· reAPI Team
Together AI deprecations: redirects, replacements and migration
Handle Together AI deprecations with a route checklist, replacement tests and recovery plan. See when IDs redirect and when migration needs an explicit change.

A deprecation notice leaves you with two different questions. Will the old model ID stop working, or will it keep working while sending requests to a newer version? Check which event Together describes before changing your application.
Together's lifecycle documentation distinguishes qualifying upgrades, which can redirect an existing endpoint after notice, from materially different new models that require an explicit ID change. It also lists serverless removals and the dedicated support available for particular entries.[1] A recommended successor, an automatic redirect and a removed route are not the same migration.
Read the effective date beside the serving route
Take the complete model string from an actual request and match it to the notice. Include the API host and endpoint you call. A friendly model name in your application settings may hide an alias or deployment name, so it is worth checking the transmitted value before assigning the change to somebody.
Together's general notice table lists two or three weeks for serverless and on-demand dedicated models, depending on usage and replacement availability. Preview models have a different policy after their initial period. The lifecycle section separately gives a three-day notice for qualifying upgrades.[1] Those periods apply to different events. Use the particular entry and account notification to establish your deadline rather than adding a generic grace period to the date you happened to read the email.
Save the notice and the date you checked it. If its removal date conflicts with an account notification, resolve that discrepancy before planning the cutover. An unresolved date is a scheduling problem; choosing the later one because it is convenient does not resolve it.
An ID that still returns HTTP 200 needs attention too. Under a documented redirect, continued access can coexist with a change in the model serving the request. A successful response checks availability at that moment. It does not establish that the answer has the same structure or that an agent will choose the same tool.
Schedules also belong to the service publishing them. OpenAI maintains its own model-specific deprecation and shutdown entries; a deprecated OpenAI model remains usable until its listed shutdown date.[2] That is the place to check an OpenAI API retirement. Neither its dates nor Together's notice periods can be transferred to a different host simply because both list a related model.
Follow one application behavior through the change
Consider a hypothetical support-ticket classifier. It receives a customer's message and returns a category that your application uses to route the ticket. The consumer expects an allowed category in a particular JSON field. This is a design example, not a reported failure of any Together model.
Now replace the model. The new one produces a clear, helpful sentence explaining the right category. A person can understand the answer, but the application still cannot read the expected field. The request completed; the routing job did not meet its contract.
That example suggests a better acceptance check than “the new model answers.” Give the candidate representative permitted requests and pass its responses through the same parser and validation the application uses. Check the chosen category against the intended decision as well. A valid JSON object can still route a ticket incorrectly.
Include cases that matter to the actual product. An ordinary request establishes the basic path. A message that could belong to more than one category tests ambiguity. A long request tests the input and output limits you rely on. A message the workflow should not classify automatically tests whether it can ask for review. These are proposed evaluation cases, not results from a benchmark conducted for this article.
Keep the request settings with each result. If you change the prompt, model, output budget and client together, an improvement or regression has several possible explanations. Begin with a documented request the candidate supports, then add the application's relevant features deliberately. Preserve any necessary differences rather than forcing incompatible options into an apparently identical request.
For an agent that calls tools, follow the operation beyond fluent text. Inspect the selected tool, its arguments and the application's handling of the returned result. A valid tool name with a missing required argument can still stop the workflow. A different action may also need a product decision, even when its arguments are valid.
Record satisfactory cases as well as failures. The purpose is to establish whether the replacement serves the application, not to prove that change is bad. If a difference does not affect the requirement, document it without turning every variation in wording into a migration blocker.
Keeping the model may mean changing how you host it
If preserving the current model is important, inspect Together's dedicated-support entry for that exact removal. Some entries refer to a particular underlying variant. Do not assume that every model removed from serverless can be kept running on demand.[1]
Together describes alternatives including another serverless model, an on-demand dedicated endpoint where supported, and a monthly reserved arrangement. Its migration page distinguishes per-minute on-demand dedicated billing from monthly reserved billing.[1] Keeping familiar model behavior can therefore introduce a different capacity and cost decision.
