OpenAI released GPT-6.1 Sol. Its performance is close to the flagship Astra, but it costs only a fifth

Just a week after the introduction of the GPT-6 Sol model, OpenAI is releasing its improved version. According to the company, GPT-6.1 Sol approaches the flagship GPT-6 Astra in programming, document processing, and computer control, while charging only a fifth of its price for processed tokens. Developers and companies that have AI agents working on dozens of tasks daily will thus be able to accomplish much more work for the same budget.

In programming, it can match Astra, while costing five times less

The new model made the biggest leap in programming. In the DeepSWE 1.1 test, which assigns long tasks to agents in real software projects, OpenAI states that it matches GPT-6 Astra at roughly one-fifth of the cost. Furthermore, it surpasses the best result of its predecessor, GPT-6 Sol, by 6.4 percentage points, even with less reasoning time.

Expert document processing has also improved. In the GDP.pdf test, the model answers questions based on complex PDFs from finance, healthcare, or law, including tables, graphs, and small print. Here, according to OpenAI, it beats Claude Opus 5.5 at less than half the cost and is close to Astra. In the AutomationBench test, where the agent navigates multi-step corporate procedures with dozens of tools from sales to HR, it is 2.2 percentage points ahead of Opus 5.5 and costs approximately a third.

How does it perform in computer control and science?

The second strong discipline is tasks where AI itself clicks and types in applications instead of the user. In the OSWorld 2.0 test, GPT-6.1 Sol lags behind Astra by only 2.1 percentage points, while costing roughly a seventh. Compared to GPT-6 Sol, it improved by seven points at less than half the cost.

In the Terminal-Bench Science test, where agents analyze data, run simulations, or estimate model parameters, the new Sol more than doubled its predecessor’s score. One task, at maximum reasoning intensity, costs an average of $5.47, while Opus 5.5 consumes $23.21 and Astra $23.80. However, for the most demanding research, OpenAI continues to recommend Astra, which remains at the top of the test with 68.1%.

Error rate slightly decreased, but Astra still leads

For deliberately challenging queries where users previously caught ChatGPT making mistakes, the new model’s responses contain a factual error in only 4.1% of cases. GPT-6 Sol was at 4.5%, Astra at 4.0%. However, according to OpenAI itself, this test does not represent typical usage, where errors are expected to be even rarer.

The company also reports progress in so-called alignment with human intentions. GPT-6.1 Sol more openly admits what it cannot do and more reliably adheres to instructions and safety limitations. For example, if its search tool stops working, it conceals this and starts guessing in only 2.8% of cases. The older Sol failed this way in 4.9%, Astra in 1.5%, and the cheapest Luna even in 28.7% of cases.

Where is GPT-6.1 Astra? Reportedly, it won’t be released at all

Safety is precisely the reason why its larger sibling is missing alongside the new Sol. According to Monday’s reports, OpenAI canceled the release of GPT-6.1 Astra, which was supposed to arrive in October. According to The Wall Street Journal, internal tests showed the model deceived more often than its predecessor – for example, it did not always admit what steps it actually performed – and accessed external tools without permission.

Saachi Jain from OpenAI’s safety team, according to Reuters, acknowledges that the model did not achieve the desired level of adherence to instructions and granted permissions. The company intends to use the same base model for future versions of GPT-6, but first needs to identify the source of the problems.

Price and availability of GPT-6.1 Sol

The pricing remains unchanged compared to GPT-6 Sol: developers using the API will pay $2 per million input tokens and $10 per million output tokens. However, the cost of repeatedly used cached context has been reduced, now costing $0.10 per million tokens, which is half the previous price. This will primarily save costs for agent tasks where the model receives the same inputs repeatedly.

Subscribers of ChatGPT Plus, Pro, Business, Enterprise, and Edu have access to the new model starting today in ChatGPT Work and Codex; it is not yet available in the regular chat. Along with it, OpenAI also introduced the GPT-6 Astra Ultrafast and GPT-6.1 Sol Ultrafast variants, which are up to eight times faster. In addition to the API, users of the new Pro tier will get them for approximately 10,000 CZK ($500) per month.

Would you opt for the cheaper Sol, or would you pay extra for Astra?

Sources: OpenAI, Reuters