GPT-6.1 Sol: near-Astra work at a fifth of the price

The dots got the headlines at DevDay on Tuesday, but the model launch in the same hour may matter more to anyone paying an AI bill. @OpenAI announced GPT-6.1 Sol as "near-Astra intelligence for a fifth of the price," and called it the most cost-efficient model for its performance you can buy today.

That's a company grading its own product, so treat it as a claim. It's still worth a look, because the list price stayed put while the model got better.

What OpenAI actually shipped

In a follow-up post, @OpenAI put the number people will care about up front: cached input now costs $0.10 per million tokens. The company says that is 95% below standard input pricing and half of what GPT-6 Sol charged for cached input.

The developer account, @OpenAIDevs, pitched the model for "complex refactors, deep codebase investigations, and long-running agents across apps." It points to stronger agentic coding and computer use as the main upgrades.

OpenAI's model page lists these details:

• $2 per million input tokens and $10 per million output, the same list price GPT-6 Sol landed at after its last cut.
• A context window of about 1.05 million tokens, with up to 128,000 tokens of output.
• A catch for long jobs: prompts over 272,000 input tokens are billed at double the input and cache rates.
• Reasoning effort settings from low up to max.

TechCrunch reports that the factual error rate at low reasoning effort fell from 11.4% to 7.7%. It also reports that across effort settings the error rate stays within 1.9 points of GPT-6 Astra. The model is live in ChatGPT Work, Codex and the API for Plus, Pro, Business, Enterprise and Edu users.

There's a backstory. OpenAI scrapped its planned GPT-6.1 Astra release after internal testing found more deception and a habit of pressing ahead without asking the user, according to Wall Street Journal reporting that TechCrunch cites. So the 6.1 upgrade shipped on the cheaper model only, and OpenAI says this Sol scores better on its alignment evaluations and is more upfront about what it can't do.

The useful tip came from a user, not OpenAI

The most practical post of the day came from @Voxyz_ai, who argued that for coding, 6.1 Sol can replace Astra outright "at 1/5 the price." Their setup runs Sol on everything in Codex and calls Astra only for a final independent review before a big change ships.

The part I'd copy is about the effort dial. Reading OpenAI's DeepSWE chart, Vox says Sol scores highest on high effort, and xhigh or max actually score lower. Medium lands about two points below high and costs over 30% less per task. So in their setup the main agent runs on high and the helpers run on medium.

I haven't checked that chart myself, so take it as one practitioner's reading. It does match something I see a lot: people set every model to its highest setting out of habit and pay for that effort on every call, whether or not the answers improve.

My read

This is mostly a pricing story, which is good news if you're the one paying. When a model close to the top tier costs a fifth as much, a lot of small jobs that weren't worth automating start to pencil out: sorting inbound email, pulling line items off invoices, drafting quote follow-ups, checking a job schedule against a calendar.

I'd pay the most attention to the cached input price. Most business automations send the same long block of instructions every time (your price list, your policies, how you want customers addressed) and then a short new message. With caching, that repeated block costs very little after the first call.

I'm wary of the "near-Astra" label, though. Near means some tasks will come out worse, and a launch chart won't tell you which ones. Run it on your own work before you switch anything over.

What this means for a business adopting AI

This applies whether you use OpenAI, Anthropic or Google: which model you pick, and how you set it, is now a cost you can manage like any other. For everyday office work, a cheaper model on a sensible effort setting, with its fixed instructions cached, will often do as well as a premium model on max. Keep the expensive model for the few decisions where a mistake costs real money.

If you aren't sure which of your weekly tasks belong on which tier, New Face Design offers a free process audit. We map where your team's hours go, identify what a model like Sol could handle cheaply, and flag what should stay with a person.