OpenAI’s “one-fifth of Astra” headline is true — and was already true in July. The only thing that actually moved is the cached-input price, and OpenAI tells you exactly why.
Two sources, kept apart. Every quotation and every evaluation result below is from OpenAI’s own product page for GPT-6.1 Sol, read directly.The price history of the earlier Sol models — and the finding that input and output have not changed — comes from the OpenRouter public models API, queried directly. OpenAI’s page does not state it, and the 272,000-token second price band is the register’s figure rather than OpenAI’s.Every evaluation result is OpenAI’s own, measured at settings OpenAI chose, and this publication has reproduced none of them. Nothing here is investment advice.
On September 29, 2026, OpenAI published a product page for GPT-6.1 Sol, describing it as “an upgrade to GPT-6 Sol that nearly matches GPT-6 Astra’s intelligence on agentic coding, computer use, and professional work at one-fifth of Astra’s standard input and output token prices.” That claim is accurate. Standard API prices are $2 per million input tokens, $0.10 per million cached input tokens, and $10 per million output tokens. What OpenAI’s page does not say — and what the OpenRouter public models API does — is that GPT-6 Sol, listed on September 22, carried identical input and output prices of $2.00 and $10.00, as did GPT-5.6 Sol, listed on July 9. Across three Sol generations, input and output prices are the same. Every eval cited in this piece is OpenAI’s own, at settings OpenAI chose; this publication has reproduced none of them.
The one price that did move is the cached read. It fell from $0.20 to $0.10 per million tokens — a 50 percent reduction from GPT-6 Sol’s cached rate and 95 percent below standard input pricing. OpenAI’s page is explicit about the target: “giving developers more room to build and run capable agents that reuse context across requests.” That is OpenAI’s own stated rationale, not an inference. A cached read is the charge incurred when a long prompt prefix is sent repeatedly — the cost shape of an agent looping over a fixed context rather than a chatbot answering isolated questions.
Two further details the OpenRouter register shows that OpenAI’s announcement does not. First, a second price band: the register lists a 1,050,000-token context window for all of these models and a separate rate above a 272,000-token prompt, where GPT-6.1 Sol’s input doubles to $4.00 and its output rises to $15.00; GPT-6 Astra carries the same structure at $20.00 and $75.00. OpenAI’s pricing paragraph does not mention that second band — that finding belongs to the register, not to OpenAI. Second, the Pro premium appears gone: the previous generation listed a gpt-5.6-sol-pro variant at double its sibling’s price, while both gpt-6-sol-pro and gpt-6.1-sol-pro are listed at the same price as their base counterparts.
The key insight: “A fifth of Astra” describes the Sol tier’s longstanding position in OpenAI’s pricing architecture, not a price cut won by this release. The only variable that moved is cached-input cost — and OpenAI’s own words make the reason plain: agents that reuse context across requests need a different cost curve than chat does.
Indexed rather than absolute, because the levels differ by a factor of a hundred and would hide the only thing that moved. Input has been $2.00 per million tokens and output $10.00 since July. The cached read halved to $0.10 this week.
The Structural Read
OpenAI’s evals contain the most honest signals on this page, and they reward careful reading. On Terminal-Bench Science 0.1, OpenAI reports that GPT-6.1 Sol costs $5.47 per task on average at maximum effort, compared with $23.21 for Opus 5.5 and $23.80 for GPT-6 Astra — delivering what OpenAI calls “substantial scientific capability at over 75% lower cost than either model.” Then, in the very next sentence, OpenAI writes: “GPT-6 Astra still achieves the highest score among the models tested at 68.1%, and should be used for the most difficult scientific research tasks.” The cheaper model is cheaper; the more expensive one is still recommended for hard problems. OpenAI wrote both sentences.
The honesty benchmark reinforces that tiering. In a test asking whether an agent discloses a broken search tool rather than guessing, GPT-6.1 Sol fails to disclose in 2.1 percent of cases; GPT-6 Sol fails in 4.9 percent; GPT-6 Astra in 1.5 percent; and GPT-6 Luna — the cheapest model in the family — fails in 28.7 percent of cases, more than thirteen times as often as 6.1 Sol. Astra is more reliable than 6.1 Sol on that measure. OpenAI’s own caveat must travel with those numbers: “Tasks are selected to elicit failures and do not represent typical usage. Effort was set to maximum.”
