OpenAI has introduced GPT-6.1 Sol, a new Sol-family model positioned much closer to GPT-6 Astra on complex coding, computer use and professional work without moving customers to Astra’s standard API price level.
The change is more important than a routine model refresh because it shifts the operating point for teams building agents. OpenAI lists GPT-6.1 Sol at $2 per million input tokens and $10 per million output tokens under standard processing, while its current model comparison lists GPT-6 Astra at $10 input and $50 output. GPT-6 Sol is also listed at $2 input and $10 output, so the basic uncached Sol price tier does not rise with the 6.1 update.
That makes the central question less about whether Sol is “better” in the abstract and more about whether a workload that previously justified Astra can now run acceptably on Sol.
A stronger model at the same Sol operating point
OpenAI describes GPT-6.1 Sol as delivering near-Astra performance for complex work at a lower cost than Astra. Its API documentation lists a 1,050,000-token context window and up to 128,000 output tokens, with tool support through the Responses API for web search, file search, code execution, computer use, MCP and other agent capabilities.
The company’s safety addendum also shows a meaningful jump over GPT-6 Sol on selected cybersecurity evaluations. On one arbitrary code-execution evaluation, OpenAI reports 21.5% success for GPT-6.1 Sol, versus 5.5% for GPT-6 Sol and 31.5% for GPT-6 Astra. On ExploitGym, it reports 35.1% for GPT-6.1 Sol, compared with 22.1% for GPT-6 Sol and 42.4% for Astra.
Those numbers are publisher-reported evaluation results, not a guarantee of production performance. They do, however, support the narrower conclusion that GPT-6.1 Sol is not merely a naming change: OpenAI is reporting a substantial capability increase within the Sol tier.
Agent workflows are where the economics matter most
The model is aimed directly at workloads where repeated tool calls and long reasoning chains can make model selection an operating-cost decision. Coding agents, computer-use systems and professional research workflows often invoke a model many times inside one user-visible task. A fivefold difference in standard input and output list pricing between Sol and Astra can therefore matter even when token efficiency varies by workload.
GitHub’s September 29 rollout makes that developer impact concrete. GitHub says GPT-6.1 Sol is generally available and rolling out in Copilot for agentic coding and terminal workflows, with access across its coding agent, command-line interface, major IDEs, github.com and mobile clients for eligible paid plans.
For Aipolix, the architectural implication is that model routing should become more empirical, not less. If Sol now reaches an acceptable quality threshold on a task that previously required Astra, routing that task downward can reduce cost materially. But if the task depends on Astra’s remaining capability margin, switching purely because Sol is cheaper can still reduce completion quality or increase retries.
The right unit of comparison is therefore cost per successful task, including retries and tool usage, rather than token price alone.
Cyber capability changes the control boundary too
The stronger capability comes with a security consequence. Under OpenAI’s Preparedness Framework, the company says it is treating GPT-6.1 Sol as Critical in cybersecurity and High for biological and chemical capability, and is applying the same safeguard stack used for GPT-6 Astra.
That classification matters for organizations deploying autonomous or semi-autonomous agents. A model that is more capable at code execution and vulnerability exploitation can improve defensive security work, but it also increases the importance of tool permissions, network controls, credential isolation, audit logs and human approval around high-impact actions.
The same model-selection decision can therefore move in two directions at once: operational cost may fall relative to Astra, while the required control environment becomes more demanding than teams may have associated with an ordinary “Sol” tier.
OpenAI’s published cyber results should not be treated as a universal measure of offensive capability. They are specific evaluations under specific conditions. The more defensible operational conclusion is that teams should re-evaluate their threat model when upgrading, rather than assume the security posture for GPT-6 Sol transfers unchanged to GPT-6.1 Sol.
The model hierarchy is becoming a routing problem
GPT-6.1 Sol does not make Astra obsolete. OpenAI still positions Astra as its most capable model, and its own cyber evaluations show Astra ahead on several measures. Availability also varies by product, plan, administrator policy and rollout state.
What changes is the space between the tiers. GPT-6.1 Sol gives developers a new point on the capability-cost curve: materially stronger than the prior Sol model in OpenAI’s published evidence, close enough to Astra to deserve direct workload testing, and priced far below Astra at standard API rates.
Aipolix’s conclusion is that teams should use the release to retest routing boundaries rather than simply replace one model name with another. Benchmark a representative set of real tasks, measure successful completion cost, include retry and tool-call overhead, and separately review the permissions granted to the model.
The upgrade is most valuable where it lets an application move demanding agent work onto a cheaper tier without increasing failure rates. The risk is assuming that stronger capability is only an economic benefit. In GPT-6.1 Sol, the same capability gain that improves agent performance also increases the importance of containment and governance.