Within two days, both major providers of AI language models have followed up – and not at the top of the performance range, but in the class below.Anthropic has released Sonnet 5.5, OpenAI GPT-6.1 Sol.Neither model is marketed by its manufacturer as a new benchmark, but as the model you actually use in everyday work.
This is a remarkable shift in communication.For years the message was: the new model can do more.Now it is: the new model can do almost as much, but costs less.For companies that integrate AI assistance productively into development and editorial processes, this is the more relevant statement.
What the two announcements say
Anthropic followed up with Sonnet 5.5 around three months after Sonnet 5.According to the manufacturer, the model comes close to Opus 5.5 – the provider's more powerful model line – but is the better choice in everyday work.The argument therefore targets less benchmark values than the ratio of output quality, speed and cost across many requests.
OpenAI positions GPT-6.1 Sol similarly according to its own account: cheaper than GPT-6 Astra, yet almost as good.Here too, the focus is on the classification between cost efficiency and performance, not on a new top score.
Both manufacturer statements are initially exactly that: statements by the manufacturers.How large the gap to the respective top model is in practice can only be assessed in your own use case.However, the short interval between the releases shows how quickly the market is currently moving: a model choice made a quarter ago may already no longer be the most economical today.
Why the model class below is often sufficient in everyday project work
In development projects, the actual workload with AI assistants rarely looks like the providers' demonstrations suggest.The majority of requests consist of recurring, well-defined tasks:
- Reading and explaining code, for example in unfamiliar or older codebases
- Generating tests and test data
- Refactorings with a clear objective
- Preparing migration steps between framework or TYPO3 versions
- Formulating documentation and commit descriptions
- Classifying error messages and forming hypotheses about the cause
For these tasks, the most expensive available model is usually not necessary.What matters is response time and reliability across many runs.Sonnet 5.5 and GPT-6.1 Sol target exactly this pattern.
Conversely, there are tasks where the leap to the top model remains worthwhile: complex architecture decisions, errors that are hard to reproduce, long dependency chains across many files.The practical consequence is therefore rarely either or, but a deliberate division.
Costs arise in operation, not in price comparison
A lower price per request only reduces total costs if the number of requests remains stable.In practice, the opposite often happens: the cheaper and faster a model responds, the more often it is called – including for tasks that no one would previously have delegated to an AI.This can be productive, but it drives up the total.
Add to that the rework effort.A model that delivers a usable result in four out of five cases is more expensive than the pure token price suggests if the fifth case is noticed in review – or worse: only in production.The reliable metric is therefore not the price per request, but the effort per completed task including review and correction.
How to organise model changes without project risk
Anyone who integrates AI assistance firmly into processes should treat the model choice as a replaceable component and not as a foundation.A pragmatic approach has proven itself:
- Introduce abstraction: Access to models runs through a dedicated layer, not scattered across the codebase.A change is then a configuration change.
- Define your own comparison cases: Ten to twenty typical tasks from real everyday project work, with a known good solution.These cases say more than any general leaderboard.
- Deploy in tiers: Routine tasks on the cheaper model, difficult cases specifically on the stronger one.Escalation instead of permanent operation in expensive mode.
- Make consumption visible: Record costs per team, per project and per task type before the bill becomes a surprise.
- Maintain the review requirement: Generated code goes through the same review and testing process as hand-written code.
Data protection and traceability remain part of the decision
The model choice is not just a cost question.Which content is transferred to which provider, which processing locations apply and how results are documented must be regulated independently of the performance class.Switching to a cheaper model from the same provider changes little – switching between providers very much does.Such checks belong before the technical changeover, not after.
Assessment for web and software projects
For projects, this means: the decisive question is no longer which model is the strongest, but which is the most economical for which task.Anyone who builds their applications and tools so that the model remains replaceable can evaluate such announcements in days rather than months in future.And anyone who maintains their own project-specific comparison cases makes this decision on the basis of their own codebase rather than on the basis of vendor promises.