On the same day, two of the largest providers of language models presented new versions – and both argue in a strikingly similar way: better results, fewer errors, lower prices. OpenAI has announced GPT-6 in the Sol and Luna variants, Anthropic the new flagship model Claude Opus 5.5.

For companies that build AI functions into websites, customer portals or internal tools, this is less a product announcement than a business consideration. Because the two factors that have most often held such projects back so far are precisely the ones that are now changing: the error rate and the cost per request.

What the two providers have announced

OpenAI promises for GPT-6 Sol and Luna half as many errors as with the previous versions – and at halved prices for the API, i.e. the programming interface through which applications call the model and which is usually billed according to the amount of text processed.

Anthropic positions Claude Opus 5.5 at the level of Claude Fable 5.1, but it is supposed to be 40 percent cheaper. In addition, the provider emphasises improvements in safety, without any robust details being available on this.

Both announcements are based on manufacturer information. How exactly the error rate was measured, with which task types and in which languages, remains open. Such figures should therefore be read as a direction, not as a guaranteed property for one's own use case.

Two different promises

On closer inspection, the messages differ. OpenAI is selling a double leap: simultaneously more accurate and cheaper. Anthropic is selling above all price – the same performance as before, significantly lower costs – flanked by statements on safety.

These are two legitimate strategies that meet different needs. Anyone who has so far worked at the quality limit and always had expenses checked is interested in the error rate. Anyone who has a working application in operation and looks at the monthly bill is interested in the 40 percent.

The timing is remarkable. The fact that two providers announce price reductions of a similar magnitude on the same day points to competition in which the price per request has become the main argument. For customers, this is good news in the short term – in the long term, it mainly means that prices remain volatile.

What changes in terms of economic viability

In many projects, AI functions fail not because of feasibility but because of the extrapolation. An assistant in a customer portal that triggers several requests per session becomes a cost block of its own with tens of thousands of sessions per month. If the price per request is halved, the limits shift significantly:

  • High-volume functions become calculable – for example, automatic summaries for every data record instead of only at the push of a button.
  • Multi-stage processes become more attractive: if a result is checked by a second pass, quality assurance no longer costs twice the previous price.
  • Powerful models across the board instead of only in a few places: where a smaller model was previously used for cost reasons, the larger one can take over.

The second point is directly related to the error rate. Part of the effort in AI projects is not caused by the model but by the safeguards around it: checking steps, approvals by employees, corrections afterwards. If the error rate falls measurably, this effort also falls – often the more expensive item, because it involves working time.

What makes sense now

An announcement does not lead to a project decision. A sober approach makes sense:

  1. Create your own measurement basis. Build a collection of real requests from your use case, with expected results. Only with this can you check whether a new model actually works better for you.
  2. Calculate costs per process instead of per request. What matters is what a completed business transaction costs – including repetitions, checking steps and manual rework.
  3. Ensure replaceability. If the model call lies behind your own interface, a change is a configuration question and not a rebuild. With price movements on a half-yearly cycle, this is the most important architectural lever.
  4. Budget conservatively. Price reductions are welcome, but not a reliable planning basis. Calculate with today's terms and take improvements as a buffer.

Limits of the reports

So far, only brief information is available on both models. There is a lack of information on availability in different regions, on context lengths, on response times and on how the models behave with subject-specific content and German-language texts. Especially response time is often more important for functions in the frontend, i.e. the visible part of a website, than the last point of accuracy.

The safety statement on Claude Opus 5.5 also cannot be assessed without further details. For regulated industries, the review of data processing, storage locations and contract processing remains a separate task, regardless of how a model performs in tests.

For web and software projects, this means: the cost side of AI functions is becoming more relaxed, and use cases with high volume are moving within reach. What remains decisive, however, is your own measurability – anyone who works without test data and without cost calculation per process will notice neither improvements nor setbacks.

And anyone who cleanly encapsulates the model call can take advantage of every further price round without touching the application.

Sources