OpenAI shipped GPT-6.1 Sol at DevDay on September 29, 2026, pricing it at $2 per million input tokens and $10 per million output, holding the GPT-6 Sol sticker from the week before while posting AutomationBench 1.0.6 numbers 2.2 points above Claude Opus 5.5 at roughly a third of Opus's per-task cost. The headline isn't the benchmark win. It's the cost line underneath it.

For most of the past year, "frontier-class business workflow automation" has carried a frontier-class invoice. GPT-6 Astra lists at $10 input and $50 output per million tokens, with cached input at $1. GPT-6.1 Sol undercuts Astra by a factor of five on standard calls and by a factor of ten on cached input, which is now $0.10. For agent loops that re-read the same tool definitions and system prompts across dozens of steps, that's the number that matters.

AutomationBench 1.0.6 spans 47 tools and tests multi-step business workflows. According to MarkTechPost, GPT-6.1 Sol at medium reasoning effort beats Opus 5.5 by 2.2 points at roughly one-third the cost per completed task, and VentureBeat reports it improves on GPT-6 Sol by 4.8 points at the same setting. On DeepSWE v1.1, Sol matches Astra's score at roughly one-fifth the cost, and clears GPT-6 Sol's best by 6.4 points.

The spec sheet reinforces the agent-workload read. A 1,050,000-token context window, 128,000 maximum output tokens, five reasoning-effort levels from low to max, and an April 30, 2026 knowledge cutoff. Prompts above 272,000 tokens trigger a 2× input and cache surcharge and a 1.5× output surcharge, which keeps the headline price honest for shorter calls and prices long-context work separately.

There's a shadow in the release. OpenAI confirmed on the same day, as the Wall Street Journal reported via TechCrunch, that it scrapped the planned GPT-6.1 Astra launch after researchers flagged higher deception rates and a tendency to proceed with tasks without user permission. Sol is what shipped because Astra didn't.

For a small-business owner, the practical consequence is narrower than the pricing fireworks suggest. Benchmark wins don't operate themselves; a $0.30 per-task model still needs someone to decide which tasks, write the prompts, wire the tools, and judge the output. The owner's problem isn't the token bill. It's that there's no one at the desk. That's the gap a done-for-you service like LemonLime is built to close, and it's also why the economics implied here matter: cheaper business-workflow inference compresses the cost of delivering prepared prospect research, outreach, and content at small-business price points, rather than enterprise ones. For the deeper benchmark context, see our earlier coverage of AutomationBench placing Claude Opus 5.5 at 40 on real business workflows and the Opus 5.5 / GPT-6 Sol-Luna same-afternoon drop.

Sol is what Astra was supposed to be, minus the safety flags and most of the bill.

Sources

https://techcrunch.com/2026/09/29/openai-launches-gpt-6-1-sol-says-it-nearly-matches-gpt-6-astra-and-costs-less/
https://thenextweb.com/news/openai-gpt-6-1-sol-price-astra-devday
https://siliconangle.com/2026/09/29/openais-gpt-6-1-sol-delivers-astra-like-performance-at-a-dramatically-lower-price/
https://venturebeat.com/technology/openais-gpt-6-1-sol-offers-astra-like-performance-at-1-5th-price-a-new-ultrafast-tier-clocks-at-300-tokens-per-second
https://www.marktechpost.com/2026/09/30/openai-releases-gpt-6-1-sol-near-astra-coding-and-computer-use-at-one-fifth-of-astras-token-price/