What is AI model lock-in?
On 30 July 2026, OpenAI cut the price of Luna, the cheapest model in its GPT-5.6 range, by 80 %. On 10 August, Anthropic dropped the increase it had announced for 1 September on Claude Sonnet 5 and turned the introductory rate into the standard one. Both cuts reached existing applications without a line of code moving: same model identifier, no migration to run. The real event was 24 July, when Claude Opus 5 shipped under an identifier nobody was calling yet.
AI model lock-in describes an application that can no longer change model without a rewrite. The model identifier, the expected response format and prompt-specific quirks are scattered through business code. Moving to a newer version takes the shape of a project to fund, settled in the next budget cycle.
The coupling stays invisible while the provider stands still. It surfaces the day a version is deprecated, or the day a more capable model ships under a different identifier for the use case that matters. In 2026, both happen several times a year.
What does an application frozen on a model identifier lose?
Nothing on price, and that is what misleads. Both of this summer’s cuts reached existing applications automatically, on the same identifier, with no migration and no contract change. Capability gains do not travel that way, because they arrive under a different name. Claude Opus 5 shipped on 24 July 2026 at the Opus 4.8 rate, with a higher capability level, and became the default model on Anthropic’s Max plan. An application calling claude-opus-4-8 pays exactly what it paid before, and will never see any of that generation.
Those cuts describe no general trend either. In its 3 September 2026 edition, covering 40 models, BenchLM’s Token Price Index puts the frontier segment up 72 % year over year and the budget segment up 36.8 %. Entry-level pricing is climbing back because a share of those rates were launch promotions, and they expire. Its third segment, mid-tier, has to be set aside: the index still carries a Claude Sonnet 5 increase that Anthropic cancelled on 10 August 2026, with no erratum published. The automatic catch-up therefore applies to the identifier that got the cut, and stops there. At the top, the headline rate of the tier above has not moved: Claude Fable 5.1, released on 1 September 2026, keeps Claude Fable 5’s $10 and $50 per million tokens, twice the blended Opus rate (Anthropic public pricing as of 10 September 2026).
The other deadline is harsher. Anthropic retired six versions from its catalogue between February and August 2026, and a request to a retired model fails. Notice runs to at least sixty days, and every active version carries a floor date for retirement: Claude Sonnet 4.5’s falls on 29 September 2026. An application delivered in spring is already running on a previous generation, and nobody can price the switch: engineering does not propose one, and the decision slides to the next budget cycle.
How do you know your architecture is coupled to one model?
Five questions settle it. Three negative answers mean the coupling is real.
- How many files contain a model identifier? Beyond a single configuration file, you are coupled.
- Can you move one use case to a different version without redeploying the application?
- Do you have an evaluation set that replays your real cases against a candidate model?
- Do you know consumption per use case, or only the monthly invoice?
- Who decides on a model change, and against which written criteria?
The last question is the one that stalls most conversations, and it is the only one of the five that cannot be settled in code. Where nobody owns it, the model in production is still the one picked by the team that shipped the first prototype, and that choice has never been reopened.
What does a model abstraction layer do?
The abstraction layer is a single point of passage between your applications and the models. It serves four functions.
- Decouple: business code calls a capability, say “summarise a claims file”. The capability-to-model mapping lives in configuration.
- Route: each use case goes to the model that fits it. Triaging inbound tickets and reviewing contracts do not demand the same reasoning depth.
- Measure: latency, failure rate, consumption and quality are attributed use case by use case, at the level where the trade-off is actually made.
- Govern: logs, guardrails and approval trails sit in one place, which makes an inventory of AI systems maintainable.
That last function carries a regulatory deadline. Article 50 of the EU AI Act has been applicable since 2 August 2026: public-facing systems must tell users they are interacting with an AI, and generated content must be marked. Systems already on the market on 2 August have until 2 December 2026 on that marking obligation alone. Either way, you need to know which model runs where. Deloitte reported in its 2026 edition that only 21 % of enterprises deploying agents had mature governance in place.
Does this mean going multi-provider?
No, and the confusion is common. Multi-provider is a means, sometimes useful, often expensive in skills and tooling. Committing to a reference model remains a defensible architectural decision: one evaluation harness to maintain, one learning curve to fund.
Our position is narrower than the usual debate. Choosing Claude as the foundation and recording it in versioned configuration is engineering. Scattering it across forty application calls is debt. Versioned configuration leaves you free to reopen the decision at a moment of your choosing.
Second guardrail: a badly designed abstraction levels everything down. If your layer only exposes what every provider has in common, you lose whatever made you pick that model, from long-context handling to structured tool calls. A useful layer exposes the advanced capabilities and documents the ones that are not portable.
Where should you start?
With an inventory, which takes a few days. List the AI applications running in production, the model each one uses, and the date that choice was last reviewed. The exercise produces the same surprise almost every time: forgotten use cases, and two or three applications running on a version the team itself would no longer pick today.
Then lift the model identifier out of the code on your single most critical application, and build the evaluation set that will tell you whether a candidate performs as well on your own data. S&P Global reported in 2025 that 46 % of organisations abandoned their prototypes before production. The ones that clear that bar treat AI as an architectural component, held to the same reversibility standards as the rest of the information system.
Our AI maturity assessment covers this dimension in three minutes, alongside governance and industrialisation, using the Koneetiv framework (2026 edition). For the wider picture, see AI agents in the enterprise and the LOOP™ AI governance methodology.