OpenAI Launched Another Model. The Pelican Drawings Look the Same.

OpenAI shipped GPT-6 Astra on September 3. It shipped GPT-6 Sol on September 22. On September 29, at DevDay, it shipped GPT-6.1 Sol. Three model launches in twenty-six days is not progress. It is what happens when a company cannot stop moving because standing still means admitting the last model was not that different from this one.

The verdict: GPT-6.1 Sol is not a new frontier. It is a markdown.

OpenAI's own framing gives it away. The pitch was "Near-Astra level intelligence at a fifth of the price." Translation: we are not claiming this beats our best model. We are claiming it is close enough that you should stop paying full price for the one that does.

The numbers back the translation. On OSWorld 2.0, which measures PC operating performance, GPT-6.1 Sol costs one-seventh the price of GPT-6 Astra while keeping the gap to a mere 2.1 percentage points. The largest factual-accuracy gain appears at low reasoning effort, where the error rate drops from 11.4% to 7.7%. Across all reasoning settings the new model's error rate stays within 1.9% of Astra. A 2-point gap at a fifth of the price is not a new tier of intelligence. It's a pricing decision wearing a research paper.

The squint test

Simon Willison runs the same test on every model release: he asks it to draw a pelican riding a bicycle, because the task is silly enough to expose real differences without a benchmark to game. His read on GPT-6.1 Sol: the pelicans "are not notably different from the GPT-6 family pelicans." A frontier release is supposed to make you say "oh." This one made a very online LLM obsessive shrug.

Where OpenAI has a real number is cost-per-task. At maximum reasoning effort, the cost per task was $5.47 for GPT-6.1 Sol versus $23.80 for GPT-6 Astra and $23.21 for Claude Opus 5.5. That's the actual headline, and it's a good one for enterprise buyers. It's also not a capability story. It's a procurement story.

Follow the money

This release makes sense only if you stop listening to OpenAI's pitch. On DeepSWE v1.1, a benchmark for complex software-engineering tasks, GPT-6.1 Sol matches Astra at one-fifth of the cost while exceeding the previous GPT-6 Sol by 6.4 percentage points. On AutomationBench, which tests multi-step business workflows, it scores 2.2 percentage points above Opus 5.5 on OpenAI's own tests.

That's a genuinely useful model for anyone running agents at volume. It's not a reason to believe OpenAI widened its lead on Anthropic or its own flagship. It's a reason to believe the mid-tier is now good enough that paying for the top tier needs better justification than it used to. GPT-6.1 Sol is rolling out in ChatGPT Work and Codex first, not consumer Chat. That tells you exactly who this release is for: the enterprise accounts OpenAI needs to keep spending, not the hype cycle.

If you run an agent fleet on Astra today, the question is no longer whether 6.1 Sol is good enough. The benchmarks say it's close. The question is why you're still paying five times more for two percentage points.

What to watch: whether enterprise customers actually downgrade workloads from Astra to 6.1 Sol in the next quarter, whether Anthropic answers with its own mid-tier price cut, and whether OpenAI's next "frontier" announcement bothers to compare itself to Astra at all.