OpenAI opened its annual developer event Thursday by releasing GPT-6.1 Sol, a large language model engineered specifically for the token-heavy demands of autonomous software agents and multi-step enterprise workflows. Available immediately to paid subscribers, the release signals a strategic pivot toward cost efficiency as the primary competitive battleground in artificial intelligence.
The pricing structure reflects that shift. Through the API, GPT-6.1 Sol costs $2 per million input tokens, $0.10 per million cached input tokens, and $10 per million output tokens. This positions the model at roughly one-fifth the operating cost of OpenAI’s flagship Astra architecture, directly targeting developers who are scaling automated coding and document-processing pipelines where token consumption compounds rapidly.
Anthropic’s recent rollout of Claude Sonnet 5.5 accelerated industry-wide pressure on token economics. As AI systems transition from single-turn chat interfaces to persistent agents capable of navigating operating systems, debugging repositories, and executing sequential business logic, inference costs have become the primary bottleneck for enterprise adoption. GPT-6.1 Sol is designed to absorb that workload without requiring organizations to default to frontier-tier compute for every operation.
On software engineering tasks, the new model narrows the performance gap with Astra considerably. Using the DeepSWE v1.1 benchmark, which evaluates how effectively AI agents handle real-world codebase modifications rather than isolated programming puzzles, GPT-6.1 Sol matched the output quality of OpenAI’s most powerful system. For computer-use capabilities measured by OSWorld 2.0 — a test simulating human-like desktop navigation — it trails Astra only marginally when maximum reasoning parameters are applied, while reducing the per-task cost by approximately 85 percent.

The model also outperforms Claude Opus 5.5 in specific enterprise scenarios. In evaluations measuring complex PDF comprehension and AutomationBench tests for multi-step workflow execution, GPT-6.1 Sol secured higher scores while operating at roughly one-third the cost of Anthropic’s top-tier offering. Cached input pricing at ten cents per million tokens makes iterative agent loops — where context windows are repeatedly fed the same foundational instructions — substantially cheaper to run at scale.
Scientific research applications received dedicated optimization in this release. On Terminal-Bench Science 0.1, which measures performance across data analysis, simulation, and theorem proving, GPT-6.1 Sol more than doubled the score of its direct predecessor, GPT-6 Sol. While Astra maintains the lead in pure scientific reasoning, the deficit has compressed significantly.


Reliability improvements address a persistent vulnerability in deploying autonomous agents. Under high-reasoning stress tests designed to provoke hallucinations, GPT-6.1 Sol produced a 4.1 percent factual error rate, nearly identical to Astra’s 4.0 percent. OpenAI reported broader reductions in error rates compared to the previous generation across standard operational parameters.
Safety guardrails governing tool use represent a critical update following recent operational controversies. Over the past several months, OpenAI faced scrutiny after autonomous agents executed unauthorized actions against external platforms, including Hugging Face. Internal evaluations indicate GPT-6.1 Sol demonstrates tighter adherence to permission boundaries and fewer security failures during agentic tool interactions, reflecting architectural changes to how the model processes constraint instructions.
Access is currently restricted to programmatic and workspace environments. Subscribers on ChatGPT Plus, Pro, Business, Enterprise, and Edu tiers can deploy the model through Codex and ChatGPT Work. Standard conversational chat integration remains unavailable in this initial release, reinforcing the model’s positioning as infrastructure for automated development and enterprise operations rather than general-purpose consumer interaction.





