OpenAI's GPT-6.1 Sol cuts computer-use cost to about a seventh. Here's which portal chores a sub can now price out
OpenAI says GPT-6.1 Sol comes within 2.1 points of its flagship on a computer-use benchmark at roughly a seventh of the cost per task. For contractors, the question is which repetitive portal and spreadsheet chores now pencil out.
OpenAI's new GPT-6.1 Sol model completes computer-use tasks within 2.1 points of its flagship Astra model at roughly one-seventh of the cost per task, according to OpenAI's own figures. For a trade sub, that moves repetitive portal and spreadsheet chores from "interesting demo" toward "worth a line in the overhead budget," as long as someone checks the output.
What did OpenAI actually launch?
At DevDay on September 29, OpenAI released GPT-6.1 Sol. Pricing is $2 per million input tokens and $10 per million output tokens. Astra charges $10 and $50. Cached input, which matters when an agent reuses the same instructions over and over, is $0.10 per million.
OpenAI's claims, as reported by TheNextWeb, MarkTechPost and others:
- Coding: matches Astra on DeepSWE v1.1 at about a fifth of the cost.
- Computer use: within 2.1 points of Astra on OSWorld 2.0 at about a seventh of the cost per task.
- Availability: ChatGPT Work and Codex for paid tiers, and the API as
gpt-6.1-sol.
These are vendor-reported benchmarks. None of them involve a construction workflow.
Why does cost per task matter for a sub?
Because most of a sub's office work is small, repetitive and slightly different every time. A model that is good at clicking through screens was already possible. Whether it was affordable at the volume a 40-person shop generates was the open question.
Here is where the arithmetic could start to work:
| Chore | Why it fits | What to check |
|---|---|---|
| Pulling certified payroll or insurance certificates into GC compliance portals | Same fields, same steps, weekly | Every entry matches the source document |
| Re-keying pay application line items between your system and a GC's | Bounded, checkable against totals | Retainage and stored materials lines |
| Checking order status across supplier sites | Read-only, low risk | Dates and quantities |
| Building bid-day comparison sheets from vendor quotes | Spreadsheet work, easy to audit | Exclusions and alternates |
Notice what is not on the list: anything that submits, signs or commits money without a person reading it first.
What are the limits?
Sol is not Astra. One cost-per-task comparison from Beri reported Sol's terminal-work task at $5.47 versus $23.80 for Astra, with a score about 11 points lower. Cheaper and near-top is still not top.
A benchmark also tests a clean sandbox. Real GC portals time out, change layouts, demand two-factor codes and sometimes forbid automated access in their terms. Check the portal's terms before pointing any agent at it, and give it its own login rather than a superintendent's or owner's credentials.
Should a mid-size sub act now?
Run one cheap test rather than a rollout. Pick the portal chore your office manager dislikes most, run it with Sol for two weeks alongside the person doing it, and log three numbers: minutes saved, errors caught, and the token bill. If errors are not near zero, the savings are fake, because someone is still re-checking everything.
Takeaway: the model price dropped, but the review step did not. Price out one chore, keep a human sign-off on anything that leaves the building, and decide from your own error log, not OpenAI's chart.
- What is GPT-6.1 Sol and how much does it cost?
- GPT-6.1 Sol is an OpenAI model launched at DevDay on September 29. API pricing is $2 per million input tokens and $10 per million output tokens, about one-fifth of GPT-6 Astra's standard rates, with cached input at $0.10 per million tokens.
- Can an AI model fill out contractor portals and web forms for me?
- Computer-use models can click through web pages and forms, and OpenAI says Sol is within 2.1 points of Astra on the OSWorld 2.0 benchmark. A benchmark is not your vendor's login page, so test on one low-risk portal and review the output before relying on it.
- Is GPT-6.1 Sol as good as OpenAI's top model?
- No. OpenAI describes it as near-Astra, and one cost-per-task comparison reported a terminal-work score about 11 points below Astra's. It is the cheaper option for repetitive, checkable work rather than the strongest model available.
- Where can contractors use GPT-6.1 Sol?
- It is available to Plus, Pro, Business, Enterprise and Edu users in ChatGPT Work and Codex, and to developers through the API as gpt-6.1-sol.