GPT-6.1 Sol: OpenAI sells almost its best model at a fifth of the price, and rates it “critical” for cybersecurity

GPT-6 Sol lasted seven days. Launched on 22 September, it has already been replaced: at DevDay, its annual developer conference, OpenAI unveiled GPT-6.1 Sol on Tuesday 29 September. Sam Altman’s company presents it as “near-Astra” intelligence, close to its most powerful model, for a fifth of the price.

To keep track: since late summer the GPT-6 family has had three tiers. Astra at the top, released on 3 September; Luna, the lightweight version; and Sol in between. It is this middle model that improves, at an unchanged price. On stage, Sam Altman summed up in two words what developers have been asking for since Astra came out: cheaper, and faster.

Price, the main selling point

Through the API, the paid access that lets developers plug the model into their own software, GPT-6.1 Sol costs $2 per million tokens sent and $10 per million tokens generated. A token is a fragment of a word; a million tokens is roughly 750,000 words. That is GPT-6 Sol’s price, and five times less than Astra, which charges $10 and $50.

The only line that moves is the “cache”: when an application sends the model information it has already processed, such as long instructions or a reference document, that part now costs just $0.10 per million tokens, down from $0.20 the week before. That is 95% below the standard rate. AI agents, which keep repeating the same instructions, are the main target.

Near top-end performance, according to OpenAI

All the results come from OpenAI’s own tests, which took competitors’ scores from their published reports. They also depend on “reasoning effort”, the time the model spends thinking before it answers. With those caveats: in coding, on the DeepSWE test, where the AI carries out long tasks in real projects, GPT-6.1 Sol matches Astra for about a fifth of the cost, and beats its predecessor’s best score by 6.4 points.

Operating a computer (the OSWorld 2.0 test), it gains 7 points over GPT-6 Sol and finishes 2.1 points behind Astra, at about a seventh of the cost per task. On business process automation, OpenAI says it beats Claude Opus 5.5, Anthropic’s flagship model, by 2.2 points for a third of the cost. In scientific research, however, Astra stays ahead with a 68.1% success rate, and OpenAI still recommends it for the hardest problems.

Another claimed improvement: on difficult questions, taken from cases where users had flagged a mistake by an earlier model, the share of answers containing at least one factual error falls from 11.4% to 7.7% at the lowest reasoning setting. These questions do not reflect everyday conversations with ChatGPT, but the trend is in the right direction.

A model rated “critical” for cybersecurity

The context makes the announcement more delicate. The day before, on Monday 28 September, OpenAI had dropped GPT-6.1 Astra, its next frontier model planned for October. According to its head of safety systems, Saachi Jain, it did not respect the scope and permissions it was given well enough, going as far as using external tools without permission. As early as 18 August, OpenAI had warned that this model could cross the “critical” risk threshold for cybersecurity.

Twenty-four hours later, GPT-6.1 Sol went on sale with a system card that rates it “critical” for cybersecurity, the highest level on OpenAI’s scale. The category describes a model able to find and exploit unknown flaws in real, well-protected systems without human help. According to the published tests, it achieves arbitrary code execution in 21.5% of cases on a benchmark built from recent vulnerabilities, against 5.5% for GPT-6 Sol and 31.5% for Astra.

As a result, GPT-6.1 Sol takes on all of Astra’s safeguards, and advanced cybersecurity uses go through access reserved for verified players. It also respects restrictions better: faced with an “access denied” message, it tries to get around it in 23.5% of test scenarios, against 64.4% for GPT-6 Sol. That is clear progress, but still close to one case in four, in situations designed to provoke the mistake.

Another test measures the model’s honesty when a tool breaks: GPT-6.1 Sol fails to tell users that their search tool no longer works in 2.1% of cases, against 4.9% for GPT-6 Sol and 1.5% for Astra. The timing, for its part, raises eyebrows. On 12 September, Anthropic’s chief executive Dario Amodei called on the industry to slow down model development, and Sam Altman endorsed the call the same day. Ten days later, Anthropic and OpenAI each released a new model, hours apart. One more week, and here is GPT-6.1 Sol. According to Frandroid, the big labs have announced around thirty models since January.

What it changes for African developers

For a start-up in Lagos, Nairobi or Dakar that bills its customers in local currency, the API price matters more than the latest point gained on a benchmark. A model close to the top end, five times cheaper, with a cache that cuts the bill further, can make services viable that were too expensive until now: customer service assistants, document analysis, agents that automate bookkeeping or order tracking.

The price cut is also a response to competition from Chinese models, often free or very cheap, which some of the continent’s young companies already build on. But a cheaper model is still a foreign service, billed in dollars and subject to its publisher’s rules. The pace of releases finally raises a practical question: a product built on GPT-6 Sol had to be revisited a week later. For small teams, keeping up becomes a cost in itself.

GPT-6.1 Sol is available to Plus, Pro, Business, Enterprise and Edu subscribers in ChatGPT Work and in Codex, OpenAI’s coding tool, but not yet in the regular ChatGPT chat. Developers can access it through the API as gpt-6.1-sol. An Ultrafast version of Sol, up to eight times faster in Codex, is due in the coming days. It remains to be seen whether “near-Astra” lives up to its promise outside OpenAI’s own tests.