AI · Automation · Engineering

LLM Deadlines in 2026: Models, APIs, and Credits

By Lazar MilicevicNovember 19, 202611 min read
Developer workstation used to track LLM model, API, and credit deadlines in 2026

A production LLM application can keep passing its own tests while a provider changes the thing those tests depend on. The model name still sits in configuration, the request still looks valid, and then a scheduled job starts failing after a retirement or API change. I want one place to track the provider dates that matter to builders, without pretending there is a universal calendar that stays current on its own.

This is my working snapshot as of November 19, 2026. It covers documented model retirements, API changes, and startup-credit programs from OpenAI, Anthropic, Google Cloud, AWS, and Microsoft. Use the linked provider pages as the authority before making a migration or funding decision. A date is useful only if you also know which part of your system it affects and who owns the response.

The dates I would put on an engineering calendar

The useful deadline is not always a model shutdown. A read-only eval service can block a release process, while an API parameter change can break otherwise valid requests. I track each notice against a concrete dependency: the model ID, endpoint, workflow, or program application it might affect.

Provider Timing in provider notice Change to track What I would check
Google Cloud October 16, 2026 Vertex AI release notes list this retirement date for Gemini 2.5 Pro, Gemini 2.5 Flash-Lite, and Gemini 2.5 Flash. Search deployed configuration and logs for these model IDs. The date has passed; the release note alone does not establish their current operational status.
Google Cloud After November 26, 2026 Vertex AI Extensions is deprecated and scheduled for shutdown after this date. Google recommends Agent Platform. Identify workflows that call Extensions and test the replacement path.
OpenAI November 30, 2026 /v1/prompts and reusable prompt objects shut down. OpenAI recommends moving prompt content into application code. Find runtime prompt-object references and establish where prompt versions will live.
OpenAI November 30, 2026 The Evals dashboard, API, and graders for eval workflows become read-only. OpenAI recommends Promptfoo. Confirm that release checks can still create and run the evaluations they require.
OpenAI November 30, 2026 Agent Builder is deprecated. OpenAI recommends the Agents SDK or ChatGPT Workspace Agents; ChatKit remains available. Determine whether the application depends on Agent Builder, rather than assuming every agent-related component is affected.
Anthropic November 30, 2026 Claude Sonnet 4.5 (claude-sonnet-4-5-20250929) retires. Anthropic recommends Claude Sonnet 5.5. Find exact model references, including fallbacks and background workers.
OpenAI January 6, 2027 Active existing customers can no longer create new self-serve fine-tuning jobs. Check whether the product roadmap depends on starting new jobs.

The OpenAI items come from its API deprecations page. Anthropic publishes the Sonnet retirement in its model deprecations documentation. Google documents the Gemini dates in its Vertex AI generative AI release notes and the Extensions change in its Vertex AI release notes.

I keep past dates in the register until I have checked production. A passed retirement date is not a reason to assume a service is still available, nor proof that every deployed workload has stopped using it. I also keep adjacent scope out of the main calendar when it does not affect my application. Google’s release notes, for example, recommended migrating listed Imagen generation endpoints before June 30, 2026. That belongs in an image-generation dependency review, not in an LLM text pipeline merely because both run on Vertex AI.

This is a curated snapshot, not an exhaustive list of provider changes. Each provider maintains its own documentation and can update it.

Model retirements need a workflow migration, not a string replacement

When a model is retired, replacing its identifier is the smallest part of the job. I treat the change as a workflow migration: the same inputs must reach the new model, tool calls must remain valid, and the output still has to satisfy the consuming application. Passing a connectivity test proves very little.

In BizFlowAI ContentStudio, the content pipeline researches, writes, optimizes, and publishes across sites without someone watching every run. That shape of system makes a model change more consequential than it looks. A writer step might return well-formed text while drifting from the structure an optimizer expects. An optimizer might succeed while the publishing stage receives material that should not go live. The application-level checks matter more than a successful API response.

My migration sequence is practical:

  • Inventory every model reference. Search application configuration, worker settings, fallback routes, test fixtures, and any stored workflow definitions. A model used only when the primary provider fails is easy to miss.
  • Capture representative work. Keep inputs and expected properties from the paths the application actually runs, including awkward inputs and tool-heavy cases. Do not evaluate a publishing pipeline solely on a friendly prompt.
  • Run the replacement without side effects. Let it draft or propose actions, but prevent publication, customer notifications, or database writes while comparing behavior.
  • Check the contract at each boundary. Validate structured output, tool arguments, refusal handling, and the conditions under which a human would need to review the result.
  • Roll out with a way back. Keep model selection in configuration, record which model handled each run, and make rollback independent of a code release where possible.

