OpenAI GPT-5.1 Release Notes
OpenAI has released GPT-5.1, a model designed to balance high-level intelligence with execution speed for agentic and coding workflows. The update introduces adaptive reasoning that adjusts thinking time based on task complexity, a dedicated mode for low-latency responses, and new developer tools for direct code and system interaction.
Adaptive Reasoning and Efficiency
GPT-5.1 utilizes an overhauled training approach that allows it to dynamically adjust the amount of time and tokens spent "thinking" depending on the complexity of the request. This results in faster responses and lower token costs for simple tasks without sacrificing reliability on complex problems.
Performance Gains
- Latency Reduction: For simple queries, such as requesting an npm command, GPT-5.1 can reduce response time from 10 seconds to 2 seconds.
- Third-Party Validation: Balyasny Asset Management reported that GPT-5.1 ran 2-3x faster than GPT-5 and used approximately half as many tokens as leading competitors on tool-heavy reasoning tasks. Pace reported that their agents ran 50% faster while exceeding the accuracy of GPT-5.
"No Reasoning" Mode
Developers can set the reasoning_effort parameter to none to disable reasoning. This mode is optimized for latency-sensitive use cases and provides several advantages over GPT-5's "minimal" reasoning setting:
- Improved Capabilities: Better performance in parallel tool calling, coding tasks, instruction following, and the use of search tools.
- Web Search: Supports web search within the API platform.
- Real-world Impact: Sierra observed a 20% improvement in low-latency tool calling performance compared to GPT-5 minimal reasoning.
Extended Prompt Caching
To further reduce latency and cost, GPT-5.1 supports extended prompt caching with retention for up to 24 hours (via the prompt_cache_retention='24h' parameter). This is a significant increase over previous limits, facilitating smoother multi-turn chats and long-running coding sessions. Cached input tokens remain 90% cheaper than uncached tokens.
Enhanced Coding Capabilities
GPT-5.1 improves upon GPT-5 in steerability, code quality, and the clarity of user-facing update messages during tool calls. It is specifically optimized to reduce "overthinking" on simple edits while maintaining high performance on complex tasks.
Benchmarks and Industry Feedback
- SWE-bench Verified: GPT-5.1 achieved a score of 76.3%, surpassing GPT-5's 72.8%.
- Developer Tooling: Companies including Cursor, Cognition, Augment Code, Factory, and Warp collaborated on the model's coding personality. Augment Code noted more accurate changes and faster iteration in multi-file projects, while Cline reported a 7% improvement on their diff editing benchmark.
New Developer Tools
Two new tools are introduced in the Responses API to enable more reliable agentic behavior:
apply_patch Tool
This tool allows the model to create, update, and delete files using structured diffs. By emitting patch operations rather than just suggesting edits, the model enables a more reliable, iterative multi-step editing workflow without the need for JSON escaping.
Shell Tool
The shell tool enables the model to propose command-line interface (CLI) commands to be executed on a local machine. This creates a plan-execute loop where the model can inspect systems, run utilities, and gather data to complete a task.
Model Availability and Variants
GPT-5.1 and gpt-5.1-chat-latest are available to developers on all paid API tiers with pricing and rate limits identical to GPT-5. In addition, OpenAI has released gpt-5.1-codex and gpt-5.1-codex-mini, which are optimized specifically for long-running, agentic coding tasks within Codex-like harnesses.
Technical Evaluation Summary
| Evaluation | GPT-5.1 (high) | GPT-5 (high) |
|---|---|---|
| SWE-bench Verified | 76.3% | 72.8% |
| GPQA Diamond (no tools) | 88.1% | 85.7% |
| AIME 2025 (no tools) | 94.0% | 94.6% |
| FrontierMath (with Python) | 26.7% | 26.3% |
| MMMU | 85.4% | 84.2% |
| Tau 2-bench Airline | 67.0% | 62.6% |
| Tau 2-bench Telecom | 95.6% | 96.7% |
| Tau 2-bench Retail | 77.9% | 81.1% |
| BrowseComp Long Context 128k | 90.0% | 90.0% |