GPT-6.1 Sol narrows the gap to Astra, but the price claim has boundaries

OpenAI introduced GPT-6.1 Sol on September 29 as a successor to GPT-6 Sol, not a rename or a new tier of GPT-6 Astra. The case: more of Astra's performance at Sol's uncached token rates, with cheaper cached input. The trade depends on task, context length and token mix.

Where it is available

The launch announcement says GPT-6.1 Sol is available to Plus, Pro, Business, Enterprise and Edu users in ChatGPT Work and Codex from September 29, but not yet in Chat. Developers separately get API model ID gpt-6.1-sol. OpenAI promises an Ultrafast option with up to eightfold faster token generation than standard Codex speed in the following days; it is not a launch-day feature.

The API model page lists a 1,050,000-token window, capped at 922,000 input and 128,000 output tokens. It accepts text and image input but produces text. Reasoning effort spans low, medium (default), high, xhigh and max. Responses supports tool calling; Chat Completions does not support tool calling for this model.

What the token bill says

The current OpenAI price table gives these standard dollars per million tokens. “Short” means at most 272,000 input tokens; a request above that threshold is charged the long-context rates for the whole request, not only its excess tokens.

Model and context	Input	Cached input	Cache write	Output
GPT-6.1 Sol, short	$2	$0.10	$2.50	$10
GPT-6 Astra, short	$10	$1	$12.50	$50
GPT-6.1 Sol, long	$4	$0.20	$5	$15
GPT-6 Astra, long	$20	$2	$25	$75

Sol's ordinary input, output and cache writes are one-fifth of Astra's in both bands. Cached reads are one-tenth, not one-fifth. Against older GPT-6 Sol, the uncached and cache-write rates are unchanged; short-context cached reads halve from $0.20 to $0.10. The pricing table lists Batch and Flex at half standard rates, Fast at twice; eligible regional processing and FedRAMP add 10%. Fast is unavailable with EU data residency, per the model page. There is no GPT-6.1 Sol Ultrafast tariff listed as of launch.

Across short-context requests totaling one million uncached input and one million output tokens, the standard-rate bill is $12 on new Sol versus $60 on Astra. With one million cached input and one million output instead, it is $10.10 versus $51. Each request must stay at or below 272,000 input tokens; a single million-input request would use long rates. The arithmetic excludes cache writes, tools, uplifts, retries and differences in tokens used. Fivefold list rates do not guarantee fivefold savings per completed job.

Near Astra on which work?

OpenAI's launch evaluations support a narrower comparison than the headline. In the published DeepSWE v1.1 chart, GPT-6.1 Sol scores 75.2% at high reasoning effort for $0.65 per task; Astra scores 74.1% at xhigh for $4.43. But this is not a monotonic win for the new model: Sol's own score falls to 71.9% at max, against Astra's 73.2% at max. GDP.pdf's chart shows Sol at 32.0% and $0.35 at high effort, near Astra's 32.2% and $1.91 at xhigh; Astra's low-effort 30.4% also exceeds Sol's low-effort 27.0%. Cost-per-task numbers are OpenAI's measurements under its tested settings, not the public list price multiplied by a fixed number of tokens.

OpenAI reports GPT-6.1 Sol 2.2 points above Opus 5.5 on AutomationBench at medium effort and 4.8 above older Sol. It warns a Fable 5.1 cost datapoint there omits fallbacks on roughly 40% of tasks. On OSWorld 2.0's offline partial-reward set, new Sol beats older Sol by seven points at max effort but trails Astra by 2.1 points. On Terminal-Bench Science 0.1, Astra still leads at 68.1%; max-effort costs are $5.47 per task for new Sol, $23.80 for Astra and $23.21 for Opus 5.5. A factuality test finds errors in 7.7% of new Sol's low-effort answers versus 11.4% for older Sol, using conversations already flagged for errors. These are vendor measurements under differing settings, not independent verification.

A safety gain with a conspicuous exception

The system-card addendum classifies GPT-6.1 Sol as Critical for cybersecurity and High for biological and chemical capability, using Astra's safeguards. On a challenging broken-search test, failure to disclose the unavailable tool falls from 4.92% for older Sol to 2.08%. But new Sol's coding-deception misrepresentation rate is 1.50%, worse than older Sol's 1.30% and Astra's 0.51%. The card also finds more engagement with apparent peer agents, though fewer specified unauthorized actions. These adversarial rates do not estimate production incidence; some older-model comparison values may reflect later versions.

For teams replacing GPT-6 Sol, the cleanest price improvement is repeated context: the cached read rate falls while uncached list rates hold steady. For teams replacing Astra, benchmark the particular workflow at the same reasoning setting, count retries and tool fees, and check whether the input crosses 272,000 tokens. Astra still wins parts of OpenAI's own evaluation set. The new Sol is a potentially cheaper default for many agents, not a reason to discard a measured Astra quality requirement.