On 29 September 2026 OpenAI released GPT-6.1 Sol and described it as near-Astra intelligence for coding, computer use and professional work at one-fifth of GPT-6 Astra's standard API token price. That ratio is the line most people will repeat. For an Australian business weighing it up against Claude, it is also the least useful number in the announcement.
The number that deserves more attention sits further down: OpenAI reports that GPT-6.1 Sol's factuality error rate fell from 11.4% to 7.7% at low reasoning effort, a relative drop of roughly 32%. That figure, and the review work it implies, is where most of the real cost of a model lives. This post explains why, with a worked example, and how to compare Claude and GPT-6.1 Sol on the measure that shows up in your profit and loss.
Is GPT-6.1 Sol cheaper than Claude for business work?
Per token, OpenAI prices GPT-6.1 Sol at US$2 per million input tokens and US$10 per million output tokens, which looks cheap. Per completed, correct task, nobody can answer that from a price sheet. The cost that matters is tokens plus retries plus the staff time spent checking and fixing output. For most business workloads the checking costs far more than the tokens, so the model that makes fewer errors on your tasks is usually the cheaper one, whatever its list price.
All of the figures in OpenAI's announcement are self-reported and, as at November 2026, we have not seen them independently reproduced. Here is what OpenAI claims, with the source kept visible:
Price: US$2 per million input tokens, US$0.10 per million cached input tokens, and US$10 per million output tokens, pitched as one-fifth of GPT-6 Astra's standard rate.
Coding: on DeepSWE v1.1 it matches Astra and beats the best GPT-6 Sol score by 6.4 percentage points.
Computer use: on OSWorld 2.0 it beats GPT-6 Sol by 7 percentage points at maximum reasoning effort.
Science tooling: on Terminal-Bench Science it more than doubles GPT-6 Sol's score at maximum effort.
Factuality: the error rate drops from 11.4% to 7.7% at low reasoning effort.
You can read the detail in OpenAI's GPT-6.1 Sol announcement. We looked at how to read the original Sol launch table, including OpenAI's own footnote, in Claude vs GPT-6 Sol: Reading OpenAI's Own Benchmarks. This post takes a different question: assuming the numbers hold, what do they buy you?
Why the error rate outweighs the token price
A 7.7% error rate is an improvement, and OpenAI deserves credit for publishing it. But read it the other way round. On the eval OpenAI used, about one answer in thirteen still contains a factual error at low effort. If the output goes to a client, a regulator or a ledger, someone has to find that one answer, which in practice means a person reads all thirteen.
That is the trap with a headline price. Cutting the token bill by 80% saves little if token spend was already a small share of the cost of the task. Cutting the error rate by a third can save a lot, because it changes how much review you need and how often work has to be redone.
A worked example in AUD
Take a mid-market firm in Sydney drafting 2,000 supplier contract summaries a month. Each one uses about 20,000 input tokens and 2,000 output tokens. Assume a reviewer costs $90 an hour fully loaded and each factual error takes 15 minutes to find and fix. These are illustrative assumptions, converted at an assumed A$1.55 to the US dollar, not a quote.
| Cost line | 11.4% error rate | 7.7% error rate |
|---|---|---|
| Token cost at GPT-6.1 Sol list price | $186 | $186 |
| Errors to find and fix | 228 | 154 |
| Fix time at 15 min and $90/hour | $5,130 | $3,465 |
| Total per month | $5,316 | $3,651 |
| Cost per correct summary | $3.00 | $1.98 |
The token line is around 4% of the total in the worse case and about 5% in the better one. The error rate moves the result by more than $1,600 a month; halving or doubling the token price would move it by less than $200. The same maths applies when you put Claude into the table: what decides the comparison is how often each model is wrong on your contracts, not whose per-token price is lower.
We made a similar argument about a different vendor in Gemini's low token price vs Claude, and the general method is in cost per outcome as an AI metric.
What to measure before you switch
A fair Claude versus GPT-6.1 Sol trial does not need to be elaborate. It needs to be run on your own work, with the measures fixed before anyone looks at the results:
Pick 50 to 100 real tasks from one workflow, such as contract summaries, support replies or code changes, with known correct answers.
Run each model at the reasoning effort you would actually pay for in production, not the maximum setting used in a launch chart.
Count tokens, retries and wall-clock time per task from the API logs rather than estimating them.
Have a reviewer score every output as correct, fixable or wrong, and time the fixes.
Divide the total cost, including review time, by the number of correct outputs.
Claude tends to do well on this measure in the builds we run because it is good at checking its own work: Claude Code, for example, runs tests and re-reads its changes before reporting a task done. That is a claim to test, not to take on faith, and the trial above is how you test it. If GPT-6.1 Sol wins on your tasks, the numbers will say so.
What not to read into the launch
Two cautions. First, the factuality figure was measured at low reasoning effort on OpenAI's own evaluation set. Your error rate on Australian contracts, Privacy Act questions or your own codebase could be higher or lower. Second, the benchmark gains are stated against GPT-6 Sol, OpenAI's previous model, not against Claude. A model that is much better than its predecessor tells you OpenAI is improving quickly, not where it sits against the alternative you are actually considering.
None of this makes GPT-6.1 Sol a bad option. It may well be the right choice for high-volume, low-stakes work where an occasional error is cheap. The point is to price it, and Claude, by the correct task rather than the token. Our ROI calculator lets you plug in your own volumes and review costs, and if you would like a side-by-side benchmark on your real tasks before committing budget, get in touch.



