GPT-6.1 Sol is OpenAI's model announced at DevDay on 29 September 2026. Per OpenAI it nearly matches GPT-6 Astra on agentic coding, computer use and professional work at one-fifth of Astra's standard prices, listed at $2 per million input tokens and $10 per million output tokens. For content teams the answer is: test it on your own drafts before switching, because the benchmarks are OpenAI's, they used deliberately hard prompts, and writing quality is not what they measured.
Key Takeaways
- API price: $2 input, $10 output and $0.10 cached input per million tokens, per OpenAI's model documentation.
- Specs listed by OpenAI: 1,050,000-token context window, 128,000 max output, April 30, 2026 knowledge cutoff, text and image input.
- A secondary source reports long-context rates roughly double for input above 272,000 tokens; verify on OpenAI's page.
- The cached rate halved versus GPT-6 Sol ($0.20 to $0.10), which matters for repeated style guides and briefs.
- OpenAI's headline gains are in agentic coding, computer use and professional work, not creative writing.
- Switch only after a blind test on your own content tasks.
What actually changed
OpenAI's launch post describes GPT-6.1 Sol as an upgrade to GPT-6 Sol. I read the post and the model page on 30 September 2026. These are the claims, all vendor-reported.
- Near-Astra intelligence at lower cost. OpenAI says it nearly matches Astra on agentic coding, computer use and professional work at one-fifth of Astra's standard prices.
- Coding. Matches Astra on DeepSWE v1.1 at one-fifth of the cost, per OpenAI.
- Computer use. Beats GPT-6 Sol by 7 percentage points on OSWorld 2.0, per OpenAI.
- Science. More than doubles GPT-6 Sol on Terminal-Bench Science 0.1.
- Factuality. About 32% fewer errors at low reasoning effort, per OpenAI.
- Cache pricing. Cached input at $0.10 per million tokens, half of GPT-6 Sol's $0.20.
The caveat from the same post: evaluations used "deliberately difficult prompts" and "challenging situations" not representative of typical usage. I could not find an independent benchmark page for it (the Artificial Analysis release page I tried returned 404), so nothing here is independently confirmed.
The specs on the model page
According to OpenAI's model documentation:
| Spec | Value |
|---|---|
| Model ID | gpt-6.1-sol |
| Context window | 1,050,000 tokens |
| Max output | 128,000 tokens |
| Knowledge cutoff | April 30, 2026 |
| Input / output | Text and image in, text out |
| Reasoning effort | low, medium (default), high, xhigh, max |
| Price per 1M tokens | $2.00 input, $10.00 output, $0.10 cached input |
| Endpoints | Chat Completions, Responses, Realtime, Batch |
| Data residency | US and EU options (Fast Mode not available with EU residency) |
It is available in ChatGPT Work and Codex for Plus, Pro, Business, Enterprise and Edu users, with an Ultrafast variant listed as coming.
Cost math for a content team
Prices are per million tokens. I am assuming Astra at $10 input and $50 output per million tokens, which is what a one-fifth ratio implies; confirm on OpenAI's pricing page.
Assume, as an illustration, that one 2,000-word draft costs about 4,000 input tokens (brief, style guide, examples) and about 3,000 output tokens. These token counts are my estimate, since real usage depends on prompt length.
- GPT-6.1 Sol: input 4,000 x $2 / 1M = $0.008; output 3,000 x $10 / 1M = $0.03. About $0.038 per draft.
- Astra at $10/$50: $0.04 + $0.15 = $0.19 per draft.
At 200 drafts a month that is about $7.60 versus $38. Neither number is large. For most content teams the model bill is not the bottleneck; editing time is. Cost matters more for high-volume jobs, such as classifying thousands of pages, generating metadata at scale or running agents that loop.
Where caching helps
If every request starts with a 3,000-token style guide, cached input at $0.10 per million means that repeated portion costs a fraction of a cent. That is where the halved cache rate shows up. Whether your setup gets cache hits depends on how you structure prompts, so test it.
Long-context pricing
A secondary source reports that above 272,000 input tokens, rates double for input and cache and rise 1.5x for output. If that is right, stuffing a whole site into the prompt costs more than the headline suggests. Check the official page before building anything around long contexts.
A fair test before you switch
Do not trust benchmarks for editorial work. Run this instead.
