GPT-6.1 Sol Claims Near-Astra Work at a Fifth of Astra’s Token Price
OpenAI’s DevDay model is not the shelved Astra successor. GPT-6.1 Sol keeps the $2 / $10 sticker and claims near-GPT-6 Astra results on coding, computer use, and professional work at about one-fifth of Astra’s standard token prices. Those benches are the lab’s.
Corrected October 14, 2026: Removed unverifiable claims that Arena had not ranked the model and an unsourced line about Astra Ultrafast availability; fixed internal links.
The model OpenAI shipped at DevDay is the cheap tier, tightened — not the frontier successor it had just pulled. GPT-6.1 Sol is an upgrade to GPT-6 Sol, priced the same, and sold as near GPT-6 Astra on agentic coding, computer use, and professional work at one-fifth of Astra’s standard input and output token prices. The reported hold on GPT-6.1 Astra is what makes “near” do commercial work: buyers are offered Astra-class tasks without the SKU safety staff would not release.
API prices, from the OpenAI post (September 29): $2 per million input tokens, $0.10 cached input, $10 output. Cached input is, OpenAI says, 95% below this model’s own standard input price and 50% below GPT-6 Sol’s cached input price. Slug: gpt-6.1-sol. It is in ChatGPT Work and Codex for Plus, Pro, Business, Enterprise, and Edu. Not yet in Chat. Ultrafast, up to 8× token generation versus standard speed in Codex, is promised in the coming days, not on day one.
The sticker matches Claude Sonnet 5.5 and GPT-6 Sol. The claim that has to be true is cost per task against Astra, not the menu price.
Lab claims#
OpenAI says competitor numbers are taken from public reports, and that its own runs were in a research environment or via API, which can differ from production ChatGPT. Effort level is doing a lot of work in every row.
- DeepSWE 1.1: matches GPT-6 Astra at roughly one-fifth the cost, and beats GPT-6 Sol’s best score by 6.4 percentage points at a lower effort and cost. The prose post does not print the absolute percent.
- GDP.pdf (professional questions over complex PDFs): higher than Opus 5.5 with fallbacks at less than half the cost per task across tested settings, and near Astra at roughly one-fifth the cost per task.
- AutomationBench 1.0.6 (47 tools): 2.2 percentage points above Opus 5.5 at medium effort, at roughly a third of the cost, and 4.8 points above GPT-6 Sol at that same setting. OpenAI again flags that a published Fable 5.1 point omits fallback cost, which hit on about 40% of tasks.
- OSWorld 2.0 offline, partial reward, release v2026.08.08: seven points above GPT-6 Sol at maximum effort and less than half the cost; within 2.1 points of Astra at maximum effort at roughly one-seventh the cost per task.
- Terminal-Bench Science 0.1: more than doubles GPT-6 Sol at maximum effort, at less than half the cost. OpenAI’s average cost per task at max: $5.47 for 6.1 Sol, $23.21 for Opus 5.5, $23.80 for Astra. Astra still leads the score, at 68.1%. OpenAI says to use Astra for the hardest scientific tasks.
- Factuality, an internal set of de-identified chats where a user had already flagged an error — not typical traffic. At low effort, the share of answers with a factual error falls from 11.4% to 7.7%, about a 32% reduction. Across tested settings, the error rate stays within 1.9 percentage points of Astra at less than one-fifth the cost per task.
On alignment, OpenAI says 6.1 Sol is closer to Astra than GPT-6 Sol was, more willing to state limits, and showed no attempts to bypass an automated safety reviewer — the same result it reports for Astra and GPT-6 Sol. One printed stress number: when the search tool is broken, 6.1 Sol fails to say so in 2.1% of cases, versus 4.9% for GPT-6 Sol, 1.5% for Astra, and 28.7% for Luna, all at maximum effort, on prompts chosen to elicit failures.
Independent checks#
Neither the launch post nor TechCrunch’s write-up cites a third-party leaderboard result.
Artificial Analysis had, the day before, put Claude Sonnet 5.5 max at Intelligence Index 56 and Opus 5.5 max at 58, and described GPT-6 Sol configurations as the cheaper comparison at lower Sonnet efforts. A settled max-effort Intelligence Index for GPT-6.1 Sol was not in those writeups. The model was hours old. TechCrunch restates OpenAI’s claims, including the shelved-Astra context. It is not a third-party bench.
Until AA or Arena prints a dated row, “near Astra” means OpenAI’s task-level cost curves, not a shared index. The Sonnet 5.5 piece is the caution: identical $2 / $10 stickers, and AA already showed max-effort Sonnet leaving the cost frontier by writing ~193,000 tokens a task.
Limits#
- Not in consumer Chat. Work and Codex are the surfaces.
- Ultrafast for this model is announced, not available.
- Several headlines are “within N points” or “one-fifth the cost,” not a full score table. Absolute DeepSWE and AutomationBench percents are not in the prose we reviewed.
- Factuality and broken-tool tests are hard, non-representative sets. OpenAI says so.
- The Fable AutomationBench cost still omits fallbacks, by OpenAI’s own note.
- No third-party leaderboard result is cited in the sources used here.