GPT-6.1 Sol launches at DevDay and goes GA on Bedrock, claiming near-Astra results at a fifth of the cost

GPT-6.1 Sol is the day's headline model. According to OpenAI, it approaches GPT-6 Astra on demanding evaluations at about one-fifth of the cost. AWS's Matt Wood wrote that on OpenAI's software-engineering evaluation, Sol 'matches GPT-6 Astra at roughly one-fifth of the cost per task,' and that it approaches Astra on complex document tasks and computer use. The model shipped on Amazon Bedrock as generally available on launch day, which shows how tight the OpenAI–AWS partnership has become.

OpenAI pitches Sol for long, multi-step agent work: understanding an unfamiliar codebase, changing databases and APIs, building interfaces and testing end to end. Independent early data is encouraging. LlamaIndex's Jerry Liu benchmarked Sol on document OCR and found table parsing and reading order clearly improved over GPT-6 Sol from a week earlier, now on par with Astra.

The competitive context is sharp. On Hacker News, the launch thread drew 805 points and 743 comments, and commenters framed the pricing as 'shots fired, half the price of Opus.' Sol arrives the same week as Anthropic's Sonnet 5.5, and while DeepSeek V4.1 Flash undercuts everyone on price. The mid-tier is now the main battleground.

Skeptics note that the original GPT-6 Sol was 'far worse' in practice than its benchmarks suggested, and they want to see refactor quality on messy real codebases. Sol's launch is also shadowed by the shelving of Astra itself. OpenAI's premium model is the reference point for Sol's claims but is not shipping. Watch for independent SWE and agent benchmarks over the next week.