Skip to content

Migration Cost Estimator

What moving to a different model costs once re-embedding, dual-running two indexes, repeated eval runs and prompt rewriting are all counted — and how long the saving takes to repay it.

One-off cost of the migration
$8,351

Repaid in 13.9 months at the saving you entered. Engineering time is 95.8% of it.

Re-embedding the corpus
$2.40
Dual indexes for 2 month(s)
$300.00
Eval tokens — input
7.20M
Eval tokens — output
1.80M
Evaluation, 12 full run(s)
$48.60
Engineering, 80 h × $100.00
$8,000
Total one-off
$8,351
Engineering share of it
95.8%
Monthly change after the move
$600.00
Payback
13.9 months
What this assumes: that the corpus is re-embedded once at the new model’s rate. If the move also changes your chunk size the token count changes with it, so re-count rather than reusing the old figure. The eval line assumes the whole suite is run start to finish on each pass, which is what actually happens — you fix a prompt, you run it all again. No rollback is costed, and a rollback means paying the storage overlap a second time.

The cheap part is the tokens

Every migration estimate that comes in low made the same mistake: it costed the re-embedding and stopped. Re-embedding a corpus is a predictable multiplication and it is usually the smallest term on this page. The expensive parts are the ones denominated in hours — prompts that were tuned against the old model’s quirks and do not port, output parsers that break on a slightly different JSON habit, and the fortnight of watching production afterwards to find out what changed that nobody tested for.

The evaluation line is the one people find surprising once it is written down. A full suite is not run once; it is run after every fix, and each run bills the whole set at both rates. Twelve runs of a five-hundred-case suite is an ordinary migration, not a bad one.

Reading the payback

If the payback is short, the decision is easy and the risk is operational rather than financial. If it runs past a year, be honest that the migration is being done for a reason other than cost — quality, a deprecation notice, a data-residency requirement — and justify it on that. Cost paybacks in this field are also perishable: prices move downward often enough that a migration justified purely on rate arbitrage can be overtaken by a price change on the model you were leaving. Deprecation deadlines do not move, and they are the reason most migrations actually happen.

Migration Cost Estimator · Multigrid