GPT Image 2.5 is live — OpenAI's newest image model, targeted edits that leave the rest of the frame alone
All resources

· reAPI Team

Together AI model not available: Catalog vs API access

Fix a Together AI model not available error by checking the exact model ID, serverless support, dedicated endpoint readiness, project scope and API response.

Together AI model not available: Catalog vs API access

In a historical Evo project issue, a Together response reported model_not_available and explained that the request was trying to use a non-serverless model. The suggested next step was a dedicated endpoint.[1] The model existed. The attempted way of calling it was the problem described by that response.

That report is not evidence of a current outage or a particular model's availability today. It illustrates why the full message matters. Together documents different model sets for serverless and dedicated inference, and a dedicated model still needs a usable deployment.[2][3] Start with the request your application actually sent, not the model card you expected it to use.

Copy the model value, host and path

Take the exact model string from the failing request. Keep its namespace, version and variant suffix. Record the resolved API host and request path beside it, along with the HTTP status and complete error body. Remove authentication secrets before sharing this material.

Then establish where the model value came from. A current serverless listing, an old SDK example and a dedicated endpoint screen can supply different kinds of identifier. A familiar display name does not make them interchangeable.

For Together's current dedicated workflow, the inference request uses an endpoint string in the form project_slug/endpoint_name. The catalog model ID identifies the model selected for deployment; the endpoint string identifies the resource serving your requests.[3] A configuration note might therefore contain these two different values:

Model selected for deployment: <catalog-model-id>
Model value used for inference: <project-slug>/<endpoint-name>

These are placeholders showing the distinction, not an endpoint you can call. Fill them from the model and deployment you actually intend to use. Copying the first value into a field that requires the second does not become correct because both refer to the same model family.

Also check the task. Together's catalog covers chat, image, vision, video, audio and other operations, with endpoint guidance for each.[2] A listing for an image model does not establish that it can serve a chat-completions request. Check the operation before changing credentials or assuming that a credit purchase will unlock it.

For serverless, use the serverless listing

If your application is calling shared serverless inference, compare the full ID with Together's current serverless model list. The broader catalog and dedicated-model list answer different questions.[2] This is also the answer to “which models are available?”: available for which task and serving mode must be part of the lookup.

Keep the source of the ID in your debugging notes. If it came from a tutorial written for another route, check that route's documentation before editing the model name by guesswork. If it came from an old example, verify that the exact variant is still served. A nearby name may identify a different model, and a successful call to that replacement does not explain the original failure.

When the documented route has been retired, the problem has moved from configuration diagnosis to migration. Use the Together AI deprecations guide to evaluate the successor and deadline. Repeating the retired ID or rotating its key does not restore a removed route.

If the exact ID is in the correct current list, preserve that evidence and continue to the caller's account and client configuration. You now have a narrower question than whether the provider has the model at all.

For dedicated inference, inspect the deployment you own

Together's dedicated workflow creates an endpoint, attaches a deployment and establishes routing. Inference is available after the deployment is ready.[3] Support for deploying a model does not mean your project already has a ready copy of it.

Find the intended endpoint in the intended project. Inspect its deployment state and the exact inference string shown for it. Check whether the application is calling that endpoint, rather than continuing to send the original catalog ID to a serverless or gateway route.

Project renames are a particularly concrete thing to check after a previously working configuration breaks. Together documents that changing a project slug immediately changes slug-qualified resource references, and the old slug does not redirect.[4] In that case the model and deployment can still exist while your saved endpoint string points to an outdated name.

Before creating another deployment, inspect the one that should already serve the workload. Dedicated inference bills running hardware by the minute, so provisioning an extra endpoint is a capacity decision with a cost attached.[3] It is a poor diagnostic shortcut when a mismatched endpoint name would explain the failure.

Pricing follows the route too. Together's serverless chat listings use token rates, other modalities use their documented units, and dedicated inference uses its hardware billing arrangement.[2][3] Establish which service the request is trying to reach before comparing a token quote with a deployment cost.

The browser's project may not be the worker's project

Together scopes resources and API keys to projects. Access to one project's resources does not automatically establish access to another project's endpoint.[4] The project visible in your dashboard is therefore only part of the evidence.

Consider a hypothetical worker still configured with a key for Project A while you inspect a ready endpoint under Project B in the browser. The dashboard view can be correct, and the worker can still be using the wrong access context. Check the safe credential identifier and the process's actual configuration rather than pasting the secret into a log.

Follow configuration precedence where your client or deployment tool offers several ways to select a project. An explicit setting, environment value and authenticated management context need to agree with the intended resource.[3][4] Include the failing worker in this check; a successful call from your laptop does not establish what a separate process loaded.

