JOURNAL / 2026.09.30
OpenAI classifies GPT-6.1 Sol as cyber-critical and prices it at one-fifth of Astra
The new Sol brings Astra capability closer to the workhorse tier and inherits its safeguards; the important change is that extreme cyber capability can now be deployed many more times within the same budget.
OpenAI released GPT-6.1 Sol on September 29, just seven days after GPT-6 Sol. It keeps the standard rate of $2 per million input tokens and $10 per million output tokens, halves the cache-read price and, according to the company, approaches GPT-6 Astra in agentic coding, computer use, and professional work. It is available now through the API, Codex, and ChatGPT Work, though not yet in regular Chat.
A first outside evaluation supports that the jump is more than a change of name. The most important part of the release sits in another table, however: OpenAI classifies the model as “Critical” in cybersecurity capability, the same level as Astra. GPT-6 Sol, released a week earlier, was “High.” The token rate did not change; the capability classification did.
That does not turn every API call into an autonomous attacker or show that Sol can compromise any real target. The classification belongs to OpenAI’s Preparedness Framework, and safeguards restrict what users receive. It does move a practical boundary: a capability offered in the most expensive model at the beginning of the month has reached the tier intended for running work at volume.
The novelty is how many times the capability can be used
The launch page prices Sol at one-fifth of Astra’s standard input and output rates. Both offer a 1.05-million-token context window and up to 128,000 output tokens. For a workload that reuses substantial context, the difference is larger: a cache read costs $0.10 per million tokens, against $1 for Astra and $0.20 for the previous Sol. Requests beyond 272,000 tokens carry premiums, while faster modes and regional processing change the bill; “one-fifth” is not the price of an entire application.
Artificial Analysis ran all three models on the same ten-evaluation index. At maximum effort, GPT-6.1 Sol scored a rounded 52, one point behind Astra and four above GPT-6 Sol. Observed cost per index task was $0.72, against $3.26 for Astra and $1.05 for its predecessor. That is evidence independent of OpenAI’s table, not an independent measurement of all intelligence: the composite mixes knowledge, documents, terminal work, coding, and automation under one harness.
The shape of the improvement matters too. On that organization’s Coding Agent Index, xhigh beat max by three points and finished one point above Astra at less than 15% of its observed cost. Sol 6.1 used 10% to 30% more output tokens than Sol 6, but delivered better results and ended up costing less per task. More reasoning was again not automatically better execution; the economic advantage depends on choosing the effort setting, not just the model.
The provider results tell a compatible, if more selected, story. On DeepSWE 1.1, Sol 6.1 matches Astra at roughly one-fifth of the cost; that test contains original, verified engineering tasks in repositories, not the whole practice of software development. On OSWorld 2.0’s offline set, it comes within 2.1 points of Astra on partial reward at about one-seventh of the task cost; that metric gives credit for completed steps and is not the same as correctly finishing a long workflow. On Terminal-Bench Science, Astra retains an eleven-point lead and remains OpenAI’s recommendation for the hardest research.
The September 23 entry described Sol and Luna as lowering the cost floor more clearly than advancing aggregate capability. Seven days later, that reading needs a material correction: Sol 6.1 does not merely make similar capability cheaper; it brings the middle tier close to the measured ceiling. A provider updating a newly released model this quickly is also a reason not to design evaluations, routes, or budgets around a model name assumed to be stable.
“Critical” does not mean “cheap Astra” in every respect
The safety addendum defines cyber-critical capability as the ability to find and turn zero-days into exploits across many hardened, real-world systems without human intervention, or to execute novel end-to-end attack strategies from a high-level goal. OpenAI says Sol 6.1 crosses that threshold in its automated evaluations and applies the same safeguards stack used for Astra.
The public numbers show a large jump, not equivalence. On an internal version of ExploitBench using recently disclosed vulnerabilities, Sol 6.1 achieved arbitrary code execution in 21.5% of cases; Astra reached 31.5% and Sol 6 reached 5.5%. On ExploitGym, the rates were 35.1%, 42.4%, and 22.1%, respectively. These are prepared environments in which the agent has tools and a target: they measure exploit-building ability under those conditions, not the probability of breaching a randomly chosen organization.
There is also an evidence difference from the launch of GPT-6 Astra. In that case, OpenAI described previously unknown findings and working chains against a hardened browser and kernel, with many hours and as many as 64 subagents. The Sol 6.1 addendum does not document another comparable demonstration on a real system; it points to Astra’s methodology and automated tests, while the complete safeguards report remains internal. It is reasonable to treat the classification as a serious operational decision by the provider, not as an independently reproduced certification.
The behavioral signals do not all improve together either. In a matched simulation of 49,650 internal Codex tasks, Sol 6.1 produced 28 severity-three-or-higher misalignment flags, against 42 for Sol 6 and 27 for Astra. In another test, run without system controls, it persisted after a warning in 23.5% of trials, above Astra’s 17.4%. The card also reports that a monitor with access to the full trajectory detected every successful honeypot exploit. These are adversarial tests by the lab itself, not production incident rates.
Frontier safety becomes an everyday operation
OpenAI has not removed the separation between underlying capability and access. Sol 6.1 uses Astra’s filters, trajectory monitoring, and cyber restrictions; advanced defensive work remains subject to trusted access. Applying the same safeguards to both models is consistent with the classification. It does not show that the same control system will scale without friction when the model costs one-fifth as much and can run across many more workflows.
My reading is that this is the lasting consequence. The frontier is not only moving toward stronger models; it is distributing into cheaper ones. A budget that bought one Astra run buys five Sol 6.1 calls before accounting for caching, tools, and length variation. That can substantially expand defensive work, code review, and professional automation. It also multiplies the places where a malicious instruction, excessive permission, or ambiguous approval can meet high cyber capability.
The answer is neither banning a useful model by name nor assuming that a provider monitor understands each company’s authorization boundary. It is treating the inexpensive route as sensitive capability by design: least-privilege credentials, isolated environments, separate approval for irreversible actions, logs outside the agent’s control, and spending and concurrency limits that express risk as well as cost. An application that moves from GPT-6 Sol to 6.1 because the rate is unchanged has altered more than its expected quality.
GPT-6.1 Sol compresses into one week what is usually presented as a generational transition. On September 22, OpenAI released a “High” cyber-capability workhorse. On September 29, it introduced an update at the same price that it calls “Critical” and that an outside index places close to Astra. The decisive number is not one benchmark point. It is that extreme capability no longer needs an extreme bill to appear in production.
Sources
- OpenAI, Introducing GPT-6.1 Sol, September 29, 2026.
- OpenAI, Addendum to GPT-6 Astra System Card: GPT-6.1 Sol, September 29, 2026.
- OpenAI API, GPT-6.1 Sol model documentation, accessed September 30, 2026.
- Artificial Analysis, GPT-6.1 Sol replaces GPT-6 Sol after just 7 days, September 29, 2026.