Two AI Labs Cut Prices on the Same Afternoon
On 22 September, OpenAI released GPT-6 Sol and GPT-6 Luna at half the API price of the models they replace, about 90 minutes after Anthropic released its own cheaper model, Claude Opus 5.5. Anthropic priced Opus 5.5 at $4 per million input tokens and $20 per million output, down from $5 and $25 for Opus 5, and says it costs about 40% less to run on typical workloads. OpenAI’s Sol came in at $2 and $10, with Luna an order of magnitude below at 10 cents and 50 cents. OpenAI attributed the cuts to better caching and more efficient inference.
Two of the largest AI companies cutting prices on the same afternoon is not a routine version update. It is a sign of where the market is heading: the cost of machine reasoning is falling fast, and the providers are now competing on it openly. For a manufacturer, that changes a few calculations in concrete ways.
For document work, the model is now a rounding error
Take a task most mid-size manufacturers still do by hand: reading incoming order confirmations, supplier quotes or delivery notes and getting the data into the ERP. Assume, as a rough illustration, a document of about 5,000 tokens in and 1,000 out.
At the cheapest of the new models, that works out to around a tenth of a cent per document. At the most capable, around four cents. Ten thousand documents a year cost somewhere between ten and four hundred dollars in model fees, depending on which tier you need.
Now compare that with the cost of a person spending three minutes on each of those documents. The model bill is not the issue anymore, at any tier. What costs money in these projects is the integration into your ERP, the handling of the documents that do not fit the pattern, and the process for someone to check the exceptions. If a document automation project was shelved last year because the running costs looked uncertain, that reason has largely disappeared. The remaining reasons are the ones worth working on.
If you chose open-weight models for price, check again
Several outlets pointed out that Luna’s pricing now undercuts DeepSeek’s V4.1 Flash, one of the most widely used low-cost open-weight models. That changes the calculation for any team that picked an open-weight model purely to save on inference.
The distinction matters. If you run open models on your own infrastructure because customer drawings, pricing or production data must not leave your control, nothing in this week’s news changes that decision. Control was the reason, and control is not what got cheaper. But if the choice was simply about cost, it deserves a fresh look, because a self-hosted setup has its own running costs in hardware, maintenance and people, and the commercial side just moved.
Do not lock in at today’s prices
The more durable lesson is about architecture rather than any single price sheet. Prices at the top of the market moved substantially in one afternoon, and they will move again. Any system you build now should let you change which model handles which task without rebuilding the workflow around it. Most operational document tasks, such as classification, extraction and routing, will run perfectly well on the cheapest tier, with the expensive models reserved for the steps that genuinely need them. Commitments priced per token over several years deserve particular scepticism in a market moving this quickly.
One caveat
These are launch-day numbers. At the time of writing, no independent head-to-head benchmark compared the three new models. The providers’ own benchmarks tell you what the models can do on their tests, not on your supplier’s badly scanned delivery note. Before you pick, run fifty of your real documents through two or three candidates and compare the error rate, not the brochure. That test now costs less than the coffee during the meeting where you discuss it.
“Content generated using AI”