GPT-6.1 Sol vs GPT-6 Astra: Near-Equal Scores, One-Fifth the Price

On September 29, OpenAI released GPT-6.1 Sol with a one-line pitch: near-Astra intelligence at a fifth of the price. For a marketing team or an agency that runs an AI language model all day on briefs, scripts, research and automated workflows, that line is a budget decision. The question is where “near” is close enough, and where the gap still costs you.

The answer depends on the task. Here is what the numbers say so far, who produced them, and a simple rule for choosing.

What OpenAI released

GPT-6.1 Sol upgrades GPT-6 Sol, the cheaper sibling of OpenAI’s flagship GPT-6 Astra. The independent benchmarking firm Artificial Analysis noted that it replaced GPT-6 Sol after just seven days. OpenAI presents it as strongest at agentic coding, computer use and professional work.

In the API, as gpt-6.1-sol, it costs $2 per million input tokens and $10 per million output tokens. GPT-6 Astra costs $10 and $50. Cached input, the discounted rate for text the model has already processed in a repeated prompt, is $0.10 for Sol and $1 for Astra. In ChatGPT, TechCrunch and VentureBeat report that Sol is available on the Plus, Pro, Business, Enterprise and Edu plans, inside ChatGPT Work and Codex, but not yet in regular Chat.

Otherwise, the two models look alike on paper. OpenAI’s model pages list a 1,050,000-token context window, 128,000 maximum output tokens, an April 30, 2026 knowledge cutoff and the same five reasoning-effort settings for both. Both charge a premium on any prompt longer than 272,000 tokens.

The same event added a speed tier. OpenAI says Ultrafast generates up to 300 tokens per second, up to eight times standard speed in Codex and up to six times in the API. It is live for Astra only, in the API and on ChatGPT’s Pro 500 and Enterprise plans. OpenAI says a Sol version arrives in the coming days.

Where Sol keeps up, and where Astra still leads

OpenAI’s own results put Sol level with Astra on DeepSWE v1.1, a software engineering test, and 2.1 points behind on OSWorld 2.0, which measures operating a computer. On factual questions, its error rate stays within 1.9 points of Astra across reasoning settings. The Decoder flags the catch: “All benchmarks come from OpenAI, which describes them as preliminary.”

Artificial Analysis provides the first outside check. At maximum reasoning effort, its Intelligence Index scores Sol at 52 and Astra at 53. Astra keeps a two- to three-point lead on tests of terminal coding, business automation, scientific coding and very hard questions. Sol ties on GDP.pdf, which asks questions about long professional PDFs, and edges ahead on long-context reasoning, 83% to 81%.

One point apart on the index, five times apart on cost

Independent scores at maximum reasoning effort, and the cost of running the full Intelligence Index

GPT-6.1 Sol and GPT-6 Astra on Artificial Analysis evaluations, maximum reasoning effort

Evaluation	GPT-6.1 Sol	GPT-6 Astra
Artificial Analysis Intelligence Index	52	53
Terminal-Bench 4.0	56%	59%
AutomationBench-AA	65%	68%
Humanity’s Last Exam	53%	55%
GDP.pdf	31%	31%
AA-LCR (long-context reasoning)	83%	81%
Cost to run the full Intelligence Index	US$1,082	US$5,324

Two areas still clearly favour Astra. On Terminal-Bench Science, OpenAI’s own scientific research test, Astra posts the top score at 68.1%, and OpenAI says it “should be used for the most difficult scientific research tasks.” The second is behaviour inside agents. In OpenAI’s safety addendum, Sol kept pushing after a warning in 23.5% of test runs, against 17.4% for Astra, and misrepresented its coding work in 1.50% of cases, against 0.51%. Those tests are built to be hard, but they matter once a model acts on your accounts.

The cost and speed math

List prices are five times apart, and measured costs roughly follow. Running the full Artificial Analysis index cost $1,082 with Sol and $5,324 with Astra. OpenAI reports Sol within reach of Astra on OSWorld at about a seventh of the cost per task, and $5.47 per Terminal-Bench Science task against $23.80 for Astra, though at a lower score.

Take a typical agency job: a 40,000-token prompt holding brand guidelines, research notes and a brief, and a 4,000-token answer. At list price, that is $0.12 on Sol and $0.60 on Astra. Run it 2,000 times a month for variations, language versions or product pages, and the bill is $240 against $1,200. Reasoning tokens are billed as output, so heavy-thinking runs cost more on both. Caching widens the gap: if 30,000 tokens of guidelines open every prompt and are cached, that part costs $0.003 per run on Sol and $0.03 on Astra.

Speed is a separate budget. Artificial Analysis measured Sol at 66 output tokens per second at standard speed and Astra at 51, so 1,000 tokens take about 15 and 20 seconds. Ultrafast brings that to roughly three seconds, at six times the standard price: $60 input and $300 output per million tokens for Astra. OpenAI’s pricing page and VentureBeat list different output prices for Sol Ultrafast, so wait for the live rate before budgeting it.

What it changes for marketing and creative teams

Scripts, briefs and copy variations are high-volume work where a one-point benchmark gap rarely shows in the result, and a person edits the draft anyway. Sol is the sensible default there, and the savings can pay for more drafts and more review time.

Research on long documents, such as decks, reports and tender packages, also suits Sol: it tied Astra on the PDF test and led on long-context reasoning in independent testing. Keep fact-checking. A 1.9-point gap in factual errors is small, and it applies to every claim.

Agents and automation are where the choice gets harder. Sol scored close to Astra on business workflows and computer use at a fraction of the cost per task, but the safety results show it more likely to push past a warning. For an agent that only drafts, Sol is fine. For one that sends emails, edits a website or spends an ad budget, Astra’s extra caution may be worth paying for, alongside human approval at each step.

A practical rule: start every task on Sol. Move a task to Astra when a mistake costs more than the model does, such as a strategy memo, a complex data analysis or an agent with real permissions. Pay for Ultrafast only where someone is waiting on the answer, like a live working session with a client.

Before you switch

Run twenty of your own real tasks through both models and have the person who usually reviews that work grade the outputs blind. Compare cost per finished task, retries included, rather than price per token. Check which reasoning setting your tools use, since cost and quality both move with it. Confirm that your team’s ChatGPT plan actually offers Sol where you work. And date the decision: Sol replaced its predecessor within a week, so plan to rerun the test at the next release.