You Didn’t Get the AI Model You Paid For
The line in the response object
You name the API. You move mannequin: “claude-fable-5”. You get again a completion, a token depend, and a area that reads “mannequin”: “claude-opus-4-8”.
Nothing errored. Nothing retried. The request was labeled earlier than technology started, matched a delicate class, and was handed to a unique set of weights totally. Anthropic documented this when it introduced Fable 5 again on July 1: blocked requests are despatched to Opus 4.8 as a substitute, and the person is notified. The swap occurs at the API layer, and the response object names the mannequin that truly ran.
That is the well-behaved model. It can also be, so far as I can inform, the solely model in huge deployment that tells you the reality in-band.
Two weeks later, Cursor shipped Router: a classifier educated on 600,000-plus reside requests that reads every question’s context, complexity, and area and dispatches it to whichever mannequin it judges finest with three early-access accounts reporting 30–50% financial savings in opposition to routing every part to Opus 4.8. Cursor printed its routing guidelines however didn’t title a selected mannequin per process kind.
And beneath each, at the aggregation layer, OpenRouter warns that some suppliers serve quantized weights at decrease costs, that output can differ from what full-precision weights would have produced, and that your logs is not going to inform you this occurred.
Three merchandise. Three completely different solutions to the query what’s a mannequin. Zero circumstances telling us which one the legislation will settle for.
Three methods a reputation can cease that means a factor
Model id is fracturing alongside three impartial axes, and they aren’t normally distinguished:
- Substitution – A unique structure, completely different weights, completely different functionality profile – dispatched by a classifier. Fable 5 → Opus 4.8. Cursor Auto → no matter the router picks this flip.
- Degradation – The similar mannequin, served at lowered precision. OpenRouter exposes a quantizations parameter exactly as a result of quantized endpoints could carry out worse on sure prompts, and by default requests are load-balanced throughout suppliers ordered by value. Same title in your request. Different arithmetic on the different finish.
- Drift – The similar title pointing at silently up to date weights. Every -latest alias in manufacturing is an unversioned dependency you’d by no means tolerate in a bundle manifest.
The engineering neighborhood treats all three as reliability issues. They are additionally id issues, and id is what contracts, warranties, disclosures, and proof guidelines are constructed on.
So what did you truly purchase?
Start with the oldest query in industrial legislation: was the description a time period of the deal?
If an API name have been a sale of products, this may be near-trivial. UCC §2-313(1)(b) makes any description of the items that types a part of the foundation of the cut price into an categorical guarantee that the items will conform to it. India’s Sale of Goods Act, 1930, §15 does the similar work by the doctrine of sale by description.
But an inference API nearly definitely isn’t items. Courts have usually handled hosted software program as a service, which pushes you out of guarantee statutes and into common-law contract — the place the reply relies upon totally on what the documentation stated and the way particularly the purchaser bargained. That is exactly the place the ambiguity lives. Enterprise agreements value per-model. Model playing cards are model-specific. Compliance artifacts title mannequin variations. And then the routing layer treats the title as a trace.
The cleanest solution to see the stakes: in case your code pins a mannequin ID moderately than expressing a functionality requirement, behaviour can shift materially with none error ever firing. A contract drafted the similar means has the similar defect. If you bargained for a reputation, substitution is breach. If you bargained for a functionality, substitution is ok — and now you want a definition of “frontier high quality” that survives cross-examination.
Nobody has written that definition. Cursor’s is instructive: Router was evaluated in a web-based A/B check optimising for person satisfaction as the reward sign. That is a sound engineering alternative and an odd contractual one. Satisfaction shouldn’t be conformity. A person who by no means observed the swap is proof of router, not of a delivered specification.
The disclosure gradient
Under FTC deception doctrine, a illustration is actionable when it’s materials and prone to mislead a client appearing fairly; goal efficiency claims moreover require an affordable foundation earlier than dissemination. Both halves chunk right here.
On the routing facet, the three merchandise sit at very completely different factors. Anthropic notifies and returns the served mannequin in the response. Cursor publishes guidelines however not per-task mannequin task. OpenRouter’s quantization variance is disclosed in documentation and controllable by parameter however it’s opt-out, and the default path is the low cost one.
