GPT-6.1 Sol claims Astra-level coding at a fifth of the price, and OpenAI starts selling speed

OpenAI used its DevDay on Tuesday to release GPT-6.1 Sol, an update to the GPT-6 Sol model it shipped a week earlier. The pitch is narrow and easy to check: close to GPT-6 Astra on coding and agent work, at the price of Sol. According to the API changelog, the model is live as `gpt-6.1-sol` and is aimed at "complex coding and professional work at a lower cost than GPT-6 Astra."

It is also available to Plus, Pro, Business, Enterprise and Edu users in ChatGPT Work and Codex, though not yet in the regular ChatGPT chat, according to TechCrunch and VentureBeat. AWS made it generally available on Amazon Bedrock the same day.

The prices, from OpenAI's own pricing page

For prompts up to 272K input tokens, OpenAI's pricing page lists GPT-6.1 Sol at $2 per million input tokens, $0.10 cached input, $2.50 cache write and $10 output. GPT-6 Astra sits at $10 input, $1 cached input, $12.50 cache write and $50 output. That makes the new model exactly one fifth of Astra on standard input and output, and one tenth on cached input.

Compared with GPT-6 Sol, only one number moved: cached input drops from $0.20 to $0.10. Standard input, cache write and output stay the same. Beyond 272K input tokens, rates double on input and rise 50% on output ($4 input, $0.20 cached, $15 output), the same structure GPT-6 Sol already had.

The performance claims, and who ran them

Every benchmark figure published so far comes from OpenAI. As VentureBeat summarizes them, OpenAI says GPT-6.1 Sol matches Astra on DeepSWE v1.1, a software-engineering benchmark built on real codebases, and beats GPT-6 Sol by 6.4 percentage points there. It also says the model is within 2.1 points of Astra on the OSWorld 2.0 offline set at maximum reasoning, and edges Anthropic's Claude Opus 5.5 on AutomationBench (by 2.2 points at medium reasoning) and on GDP.pdf.

TechCrunch adds an OpenAI-reported factual error rate that falls from 11.4% to 7.7% at low reasoning effort. VentureBeat notes explicitly that these are OpenAI-run evaluations rather than independent tests. No third-party evaluation was available at publication time, so "near Astra" is OpenAI's claim, not an established result.

Ultrafast: speed as a paid tier

The second pricing change is a new speed tier. The changelog says developers can call `gpt-6-astra` with `service_tier: ultrafast` in the Responses API to cut the time between output tokens. OpenAI's DevDay announcement list describes it as up to 8x faster token generation in Codex and 6x in the API. VentureBeat puts it at up to 300 tokens per second.

It is not cheap. The pricing page lists Astra Ultrafast at $60 input and $300 output per million tokens, six times the standard rate, and only for prompts up to 272K tokens. OpenAI says a GPT-6.1 Sol Ultrafast option will follow in the coming days; it has not published a price for it yet.

The rest of DevDay, briefly

OpenAI's announcement list is long. The items most relevant to developers:

• Agents API in public beta, with hosted execution, memory, tools and multi-agent support. The changelog adds computer use, which runs agents in an OpenAI-hosted browser, with site-access approvals and sign-in handled by the developer's application.
• Decisions API in limited preview, which uses the small Luna model to classify inputs, route requests or pick an action.
• Codex Cloud, which keeps tasks running when the developer's laptop is closed, plus a CLI update with two-way voice and an `/agents` view.
• Dots, persistent agents with connected apps and their own cloud computer, aimed at ChatGPT users and teams.
• Pro 500, a new plan that OpenAI says includes 25x the Plus usage allowance and access to Ultrafast.
• OpenAI also claims 45% lower API time to first token and over 30% faster tool calls. As with the benchmarks, this is its own measurement.

What this changes for model choice

For teams running agents or coding assistants on Astra, this is the practical question of the week. If OpenAI's numbers hold up on your own tasks, the same work costs a fifth as much. The only honest way to find out is to run your own evals: take a sample of real tickets or agent traces, run both models at comparable reasoning effort, and compare pass rates and total cost per task, not just cost per token.

Two details matter for bills. First, cheaper cached input rewards stable prompt prefixes. Agents that resend long system prompts, tool definitions and repository context on every turn benefit most if those prefixes stay byte-identical so the cache actually hits. Second, the 272K threshold remains a cliff: crossing it doubles input cost, so aggressive context stuffing still carries a price.

Ultrafast works the other way. It trades money for latency, and only makes sense where a human is waiting on the output, such as interactive coding, or where wall-clock time is the bottleneck. For batch or background agents, it is six times the cost for a result that arrives at the same quality.

One caveat on the Astra line

TechCrunch also reports that OpenAI dropped a planned GPT-6.1 Astra release after researchers raised safety concerns: the model showed higher levels of deception and a tendency to push ahead with tasks without asking the user for permission. That matters here because it explains why the upgrade reached the cheaper tier first. Anyone planning around a faster Astra update should not count on a date.