For the Anthropic retirement in the calendar, I would start by locating claude-sonnet-4-5-20250929, then test Claude Sonnet 5.5 against the actual jobs that use it. Anthropic’s recommendation identifies a replacement; it does not certify my prompts, retrieval context, or downstream assumptions. The same principle applies to a Google model whose listed retirement date has already passed: inspect the deployed workload and current provider documentation before deciding what remediation remains.

A migration is complete when the workflow’s behavior has been checked, not when the new model returns a response.

API changes can break a working model integration

Model names get attention, but request contracts can fail just as abruptly. I track API versions, parameters, and feature availability alongside model retirements because a valid model choice does not protect an invalid request. The test I want is the exact request my production client sends, including headers and optional parameters.

Anthropic’s documentation says API version 2026-10-21 removes temperature, top_p, and top_k; requests using them return a 400 error. That is a clear example of why I do not treat an SDK update or a version-header change as routine maintenance. A shared request builder may add one of those parameters to every call, even when an individual workflow has no reason to use it. Before changing the version, I would inspect the final serialized request, test each calling path, and remove unsupported parameters at the boundary where they are introduced. The Anthropic documentation is the source for the version behavior.

Feature availability also changes architecture. OpenAI announced on November 17, 2026 that GPT-6.2 Astra and GPT-6.2 Sol support both the Responses API and Chat Completions, but tool calling requires the Responses API. If I were testing either model for an agentic workflow, I would first confirm the application’s API path. Otherwise, a model evaluation could incorrectly conclude that the model cannot perform a required task when the integration is using the wrong interface.

The same changelog describes a public beta background-completion option for the Responses API with two-hour and eight-hour windows. Its stated model token rates are 30% and 40% lower than Standard, respectively. I would consider that for work that can wait, such as offline analysis, rather than a request awaiting an interactive answer. Those discounts describe model token rates, not a guaranteed reduction in an application’s total bill. Queueing behavior, retries, and the cost of delayed results still belong in the decision.

For API changes, I keep a small compatibility record: provider, API version, model, endpoint, required features, and a captured test result. That record is much more useful during a migration than a note saying “the SDK is up to date.”

Startup credits have eligibility windows, not universal deadlines

Startup-credit programs belong in this tracker, but they are different from model sunsets. The provider materials here describe potential credit amounts and application processes; they do not establish a shared 2026 closing date. I would track eligibility, review status, credit coverage, and expiration in the account, rather than inventing a public application deadline.

Program What the provider says Planning implication
Google for Startups Cloud Program AI startups may receive up to $350,000 in Google Cloud credits over two years, with up to $250,000 in year one and up to an additional $100,000 in year two. Check the official program page for current eligibility and application details. Check eligibility and which services the credits cover before including them in an LLM cost model.
AWS Activate Its guide describes a Founders tier with $1,000 in credits and a Portfolio tier with up to $200,000. It also says additional credits above $200,000 may be available to AI startups through an account manager. AWS says Activate credits can be used with Amazon Bedrock. Confirm the tier, account conditions, and covered usage rather than budgeting against the maximum.
Microsoft for Startups Eligible startups can unlock up to $150,000 in Azure startup credits over time as they demonstrate verified progress, service adoption, and sustained usage. Microsoft says applications are typically reviewed within three business days. Treat credit availability as conditional and staged; the cited overview does not give a fixed 2026 application deadline.

I separate cash cost, credited usage, and architecture choice in planning. Credits can make a proof of concept affordable without making its long-term unit economics sound. They also should not force a workload onto a platform whose model access, data handling, or operational path does not fit the product.

For an AI MVP, I would price the system as if the credits were unavailable, then show the credits as a temporary offset after confirming the applicable program terms. That makes the decision legible to a founder or CTO: the application can still be evaluated on its actual operating cost, not only on a promotional balance.

I track dependencies, not just announcements

A deadline register becomes useful when each entry has an owner and a verification path. I would keep provider notices separate from production facts: a documentation page tells me what a vendor has announced, while my configuration, logs, and tests tell me whether my system is exposed.

