 is designed as a coding collaborator capable of fixing bugs, editing code, and navigating complex codebases. It outperforms previous models, including o3, across key software engineering benchmarks.

• SWE-bench Verified: GPT-5 scored 74.9%, surpassing o3's 69.1%. This was achieved with 22% fewer output tokens and 45% fewer tool calls than o3 at high reasoning effort.
• Aider Polyglot: The model achieved a record 88% score, representing a one-third reduction in the error rate compared to o3.
• Frontend Development: In internal testing, GPT-5 was preferred over o3 70% of the time for frontend web development due to improved aesthetic sense and code quality.

Early adopters, including Cursor, Windsurf, and Vercel, have reported that GPT-5 is the most intelligent coding model they have used, noting its ease of steerability and high performance in aesthetic frontend tasks.

Advanced Agentic Intelligence and Tool Use

GPT-5 demonstrates significant improvements in executing long-running agentic tasks and chaining complex tool calls without losing context.

• Tool Calling Performance: GPT-5 scored 96.7% on the τ 2-bench telecom benchmark, a substantial increase over previous frontier models where no model had previously scored above 49%.
• Instruction Following: The model scored 69.6% on Scale MultiChallenge (graded by o3-mini) and 99% on COLLIE.
• Long-Context Retrieval: GPT-5 outperforms o3 and GPT-4.1 on the OpenAI-MRCR benchmark, with performance gains increasing at longer input lengths. On the BrowseComp Long Context benchmark (128K–256K tokens), it provides correct answers 89% of the time.
• Context Window: All GPT-5 models support a maximum of 272,000 input tokens and 128,000 reasoning and output tokens, totaling a 400,000 token context length.

New API Features for Developer Control

OpenAI has introduced several parameters and tool types to give developers more granular control over model behavior:

• reasoning_effort Parameter: Developers can now select from minimal, low, medium (default), and high. The minimal setting minimizes reasoning time to return answers faster.
• verbosity Parameter: A new parameter with values low, medium (default), and high controls the length of responses. Explicit prompt instructions override this setting.
• Custom Tools: GPT-5 now supports calling tools via plaintext instead of JSON. This reduces errors associated with escaping control characters in long outputs. Developers can constrain these tools using regex or context-free grammars.
• Preamble Messages: When instructed, GPT-5 can output visible messages before and between tool calls to communicate its plan and progress to the end user.

Factuality and Safety

GPT-5 shows a marked reduction in hallucinations compared to previous models. On LongFact and FactScore benchmarks, GPT-5 makes approximately 80% fewer factual errors than o3. The model has also been specifically tuned for higher accuracy on health-related questions and increased self-awareness regarding its own limitations.

Model Availability and Pricing

GPT-5 is available in three sizes to allow developers to balance performance, cost, and latency. It is also available via the Responses API, Chat Completions API, and as the default in Codex CLI.

Model	Input Price (per 1M tokens)	Output Price (per 1M tokens)
gpt-5	$1.25	$10.00
gpt-5-mini	$0.25	$2.00
gpt-5-nano	$0.05	$0.40

Additionally, the non-reasoning version used in ChatGPT is available as gpt-5-chat-latest ($1.25 input / $10.00 output). GPT-5 is also deploying across Microsoft platforms, including Azure AI Foundry, GitHub Copilot, and Microsoft 365 Copilot.