Your model is being shut down, and you don't know every place it's called.
We find every model ID in your code, configs and gateways, move them to one place, migrate the calls, and replay 50 real prompts on the old and new model before anything ships.
You have this problem if…
- Calls started returning
404 model_not_foundon October 23, or will on November 30, December 11, January 6 or January 14. - Model names are spread across code, environment files, prompt templates, no-code steps and an LLM gateway, and nobody has the full list.
- The replacement is a reasoning model or uses a different transport, so changing the model name breaks tool calls, parameters or streaming.
What you get
- An inventory of every model ID and endpoint in your code, environment files, gateway config and no-code tools, with the shutdown date and replacement for each
- All model names moved into one config, so the next retirement is a one-line change
- Calls migrated to the replacement, including parameter, tool-call and transport changes
- A replay of 50 real prompts on the old and new model, compared on output quality, latency and cost per call
- A rollback plan and a short runbook for the next deprecation notice
Check your exposure
Paste code, an .env file, a gateway config or grep output. The check runs in your browser: nothing you paste is sent anywhere.
Every date the check knows about is on the model-retirement calendar.
How it is done
- 01
Inventory
We scan the repository, environment files, gateway and no-code configs for model IDs, endpoints and SDK versions, and match each against the providers' published retirement dates.
- 02
Choose replacements
For each call we choose the replacement by task, not by name, and write down what changes: parameters, tool calling, streaming, transport and price.
- 03
Capture a replay set
We take 50 representative prompts from your logs, with the outputs your users get today. If there are no logs, we write them with you from real cases.
- 04
Migrate behind one config
Model names move into one place, calls are updated, and a flag lets the old and new paths run side by side.
- 05
Replay and compare
Old and new run on the same 50 prompts. You see the differences in quality, latency and cost, and every regression is fixed or written down before release.
- 06
Release with a rollback
We ship behind the flag, watch error rates, and keep the rollback path until you sign off.
What usually goes wrong, and how we handle it
Anthropic API version 2026-10-21
It rejects temperature, top_p and top_k with a 400 error. The new SDK major versions (Python 2.0.0, TypeScript 1.0.0) default to it, so upgrading the SDK alone can break every request.
gpt-6-luna is a reasoning model
It replaces GPT-4o and GPT-4.1 on January 14. Tool calls over Chat Completions need reasoning_effort: none, and temperature and max_tokens can be rejected.
The tts-1 replacement is a WebSocket session
The replacement for /v1/audio/speech is gpt-realtime-2.1-mini, a Realtime model. A few lines of REST code become a session with connection, streaming and reconnect handling.
The quick fix that retires next
Several teams fixed the October 23 errors by moving to gpt-4o or gpt-image-1-mini, which shut down on January 14 and December 1.
Azure dates are not OpenAI dates
Non-fine-tuned GPT-4o and GPT-4.1 run on Azure until April 14, 2027. Standard Azure deployments of gpt-4o 2024-05-13 are upgraded automatically on December 9; Provisioned deployments are not, so they break.
Why you can trust us with this
UnvibeCode, the open-source tool our founder built, traces where behavior lives in a codebase. The same tracing finds every place a model is called, including the ones hidden in helpers and configs.
The replay comparison applies the method behind the quality gate our founder built, which checks about 50,000 AI-generated items before they reach users.
UnvibeCode: what AI-written code actually does
An open-source tool that reverse-engineers a codebase into business workflows, a connected-code map and evidence-backed risks.
Case studyA quality gate for AI-generated content: 50,000 items checked, about 1 in 10 caught
Deterministic checks on every AI-generated item before it reaches a student. Failing items are pulled or regenerated automatically.
What we don't do on this package
- No silent model swaps. Nothing changes in production until you've seen the replay.
- No migration of fine-tuned models without your evaluation data.
- No engagements that involve health information (PHI).
Questions
Can we buy time instead of migrating now?
Sometimes. Azure keeps non-fine-tuned GPT-4o and GPT-4.1 until April 14, 2027, and some providers offer extended access at higher prices. The audit tells you which calls can wait and what waiting costs.
What if the new model is worse for our use case?
You'll know before release, because the replay runs first. We tune prompts and parameters until the replay passes, pick a different replacement, or tell you plainly that the task needs a design change.
We don't keep prompt logs. Can you still build a replay set?
Yes. We write 50 prompts with you from real cases, along with the outputs you consider correct.
Who does the work?
Suresh Rajendran, our founder, does the engineering on every engagement himself. Nothing is handed to a junior engineer or a subcontractor.
Where does our data go?
We work inside your accounts: your cloud, your repository, your model-vendor keys. Where a vendor offers zero data retention we turn it on, and your data is never used to train a model. Our access is removed and credentials are rotated at handover.
How do we pay?
The first step is paid up front. Builds are 50% to start and 50% on acceptance. We invoice from Destello AI Incorporated in US dollars, payable by ACH or card. If you go ahead with a build within 30 days of the first step, its fee is credited.
What if the result doesn't meet the bar?
Acceptance criteria are agreed in writing before work starts, and the second payment is due only when they're met.
Will you sign our NDA?
Yes, usually within one business day.
Start with the migration audit.
$1,200, credited to the build if you go ahead within 30 days. Book a 20-minute call, and you'll have a written scope with acceptance criteria within one business day.
Or email hello@destello.us or call (628) 202-5322. We reply within one business day.