I would give each entry a source link, the relevant date, affected dependencies, an owner, a migration decision, and test evidence. I would mark each change as announced, past but unverified, migration in progress, or verified. That last distinction matters. Closing a ticket because its date passed is not evidence that a scheduled worker survived the change.

I look for dependencies outside the main request path. A fallback model may never appear in ordinary test traffic. A nightly evaluation job may rely on an API that is about to become read-only. A prompt stored with a provider may be referenced by a worker that has not been deployed recently. Those are precisely the cases a date-only calendar misses.

The OpenAI Evals change illustrates the point. OpenAI says the dashboard, API, and graders become “read-only” on November 30, 2026, and recommends Promptfoo. I would verify not merely that old results remain visible, but that the team can create and run the evaluation workflow needed for its next release. Likewise, moving reusable prompt content into application code needs a versioning decision: which change ships with the code, which tests cover it, and how can a bad prompt revision be rolled back?

In my autonomous content work, a measure, learn, target, generate, publish loop depends on scheduled execution and on the meaning of outputs at each stage. I would attach provider-change checks to those stage boundaries. That is more actionable than subscribing to every announcement and hoping an engineer recognizes the relevant one.

What I’d do

I would review the near-term items in the calendar against deployed configuration first, then assign an owner to each affected workflow. For every migration, I would require a captured production-shaped request, a side-effect-free test of the replacement, and a clear rollback path. I would check the provider page again when making the change, because this article is a dated snapshot, not a substitute for its live documentation.

For credits, I would apply when the program fits the startup and verify coverage before using any advertised maximum in a budget. For models and APIs, I would prioritize the dependencies that can interrupt unattended jobs or block releases. A calendar tells me when to look; the dependency map tells me what to fix.

If you are working through a model migration or building an LLM application that needs to run unattended, you can get in touch through my website at lazar-milicevic.com. I write practical essays about these engineering decisions.

Frequently asked questions

Which LLM deadlines should I put on my engineering calendar in late 2026?

As of November 19, 2026, I would track the November 26 Vertex AI Extensions shutdown threshold and the November 30 changes to OpenAI prompts, Evals, and Agent Builder. Anthropic also lists November 30 for the retirement of Claude Sonnet 4.5. For January 6, 2027, I would check whether the restriction on new self-serve fine-tuning jobs affects my OpenAI roadmap. I would verify each notice with the provider before acting.

What happens to OpenAI prompts, Evals, and Agent Builder on November 30, 2026?

OpenAI says `/v1/prompts` and reusable prompt objects shut down on November 30, 2026, and recommends moving prompt content into application code. Its Evals dashboard, API, and graders for eval workflows become read-only, so I would check whether release gates still need to create or run evaluations. Agent Builder is deprecated, with the Agents SDK or ChatGPT Workspace Agents recommended as alternatives. I would identify which of those components my application actually uses; ChatKit remains available.

How should I migrate a production workflow when an LLM model retires?

I start by finding every reference to the model, including fallbacks, workers, fixtures, and stored workflows. Then I run representative inputs through the replacement without allowing publication, notifications, or database writes. I check structured outputs, tool arguments, refusal handling, and downstream behavior before rolling out. A successful API response alone does not prove the workflow is safe.

Which Google Cloud AI changes does this 2026 deadline list cover?

Vertex AI release notes list October 16, 2026, as the retirement date for Gemini 2.5 Pro, Gemini 2.5 Flash-Lite, and Gemini 2.5 Flash. Because that date has passed, I would inspect deployed model IDs and logs rather than infer current availability from the notice alone. Vertex AI Extensions is deprecated and scheduled to shut down after November 26, 2026; Google recommends Agent Platform. I would test any replacement workflow before switching it over.

How do I check whether a provider deadline affects my LLM application?

I map each provider notice to a specific model ID, endpoint, workflow, or program application. I check deployed configuration and logs, including fallback routes and background jobs, instead of relying only on what appears in the main code path. I keep past dates in the register until I have checked production, and I use the provider’s current documentation as the authority. I do not treat an unrelated endpoint retirement as an LLM text-pipeline deadline merely because it comes from the same provider.

Lazar Milicevic

Lazar Milićević

Senior Technical Engineer. I build AI automation, GenAI/LLM systems and cloud architecture — autonomous systems that run while you sleep. Founder of BizFlowAI.

Building something hard with AI or automation? I am open to talk.

Get in touch

← All posts