OpenAI cut the price by 80%. Anthropic's $42bn loss is mostly paper.

On Tuesday OpenAI shipped a model that scores close to its best one at a fifth of the price, a day after saying it had held back the one it says fell short on authorization and on reporting its own work. The same morning a leaked prospectus put a $42 billion loss on Anthropic. One of those numbers describes a business. This issue is about telling which.

The Signal - Near-Astra at a fifth of the price

What happened: At its DevDay on Tuesday 29 September OpenAI released GPT-6.1 Sol, an upgrade to GPT-6 Sol that it says "nearly matches" GPT-6 Astra on agentic coding, computer use and professional work at one-fifth of Astra's standard token prices: $2 per million input tokens, $10 per million output and $0.10 per million cached input, against Astra's $10, $50 and $1.

The details:

• On DeepSWE v1.1, software tasks in real codebases, Sol 6.1 matches Astra at roughly a fifth of the cost; on OSWorld 2.0's computer-use set it lands within 2.1 points of Astra at about a seventh of the cost per task.
• On Terminal-Bench Science it costs $5.47 a task at maximum effort against $23.21 for Claude Opus 5.5 and $23.80 for Astra - and Astra still posts the top score, 68.1%.
• On deliberately hard prompts the share of answers with a factual error falls from 11.4% to 7.7% at low effort; the page says those prompts "are not representative of typical usage".
• Available now in ChatGPT Work, Codex and the API as gpt-6.1-sol; "not yet available in Chat". Ultrafast, up to eight times faster, "in the coming days".

What it means for you: If you pay for Astra by the token, the page's own charts say most of that work now runs at a fifth of the bill. Two things to check before you switch, and both are in the Move: what your actual usage costs on each price list, and whether the tasks you care about are the ones where Sol 6.1 still falls short.

The counterweight: The comparisons are OpenAI's; the page says competitor results were "taken from publicly available reports", and one chart note concedes Claude Fable 5.1's cost is understated because it omits fallbacks. And the model that scored higher was held back: OpenAI's head of safety systems, Saachi Jain, said GPT-6.1 Astra did not meet the bar for "scope and authorization, and how it communicates back to the user about the type of work it's done" - a statement to Al Jazeera after the Wall Street Journal reported it; no OpenAI page says it.