OpenAI's pitch for GPT-6.1 Sol, released Tuesday at its DevDay event, fits on a sticker: near-Astra intelligence for a fifth of the price. Astra is the company's top model, and Sol is its mid-tier one. The claim has two halves, and they deserve different verdicts.
The price half is arithmetic anyone can check. The intelligence half rests almost entirely on tests OpenAI ran itself, plus one independent index that has so far reached the public only secondhand.
The price: supported
OpenAI's launch post lists GPT-6.1 Sol at $2 per million input tokens and $10 per million output tokens. Tokens are the small chunks of text a model reads and writes, and they are how developers are billed. GPT-6 Astra costs $10 and $50. Cached input, meaning text the system has already processed and can reuse, costs $0.10 for Sol against $1 for Astra.
That is one-fifth on standard rates, exactly as advertised. Some early coverage presented $0.10 as Sol's headline price. It is the cached rate.
Token price is not task cost, though. In OpenAI's own figures as reported, the saving per task is roughly one-fifth on DeepSWE v1.1, one-seventh per completed task on OSWorld 2.0, and $5.47 against $23.80 on Terminal-Bench Science 0.1. MarkTechPost also reports a surcharge on prompts over 272,000 input tokens: double the input and cache rates, and 1.5 times the output rate.
Near Astra: mostly OpenAI's own numbers
The headline line of OpenAI's announcement says Sol nearly matches Astra on agentic coding, computer use and professional work. The detail further down is narrower: Sol approaches Astra on several of OpenAI's evaluations, not all of them.
The evidence, all OpenAI's figures as relayed by outlets including Vellum and MarkTechPost:
- On DeepSWE v1.1, a software-engineering test, Sol matches Astra at roughly a fifth of the cost. It scores 75.2% at higher reasoning settings, against 68.8% for GPT-6 Sol at maximum effort, so the two figures come from different settings.
- On AutomationBench, which measures multistep business workflows, Sol falls just short of Astra while beating Anthropic's Claude Opus 5.5 by 2.2 points at about a third of the cost.
- On OSWorld 2.0, a computer-use test, it lands within 2.1 points of Astra.
- On a factuality test that Vellum says was built from errors users had flagged, the share of responses with a factual error at low reasoning effort fell from 11.4% to 7.7%. Across tested settings, Sol's error rate stays within 1.9 points of Astra's.
OpenAI also says Astra still leads on the hardest science tasks. None of these results has been independently replicated in the reporting so far, and the reporting reviewed does not give Astra's exact scores for several of the comparisons.
The one outside check
Independent benchmarker Artificial Analysis gives Sol 51.8 on its Intelligence Index at maximum effort, against 52.7 for Astra and 47.5 for GPT-6 Sol, according to a blog post by router company OrcaRouter. The same post puts the cost per index task at $0.72 for Sol and $3.26 for Astra.
That broadly backs OpenAI. But OrcaRouter also reports that Astra wins six of the individual evaluations, Sol two, with two ties. A gap of under a point in aggregate can sit on top of a model that loses more often than it wins. And the figures arrive through a router company's blog, not from Artificial Analysis directly.
Free newsletter
Get the morning briefing
Start each day with the AI stories that matter and why — a short, free email from our newsroom.
Developer Simon Willison's informal test, asking models to draw a pelican, turned up nothing new. "They're not notably different from the GPT-6 family pelicans," he wrote. That is an anecdote, not a benchmark.
Cheaper is not ultrafast
Sol is a cheaper model, not an ultrafast one. Ultrafast is a speed mode for GPT-6 Astra: NVIDIA said on Thursday that GPT-6 Astra Ultrafast runs on its Blackwell GPUs and generates tokens up to 8x faster than Astra's standard mode. That is an NVIDIA and OpenAI claim with no independent check found, and NVIDIA's post gives no price.
Sol itself, by OrcaRouter's account of Artificial Analysis data, outputs 66.8 tokens per second. That beats Astra's 57.0 but trails the 76.2 of GPT-6 Sol, the model it replaces.
There are setup costs too. DataCamp, citing OpenAI's model page, reports that Sol offers no none or minimal reasoning effort, the setting that controls how much computation a model spends working through a problem before answering. Reasoning is always on. Tool calling requires OpenAI's Responses API, and moving from GPT-6 Sol requires code changes.
The launch also came a day after OpenAI held back GPT-6.1 Astra, after internal testing found it carried on with tasks without asking permission and used potentially unsafe external tools. No new date has been set. OpenAI rates Sol itself Critical for cybersecurity under its Preparedness framework, and its claim that Sol's alignment improved comes from its own evaluations.
Verdict
The price claim is supported. The near-Astra intelligence claim is supported as an average and overstated as a blanket promise: the strongest independent number is secondhand, and Astra still wins most of that index's individual evaluations, by OrcaRouter's account.
What would settle it: Artificial Analysis confirming its figures directly, a clear side-by-side of Astra's and Sol's scores on each benchmark OpenAI cites along with who maintains those tests, and independent evaluations on the long-running coding and document workloads OpenAI recommends Sol for. Until then, teams weighing the switch should run their own tasks on both models and compare the cost of a finished job, not the price of a token.