- Pick five tasks you do repeatedly: a blog outline, a meta description batch, a rewrite of a weak section, a summary of a long source, and a fact-check pass.
- Run each on Sol and your current model with identical prompts, 10 samples per task.
- Blind-score outputs: an editor rates them without knowing which model wrote them, on accuracy, voice, and edit time needed.
- Record cost and latency for each.
- Decide per task, not overall. It is common for a cheaper model to win on structured tasks and lose on voice.
I would also watch factuality carefully. OpenAI says errors fell about 32% at low reasoning effort, but that still leaves errors, so keep human verification on any claim that goes public. Google's spam policies are aimed at low-value scaled content, so cheaper drafting raises, not lowers, the need for original insight in each piece.
Where Sol would fit in a content workflow
If I were slotting it in, I would use it for the structured, repeatable steps first: turning a brief into an outline, drafting metadata variants, tagging or classifying existing pages, and summarising sources for a writer. These are tasks where output is easy to check and mistakes are cheap. I would keep final voice, original data and claims about competitors or regulations under human control, whatever model is in the loop.
Who should switch, and who should wait
| Team | Suggestion |
|---|---|
| Agencies running agents or bulk classification | Test now; cost and cache pricing matter most |
| Solo writers using ChatGPT interactively | Try it in ChatGPT if your plan includes it; no API work needed |
| Teams on a fixed vendor contract | Run the blind test, then renegotiate or adjust |
| Regulated or client-confidential work | Read OpenAI's data residency and retention terms first |
| Teams happy with current quality | Waiting costs nothing; new models arrive monthly |
I am cautious about switching for its own sake. When I worked on social content growth for Masai School, the constraint was never generation speed; it was knowing what to say that audiences would actually engage with. Tools change quickly; that constraint does not.
What I could not verify
- Independent benchmark results; the ones above are OpenAI's.
- Whether Sol's writing quality beats other current models for long-form content.
- Ultrafast Sol pricing and availability, listed as coming.
- The long-context pricing detail, which I only found in a secondary source.
- Exact naming and relationship with earlier GPT-6 models; I have relied on OpenAI's pages and did not verify anything beyond them. Google's Project Astra is a separate, unrelated research prototype.
For the wider event context, see my DevDay roundup for marketers, and for a decision method across vendors, my model-choice framework.
FAQ
What is GPT-6.1 Sol?
It is an OpenAI model announced at DevDay on 29 September 2026 as an upgrade to GPT-6 Sol. OpenAI says it nearly matches GPT-6 Astra's intelligence on agentic coding, computer use and professional work at lower cost.
How much does GPT-6.1 Sol cost?
The model page lists $2.00 per million input tokens, $10.00 per million output tokens and $0.10 per million cached input tokens. A secondary source reports higher rates above 272,000 input tokens, so check OpenAI's page for long-context jobs.
What is its context window?
OpenAI's model page lists a 1,050,000-token context window and a 128,000-token maximum output. The knowledge cutoff is April 30, 2026.
Is Sol better than Astra?
Not according to OpenAI: it says Sol nearly matches Astra, at one-fifth of the price. That implies Astra remains ahead on some tasks. I have no independent results to add.
Is it good for writing blog content?
OpenAI's published gains focus on coding, computer use and professional tasks, not creative or long-form writing. Test it on your own drafts with a blind edit comparison.
Who can use it?
It is available in the API as gpt-6.1-sol and in ChatGPT Work and Codex for Plus, Pro, Business, Enterprise and Edu users, with rollout details on OpenAI's pages.
Does it have an Ultrafast version?
OpenAI listed GPT-6.1 Sol Ultrafast as coming soon. GPT-6 Astra Ultrafast was available on Pro 500 and Enterprise plans at announcement.
Is this related to Google's Project Astra?
No. Google's Project Astra is a separate research prototype. Only the name overlaps.
Should a small content team switch today?
Only after a test. The bill is small at typical volumes, so the deciding factor is edit time and accuracy, not price.
Work with me
I am Younus Fardeen, an organic growth, SEO and content strategist with 4+ years of marketing experience across edtech and startups. If you want a practical review of which models belong in your content workflow, see my work and contact me through the form at https://younusfardeen.in.
Sources: OpenAI, "Introducing GPT-6.1 Sol" and the GPT-6.1 Sol model page; Mixed, "GPT-6.1 Sol pricing" (secondary). Verified as of 30 September 2026.