If you are using reAPI or another host instead, use that destination's model identifiers and credential format. The reAPI authentication reference describes its own credentials; it cannot activate a Together project or make a Together key transferable.[5] Keep host and credential changes together when comparing configurations.

Check whether the SDK is calling Together directly

The AI SDK's Together integration documents two paths: AI Gateway and the direct @ai-sdk/togetherai provider. In the documented SDK behavior, bare model strings use Gateway by default; the direct provider creates a model instance configured for Together.[6]

This configuration-only example compares Gateway with a direct provider. The placeholder IDs must be replaced with supported IDs for their respective catalogs; it does not send inference requests. It is not a complete dedicated-endpoint setup: that workflow also needs its documented inference host or base URL and the project/endpoint string.

import { togetherai } from '@ai-sdk/togetherai';

const gatewayOptions = {
  model: 'MODEL_ID_IN_GATEWAY_CATALOG',
};

const directOptions = {
  model: togetherai('MODEL_ID_IN_TOGETHER_CATALOG'),
};

The documented Gateway setup uses AI_GATEWAY_API_KEY; the direct provider uses TOGETHER_API_KEY by default. Their model and feature availability can differ.[6] Supplying a Together credential to one part of your environment is not evidence that a bare-string request uses it.

The Gateway provider-order option is another detail worth reading precisely. Its documented order expresses a preference and permits fallback; it does not restrict all requests to the named provider.[6] A successful Gateway request therefore does not, by itself, prove that your direct Together project can serve the model.

Check your installed SDK version and resolved connection settings against the documentation. A client can apply a base URL or other default that is not visible beside the model name in your own code. Compare the constructed request before replacing packages or editing unrelated application logic.

Reduce the diagnostic request to the documented operation and required fields while keeping the intended host, account and model fixed. Then compare it with the application's request. If the minimal form works but the full form fails, restore optional features one at a time so the difference remains visible. If both fail with the same availability message, changing the prompt is unlikely to explain a route mismatch. Any inference you choose to run for that comparison can incur charges; keep it within your normal approved testing budget rather than putting an unresolved request into an automatic retry loop.

For an image, embedding or reranking task, also check that you are using the corresponding provider factory and operation documented by the SDK.[6] A model string passed through the wrong operation is still the wrong request, even if its spelling matches the catalog.

Let the remaining error decide the next change

Once the route and caller agree, read the actual error rather than continuing to interpret every rejection as missing model support. Together's reference distinguishes authentication failures, funding cut-offs, restricted access, invalid URLs or model names, limiting and temporary service unavailability.[7]

A 401 points toward authentication. A 402 can indicate a spend cut-off or positive-credit requirement. A 403 needs the named access reason. A 429 needs its operation context: serverless limiting and a dedicated management request exceeding GPU quota are different cases. The historical Evo response also shows why model_not_available should not be hard-coded as always arriving with 404.[7][1]

Do not use a catalog listing as proof of free hosted access. Together's credit policy reviewed on November 16, 2026 says there is no free trial, requires an initial purchase of at least $15, and says free credits do not satisfy that minimum. Continued prepaid access requires a positive balance; active Scale and Enterprise contracts have their own terms.[8] Those conditions are distinct from a wrong model ID or route. Paying again cannot correct either string.

If the configuration and documented availability still contradict the response, send support that precise contradiction: the exact ID or endpoint, host and path, intended project, timestamp, SDK version, HTTP status and full sanitized error. Include a returned request ID where available. State which current listing or ready deployment you checked. That lets support investigate the request that failed, instead of trying to infer it from a screenshot of a model card.

References

  1. Evo project, GitHub issue 95, model not available on api.together.ai, including responses. Historical report reviewed November 15, 2026.
  2. Together AI, Available models: serverless catalog, task types and dedicated-model distinction. Reviewed November 15, 2026.
  3. Together AI, Dedicated model inference overview: project selection, readiness, endpoint strings and hardware billing. Reviewed November 15, 2026.
  4. Together AI, Projects: resource boundaries, project API keys and slug changes. Reviewed November 15, 2026.
  5. reAPI, Authentication. Reviewed November 15, 2026. External source URLs and full snapshots are retained in the research archive.
  6. AI SDK, Together.ai provider: Gateway and direct-provider configuration. Reviewed November 15, 2026.
  7. Together AI, Error codes. Reviewed November 15, 2026.
  8. Together AI, Credits: free-trial policy, initial purchase requirement and contract exceptions. Reviewed November 16, 2026.
All resources