OpenAI has launched GPT-6.1 Sol, an upgraded model in the GPT-6 series designed to bring capabilities closer to GPT-6 Astra while keeping API costs substantially lower. The model improves on GPT-6 Sol across coding, professional workflows, computer use and factuality, with OpenAI reporting near-Astra performance on several agentic benchmarks.
The release arrived on September 29, 2026, less than a week after GPT-6 Sol was introduced. OpenAI is positioning GPT-6.1 Sol primarily around complex, multi-step work where AI agents need to operate across codebases, applications and business workflows rather than simply generate individual responses.
Quick Summary
- OpenAI launched GPT-6.1 Sol as an upgrade to GPT-6 Sol.
- The model focuses on agentic coding, computer use and long-running AI workflows.
- OpenAI reports 75.2% on DeepSWE v1.1, compared with 68.8% for GPT-6 Sol.
- On OSWorld 2.0, GPT-6.1 Sol scored 71.4%, compared with 73.5% for GPT-6 Astra.
- API pricing is $2 per million input tokens and $10 per million output tokens.
- Cached input costs $0.10 per million tokens, according to OpenAI.
- OpenAI reports a 32% reduction in factual-error responses on its difficult factuality evaluation.
- The model is positioned as a lower-cost option for sustained, multi-step agentic workloads.
What Is GPT-6.1 Sol?
GPT-6.1 Sol is an upgrade to GPT-6 Sol within OpenAI’s GPT-6 model family. OpenAI says the new model is intended to provide a balance between capability and cost, particularly for sustained agentic workloads.
The company reports improvements in several areas, including software engineering, document-based professional tasks, computer interaction and scientific workflows. On a number of evaluations, GPT-6.1 Sol comes close to GPT-6 Astra while operating at substantially lower cost.
That positioning makes the model particularly relevant to developers building AI agents that may need to make repeated API calls, reuse context and complete long-running tasks.
GPT-6.1 Sol Improves Coding and Agentic Work
Coding is one of the main areas where OpenAI reports a significant improvement over GPT-6 Sol.
On DeepSWE v1.1, a benchmark focused on long-horizon software-engineering tasks in real codebases, GPT-6.1 Sol achieved a 75.2% score at high reasoning effort. GPT-6 Sol’s reported best score was 68.8%, putting the newer model 6.4 percentage points higher. OpenAI also says GPT-6.1 Sol reaches comparable performance to GPT-6 Astra on the benchmark at roughly one-fifth of Astra’s cost.
The benchmark is designed around AI agents solving original software-engineering problems rather than isolated coding questions. That distinction matters because long-running coding agents need to inspect existing code, make changes, run tools and maintain context across multiple steps.
Computer Use Moves Closer to GPT-6 Astra
GPT-6.1 Sol also targets computer-use workloads, where AI agents interact with applications and complete multi-step tasks.
On the offline set of OSWorld 2.0, OpenAI reports a 71.4% score for GPT-6.1 Sol at maximum reasoning effort. GPT-6 Astra scored 73.5%, leaving a 2.1-percentage-point difference between the two models.
The cost difference is considerably larger. OpenAI says GPT-6.1 Sol reaches its reported OSWorld performance at roughly one-seventh of Astra’s cost per task.
Compared with GPT-6 Sol, the new model improves by seven percentage points on the same evaluation setting. The result suggests that OpenAI is targeting computer-use agents as one of the main workloads where the improved Sol model can deliver higher capability without the cost associated with its larger Astra model.
Business Workflow Performance Also Improves
The model’s improvements extend beyond coding and computer control.
On AutomationBench, which evaluates agents completing end-to-end business workflows using tools across areas such as sales, marketing, operations, support, finance and HR, GPT-6.1 Sol improves by 4.8 percentage points over GPT-6 Sol at medium reasoning effort. OpenAI also reports that GPT-6.1 Sol scores 2.2 points higher than Opus 5.5 in the same setting, at roughly one-third of the cost.
This is relevant for applications where an AI agent has to perform several connected actions rather than answer a single prompt. Examples include updating records, navigating business applications, processing documents or coordinating tasks across multiple tools.
