Model Releases
OpenAI Releases GPT-6.1 Sol Coding Model
OpenAI's new GPT-6.1 Sol delivers agentic coding and computer use capabilities nearly matching GPT-6 Astra at one-fifth of the cost.
September 22, 2026
Model ReleaseGPT-6.1 Sol
Introduction
On September 22, 2026, OpenAI shifted the economics of high-end AI development with the release of GPT-6.1 Sol. While the industry has been chasing the raw power of flagship models like GPT-6 Astra, the developer community has long struggled with the prohibitive costs of running agentic workflows and complex computer-use tasks at scale. GPT-6.1 Sol addresses this friction directly, offering a massive leap in efficiency and reasoning for coding-centric workloads.
This isn't just a incremental patch; it is a strategic pivot toward making 'agentic coding' economically viable. By narrowing the performance gap between mid-tier and flagship models, OpenAI is enabling developers to build autonomous software engineers and complex automation agents that were previously too expensive to run in production environments.
- Targeted at agentic coding and professional computer use
- Significant performance uplift over GPT-6 Sol
- Extreme cost-efficiency for high-reasoning tasks
Key Features & Architecture
GPT-6.1 Sol is engineered specifically for high-reliability reasoning and long-context technical tasks. While OpenAI has not disclosed the exact parameter count, the model demonstrates a highly optimized architecture designed to handle the iterative nature of software development, where context must be maintained across multiple file changes and terminal executions.
A standout feature of this release is the substantial improvement in alignment and safety. Unlike previous iterations, GPT-6.1 Sol shows a marked resistance to prompt injection and attempts to bypass automated safety reviewers. This makes it a much more reliable candidate for autonomous agents that interact with live terminal environments and sensitive file systems.
- Enhanced alignment and safety protocols (System Card Addendum available)
- Optimized for agentic workflows and computer-use tasks
- High-efficiency reasoning with reduced error rates on difficult prompts
Performance & Benchmarks
The benchmark data for GPT-6.1 Sol is nothing short of transformative for the coding category. In the DeepSWE v1.1 benchmark, which measures software engineering capabilities, the model matches the performance of the flagship GPT-6 Astra while operating at roughly one-fifth of the cost. It notably eclipses its predecessor, GPT-6 Sol, by a massive 6.4 points at lower reasoning effort levels.
When evaluated on OSWorld 2.0 for offline computer use, GPT-6.1 Sol outperforms GPT-6 Sol by 7 points at maximum effort. Most impressively, it reaches within 2.1 points of the state-of-the-art Astra model, but achieves this at only one-seventh of the cost per task. In the specialized Terminal-Bench-Science 0.1 benchmark, the model averages just $5.47 per task, compared to the $23.21 required by Claude Opus 5.5 and $23.80 for Astra, representing a 75% cost reduction.