Disclosure buried in a docs web page, defaulted in opposition to the person, is strictly the reality sample regulators have been calling a darkish sample in each different client vertical.
On the substantiation facet, the publicity is the declare, not the routing. Commentators have already flagged {that a} “60% cheaper, no high quality loss” declare arrives with out printed methodology, and Anthropic’s personal disclosure is a mannequin of what candour prices: it said plainly that the retrained classifier flags benign requests extra usually throughout routine coding and debugging. That sentence is a legal responsibility protect. The distributors who don’t write it are the ones to observe.
There is a competition-law tail right here too. Working papers on vertical foreclosures in inference markets are already proposing a conduct framework constructed on routing transparency, quality-of-service parity, and FRAND-style non-discrimination. A router that can also be a first-party mannequin vendor is a self-preferencing engine sporting a cost-optimisation costume.
The half no one is : authentication
This is the place I believe the actual struggle lands, and it has nothing to do with billing. Federal Rule of Evidence 901(b)(9) authenticates output by describing the course of or system that produced it and displaying that the course of produces an correct consequence. FRE 902(13) and 902(14) go additional, letting data generated by an digital system, or information verified by hash, self-authenticate on certification. India’s Bharatiya Sakshya Adhiniyam, 2023, §63 does the analogous work, conditioning admissibility of an digital report on a certificates figuring out the report and the method of its manufacturing.
Every certainly one of these provisions assumes you’ll be able to title the system.
Now run the state of affairs – A lawyer recordsdata a short with a fabricated quotation. Sanctions proceedings start. The courtroom asks which mannequin produced it. The agency’s logs say fable-5. The supplier’s logs say a classifier fired and Opus 4.8 answered. Or the request went by an IDE router that chosen a mannequin the agency can not reconstruct, at a precision it by no means specified, on a model that has since been retired.
The chain of custody breaks at the router. Not at the mannequin at the dispatcher. And it breaks in each instructions: the celebration making an attempt to authenticate its personal output can’t, and the celebration making an attempt to assault the different facet’s output has a free-standing reliability argument that no knowledgeable can rebut.
Add the discovery downside. Routing selections are actually dispositive information. That means router classifier logs, per-request mannequin assignments, and precision metadata are discoverable from a vendor who presently has no obligation to retain them.
The repair is a signature, not a clause
You can not draft your means out of this, as a result of the illustration you want is a few runtime occasion, and contracts are static.
What the legislation will ultimately ask for is attestable mannequin id: a signed assertion, returned with every response, binding the completion to a tuple of served mannequin identifier, weights hash, precision, and system-prompt hash signed by a key rooted in {hardware} attestation, which confidential-computing GPU platforms already help.
The primitive is half-built. Response objects already carry the served mannequin title. What is lacking is the property that makes it legally helpful: that it can’t be asserted falsely, and that the purchaser can confirm it with out trusting the vendor. A hash in a response header does for mannequin id precisely what FRE 902(14) already does for copied information turns a contested factual query right into a certificates.
Routing is nice engineering. It can also be, proper now, an unlogged, unsigned, unverifiable substitution of the factor you bargained for. The mannequin ID has quietly turn into a authorized identifier. It is time anyone determined what it identifies.
Sources:
- Cursor, Introducing Cursor Router (22 July 2026) — cursor.com/blog/router
- MarkTechPost, Cursor Releases Cursor Router: A Request-Level Classifier Delivering Frontier Coding Quality at 30–50% Lower Cost (22 July 2026) — marktechpost.com
- Anthropic, Redeploying Claude Fable 5 (30 June 2026) — anthropic.com/news/redeploying-fable-5
- Espressio, Claude Fable 5 Safeguards: The Opus 4.8 Fallback Explained (22 June 2026) — espressio.ai
- Digital Applied, Why Claude Just Got More Cautious About Your Code (2026) — digitalapplied.com
- OpenRouter, Provider Routing — Provider Selection (docs) — openrouter.ai/docs
- OpenRouter, Lowest-Cost LLM Inference: The Complete OpenRouter Guide (12 June 2026) — openrouter.ai/blog
- OpenRouter, How OpenRouter Model Routing Works (12 June 2026) — openrouter.ai/blog
The publish You Didn’t Get the AI Model You Paid For appeared first on MarkTechPost.