On factuality, at low reasoning effort, GPT-6.1 Sol reduces the share of responses containing a factual error from 11.4 percent to 7.7 percent — approximately a 32 percent reduction — and its error rate stays within 1.9 percentage points of GPT-6 Astra’s at less than one-fifth the cost per task. OpenAI states plainly: “These deliberately difficult prompts are not representative of typical usage.” The test was built from de-identified ChatGPT conversations where users had already flagged a factual error from a prior model. Elsewhere the page reports a 6.4-percentage-point gain over GPT-6 Sol on a real-codebase software-engineering eval; a 2.2-percentage-point lead over Opus 5.5 at medium effort on a 47-tool business-workflow eval (and 4.8 points above GPT-6 Sol at the same setting); and a seven-percentage-point gain over GPT-6 Sol on a computer-use offline set at maximum effort, while coming within 2.1 points of Astra. On a competitor datapoint, OpenAI adds — quoted here rather than endorsed — that “the datapoint for Claude Fable 5.1 understates its actual cost, as it omits the cost of fallbacks, which occurred on approximately 40% of tasks.”
The launch surface choice is as legible as the price change. GPT-6.1 Sol is available today to Plus, Pro, Business, Enterprise, and Edu users in ChatGPT Work and Codex. GPT-6.1 Sol is not yet available in Chat. A model pitched at agentic and professional work landed in the work surfaces first — not in the consumer interface. Ultrafast token generation, described as up to eight times faster in Codex, is coming in the following days rather than shipped at launch.
Three Implications
Agent economics, not chat economics
Halving the cached-read price targets a specific cost structure: long, repeated prompt prefixes sent by agents looping over fixed context. Developers building that class of system get a 50 percent reduction on their dominant cost line. Developers running standard chat workloads see no change. The pricing change is narrower than the announcement reads, and more precise.
Tiering is explicit and self-described
OpenAI’s own evals position Sol as a cost-performance sweet spot and Astra as the correct choice for the hardest scientific and reliability-sensitive tasks. That is not analyst inference — it is OpenAI’s text. Enterprises evaluating routing decisions now have an OpenAI-authored decision tree: Sol for volume and cost efficiency, Astra where failure costs are high.
The undisclosed second band matters for large-context workloads
The register, not OpenAI’s announcement, surfaces a second price band above a 272,000-token prompt: input rises to $4.00 and output to $15.00 for GPT-6.1 Sol. Whether that matters depends entirely on prompt length, and a budget projected from the headline prices alone understates the cost of any workload that runs past 272,000 tokens.
Business Engineer Framework
Map of AI — Where Sol Sits in the Stack
The Map of AI tracks more than 200 companies across nine layers of the AI stack, from infrastructure to applications. GPT-6.1 Sol’s cached-read cut is a layer-level pricing signal: it moves the cost curve for inference workloads that are structurally different from chat. Understanding where a model sits in that nine-layer map — and which cost shapes each layer rewards — is the analytic frame that turns a pricing footnote into a deployment decision.
91,000+ executives read Business Engineer for the AI strategy frameworks cited by ChatGPT, Claude, and Perplexity.
Two sources are used above and kept apart. Every quotation and every evaluation result is from OpenAI’s own product page for GPT-6.1 Sol, read directly by this publication.The price history of the earlier Sol models comes from the OpenRouter public models API, queried directly on 29 September 2026, which returned 464 models. OpenAI’s page does not state what GPT-6 Sol or GPT-5.6 Sol cost, and the second price band above a 272,000-token prompt is the register’s figure rather than OpenAI’s. The observation that input and output are unchanged across three Sol generations therefore rests on the register, not on any OpenAI statement.Every evaluation figure is OpenAI’s own, measured in OpenAI’s environment at reasoning settings OpenAI selected, and this publication has reproduced none of them. OpenAI’s own caveats are quoted alongside its numbers rather than paraphrased: the broken-search-tool tasks are selected to elicit failures and do not represent typical usage with effort set to maximum, and the factuality prompts are deliberately difficult and not representative of typical usage. OpenAI’s note that a competitor datapoint understates its cost is quoted above rather than endorsed, and this publication has not examined that datapoint. GPT-6.1 Sol Ultrafast is described by OpenAI as coming in the following days and is not reported here as shipped. Nothing above predicts anything about pricing, adoption or any future model, and nothing here is investment advice.
Gennaro is the creator of FourWeekMBA, which reached about four million business people, comprising C-level executives, investors, analysts, product managers, and aspiring digital entrepreneurs in 2022 alone | He is also Director of Sales for a high-tech scaleup in the AI Industry | In 2012, Gennaro earned an International MBA with emphasis on Corporate Finance and Business Strategy.
Scroll to Top
Discover more from FourWeekMBA
Subscribe now to keep reading and get access to the full archive.