Confirm the supported variant, the endpoint you would actually call, and whether the deployment can become ready for your account. A model still appearing in a catalog does not prove that the removed serverless route remains usable. The model-not-available guide covers that distinction and the configuration values to inspect.
Bring the classifier example back into this choice. If only a particular model meets its acceptance requirements, dedicated hosting may warrant evaluation. If a supported serverless successor passes the same checks, maintaining a deployment may add work you do not need. The documentation narrows the options; your observed application results and expected workload decide which option fits.
If you instead consider another provider, check the candidate's exact model and request contract. The reAPI catalog is a starting point for supported model discovery, and its authentication documentation describes the credentials required there.[3] A Together key or balance does not transfer. Nor does this guide establish that reAPI hosts every retired Together version or keeps any version indefinitely.
Separate the model decision from the host decision. Changing both can affect identifiers, supported settings, error handling and usage reports at the same time. An OpenAI-compatible request shape is helpful for integration, but it does not certify the classifier's decisions or an agent's tool behavior.
Make rollback usable after the deadline
Before the old route disappears, reverting a configuration value may restore it. After removal, that value points to a route you already know is unavailable. A rollback plan that ignores the removal date can work during rehearsal and fail when you most need it.
For an upgrade that redirects the old ID, changing back to that string may still reach the newer version. Confirm which version the recovery destination actually serves; the unchanged name alone cannot restore the earlier behavior.[1]
Choose a recovery destination that will still be available when the migration is complete. It might be a second supported model that has passed the essential checks, a supported deployment, or a product flow with reduced functionality while an operator intervenes. Those are design options to evaluate, not guarantees that a provider supplies each of them.
For the hypothetical classifier, a reduced flow could send uncertain tickets for human review instead of automatically routing them. That trades automation for a controlled operational burden. Decide whether the team can handle that burden before describing it as recovery. For a workflow that cannot tolerate such a fallback, the alternate model or deployment needs stronger preparation.
Then separate candidate validation from production rollout. If your application supports a limited traffic change, use it to observe the consumer as well as the provider response. Watch parsing failures, incorrect decisions, incomplete answers and the cost of the migrated workload. Set stop conditions from the application's requirements and normal behavior; there is no universal error-rate or latency threshold supplied by this article.
If you rely on managed rollout controls, read exactly what pause and cancel mean. Do not infer that either action restores all traffic to the previous model. Confirm the current traffic destination and the operation needed to move it. The same caution applies to an application fallback: its destination must accept the request and produce a response your consumer can use.
Keep the effective date visible during this work. A small rollout that starts before retirement but leaves most callers on the old route afterward has not met the deadline. The schedule needs room for correcting a failed comparison or arranging a different recovery path.
The last caller may be a scheduled job
After the main application has switched, look at the actual model strings still being sent. Check workers, scheduled reports, saved agent configurations and workflow presets that use the retiring ID. A source-code search is useful, but it will not find a model selection stored in an external automation or database setting.
Give each caller an owner and an expected run time. A nightly report cannot demonstrate a successful migration during an afternoon test of the web interface. Observe its next relevant execution, or validate that caller through the team's normal approved procedure, before marking it complete.
Suppose the hypothetical support classifier has migrated, while a weekly quality report still requests the old model to summarize classifications. The report is a separate caller even though it belongs to the same product. Changing the classifier's central setting only covers it if the report actually reads that setting. Check the request metadata rather than inferring shared configuration from shared ownership.
Keep the notice, accepted candidate results and final caller list together. Retain obsolete settings until their owners establish they are no longer used, then remove them through the normal change process. For a video-specific retirement, the Sora migration guide addresses the additional asset concerns. For this migration, the decisive evidence is the request from that last caller reaching the intended supported route and producing a result its application can use.
References
- Together AI, Deprecations: model lifecycle, notice periods, removal tables and migration guidance. Complete official page reviewed November 15, 2026; source URL and snapshot retained in the research archive.
- OpenAI, Deprecations: lifecycle states and model-specific shutdown schedule. Reviewed November 16, 2026.
- reAPI, Authentication: credentials and authentication errors. Reviewed November 15, 2026.