Whether to pin a model version, and what each choice costs

The instinct is that pinning is the careful choice and floating is the risky one, and for behaviour that's right. A floating alias can move you onto a model that tokenizes differently or answers differently, and you find out from a metric rather than a changelog. What that instinct misses is the direction prices move. Model rates have fallen sharply between generations, so the pinned version is usually the expensive one, and the gap compounds every month nobody revisits it. Anthropic's pricing page makes this concrete by listing deprecated and retired models alongside current ones: Claude Opus 4.1, marked deprecated, is priced at three times Claude Opus 5. Worth separating from a related problem: none of this is model mix drift, where traffic moves between models with the code unchanged. Here the identifier in your code stays exactly the same, and what it means, or what it costs, changes around it. The fix is opposite too. Drift is fixed by pinning. This is fixed by reviewing the pin.

• A model version was pinned once and nobody has revisited it since.
• You're running a model the provider lists as deprecated.
• Nobody knows which model versions are live in production right now.
• Cost per call is flat while newer models have got cheaper.
• A floating alias moved you to a new generation and the token counts jumped.

1. List every model version live in production right now, from your own call records.
2. Check each against the provider's list for a deprecated or retired marker.
3. Re-price last month's volume at the current generation's rates.
4. Decide whether each pin is deliberate, and put a review date on the ones that are.

Illustrative example

Anthropic's pricing page, read 2026-08-05, lists Claude Opus 4.1 as deprecated at $15.00 input and $75.00 output per million, alongside Claude Opus 5 at $5.00 and $25.00. A modelled workload of 10 million input and 2 million output tokens a month is priced at both. The rates and the deprecation markers are published, the volume is modelled.

Three times, and the input and output multiples are both exactly three, which is what makes this easy to miss: nothing about the shape of the bill changes, only its size. The pin was probably the right call the day it was made. It stopped being right the day the successor shipped at a third of the price, and nothing announced that.

Which model version actually served a call is knowledge that exists at the moment of the call and nowhere afterwards, particularly when an alias resolved to something. An invoice groups by model name as the provider records it, which tells you what you were billed for and not what your code asked for, and the two differ precisely when a float has moved. Culpa records the version on every call alongside the feature, customer and run, prices it from a versioned price book in exact decimal, and keeps the ledger on infrastructure you control. That makes a version change visible as a step in cost per call on a specific date rather than as an unexplained movement, and it makes the review answerable: which versions are live, what each is costing, and what the current generation would cost on the same traffic.

• Prompt Version Cost Increase: Finding the Edit Behind It
• What is cost per prompt version
• Model Upgrade Cost Increase: Same Workload, New Rate
• Fine-Tuning Cost vs Base Model: the 6:1 Crossover