OpenAI also reports strong performance on GDP.pdf, a benchmark involving complex professional documents containing tables, charts, diagrams and fine-print information. The company says GPT-6.1 Sol approaches GPT-6 Astra’s performance at around one-fifth of Astra’s cost per task in that evaluation.
Factuality Improves at Lower Reasoning Effort
GPT-6.1 Sol also reports an improvement in factual accuracy compared with GPT-6 Sol.
At low reasoning effort, OpenAI’s difficult factuality evaluation found that the share of responses containing at least one factual error fell from 11.4% with GPT-6 Sol to 7.7% with GPT-6.1 Sol. That represents an approximately 32% reduction in the measured error rate.
There is an important limitation to this result. OpenAI says the evaluation uses de-identified conversations where users had previously flagged factual errors. The company explicitly notes that these deliberately difficult prompts are not representative of typical usage, where factual errors are less common.
The figure should therefore be understood as a benchmark result rather than a general error rate for everyday GPT-6.1 Sol usage.
GPT-6.1 Sol API Pricing
One of the most significant parts of the release is the model’s API pricing.
| Pricing | GPT-6.1 Sol |
|---|---|
| Input | $2 per million tokens |
| Cached input | $0.10 per million tokens |
| Output | $10 per million tokens |
OpenAI says cached input is priced at just $0.10 per million tokens, representing a 95% reduction from its standard input price and a 50% reduction compared with GPT-6 Sol’s cached-input pricing.
Cached input can be particularly relevant to agents that repeatedly reuse information across requests. For applications involving long-running workflows or large codebases, reducing the cost of repeated context can have a meaningful effect on overall API expenditure.
Availability
GPT-6.1 Sol is available through the OpenAI API under the model ID gpt-6.1-sol. OpenAI also made it available to Plus, Pro, Business, Enterprise and Edu users in ChatGPT Work and Codex at launch.
However, the model was not yet available in regular ChatGPT at the time of the announcement.
OpenAI also said it plans to offer an Ultrafast version in the following days, with up to eight times faster token generation compared with the standard speed in Codex.
GPT-6.1 Sol Safety and Deployment
OpenAI says GPT-6.1 Sol underwent additional safety and alignment evaluations before deployment.
The company reports improvements over GPT-6 Sol in areas including transparency about broken tools, following explicit restrictions and avoiding unauthorized outcomes during agentic tasks. It also says the model made no attempts to bypass an automated safety reviewer in the tested evaluation.
OpenAI’s accompanying safety documentation provides additional context. GPT-6.1 Sol uses the same safeguards stack as GPT-6 Astra and is classified by OpenAI’s Preparedness Framework as Critical for cybersecurity capabilities and High for biological and chemical capabilities.
The safety evaluations are deliberately challenging tests rather than measurements of ordinary user traffic, so their results should be interpreted within that testing context.
Why GPT-6.1 Sol Matters for AI Agents?
The central change with GPT-6.1 Sol is not simply a higher benchmark score. OpenAI is combining improved agentic capabilities with significantly lower token and task costs.
The model is aimed at workloads that can consume large amounts of context and require repeated tool interactions, including software engineering, business automation, document analysis and computer-use agents.
For developers, the combination of stronger performance and a $0.10-per-million-token cached-input price could make repeated-context agent workflows more economical. At the same time, OpenAI’s benchmark results are company-reported and were produced in its research environment or through its API; the company notes that evaluation conditions can differ from production ChatGPT because of differences in system prompts, tools and reasoning settings.
GPT-6.1 Sol therefore represents an incremental but substantial upgrade within the GPT-6 family, with OpenAI emphasizing agentic coding, computer use and professional workflows at a lower operating cost rather than introducing a completely new model class.
Also Read-
GPT-6 Astra: Features, Capabilities, Performance & What’s New?
GPT-6 Models Guide: Compare Capabilities, Performance & Pricing
GPT-6 Explained: Everything to Know About OpenAI’s Next AI Model
GPT-6 Sol and Luna Launch With Lower API Costs
Source
GPT-6.1 Sol System Card Addendum
