Late July and August Will Retire 19 AI Models and APIs. Here's Your No-Code Survival Guide

If you build with AI APIs and you haven't checked your model IDs recently, the second half of summer is going to cost you. Across five providers (OpenAI, Anthropic, Google, Mistral, and Microsoft Azure), 19 models and API endpoints are being switched off between 30 July and 30 August. Five of them go before the calendar flips to August.

The biggest one isn't a model at all: OpenAI removes the entire Assistants API on 26 August. Every call to /v1/assistants, /v1/threads, and /v1/threads/runs will return an error. No soft deprecation. No grace period.

The concentration isn't a coincidence. AI providers are on overlapping six-to-eighteen-month model lifecycles, and the end-of-quarter retirement window has become standard. What makes this window unusual is the Assistants API joining a routine model refresh cycle, turning a manageable ops task into an architectural migration for anyone who built agent workflows around it.

Here's what's dying, when, and what to do about it.

Which models and APIs are actually going?

Mistral (30 and 31 July)

Mistral Small 3.2, Magistral Medium 1.2, Devstral 2, and Mistral Nemo 12B all retire on 30 July. Magistral Small 1.2 retires on 31 July. Replacements: Mistral Small 4 for the Small, Magistral, and Devstral line; Ministral 3 8B for Nemo. Mistral Medium 3 and Mistral Medium 3.1 retire on 30 August, replaced by Mistral Medium 3.5.

Anthropic (5 and 23 August)

Claude Opus 4.1 (`claude-opus-4-1-20250805`) retires on the Anthropic API on 5 August. The recommended replacement is Claude Opus 4.8. On 23 August, Claude 3 Haiku on Google Cloud shuts down. This is a Google Cloud specific retirement on a separate schedule from the direct Anthropic API. If you checked Anthropic's own deprecation page and saw Haiku 3 listed as already retired, that was for the direct API. The Google Cloud instance stayed live longer.

Google (10 and 17 August)

`embedding-2-preview` retires on 10 August. The replacement is `gemini-embedding-2`, the generally available multimodal embedding model. This is a preview-to-GA migration: the embedding dimensions are compatible and neither re-embedding nor re-indexing is required. You do still need to update your model ID, but you can skip the multi-day vector rebuild that other embedding migrations have demanded. On 17 August, all three Imagen 4.0 variants (`imagen-4.0-generate-001`, `imagen-4.0-ultra-generate-001`, and `imagen-4.0-fast-generate-001`) are retired. The recommended replacement is `gemini-3.1-flash-image`.

OpenAI (10 and 26 August)

`gpt-5.2-chat-latest` and `gpt-5.3-chat-latest` aliases die on 10 August. If your no-code platform pins one of these behind the scenes, your AI steps may silently break. Replacement is `gpt-5.5`.

On 26 August, the Assistants API is removed entirely. `/v1/assistants`, `/v1/threads`, `/v1/threads/runs`: all gone. OpenAI announced this on 26 August 2025, giving a full year, but the migration from Assistants to the Responses API is an architectural change, not a model swap. Assistants become dashboard-only Prompts. Threads become Conversations. Runs become Responses. OpenAI has confirmed there is no automated tool for migrating Threads to Conversations.

A note on Prompts: reusable prompt objects (`/v1/prompts`) are themselves deprecated, with shutdown scheduled for 30 November 2026. Do not build your migration around them. The durable path is the Responses API with prompt content stored in your own application code.

Azure (5 and 28 August)

Three `gpt-chat-latest` preview aliases retire: the 5 May 2026 preview on 5 August, and the 28 May and 24 June previews on 28 August. The Microsoft retirement table lists no GA replacement for these. Every listed `gpt-chat-latest` version is Preview and the retirement schedule shows the replacement column as blank. Azure users should refer to Microsoft's model retirement documentation directly and pin a specific, non-preview model version that is confirmed GA for their deployment region before the deadline.

What does the Assistants API shutdown mean for no-code builders?

This is the one that'll break things without warning if you're not paying attention. The impact depends entirely on how your platform handled the migration.

Zapier deprecated its legacy ChatGPT steps that used the Assistants API. The main "Conversation With Assistant (Legacy)" action is being auto-migrated: Zapier will migrate affected Zaps automatically, then leave them switched off for you to review field mappings and turn back on. "Create Assistant" and "Find Assistant" actions require a manual rebuild. There is no auto-migration for those. Separately, the legacy Chat Completions actions (Summarise Text, Classify Text, Sentiment Analysis) are also being deprecated, but these are distinct from the 26 August Assistants shutdown and follow their own timeline.

Make has not published a direct Assistants API migration path. Its community forum shows users asking how to replace "Message an Assistant" and "Create Assistant" modules, with no official replacement module available. If you built Make scenarios around these modules, expect to redesign them at the API level or rebuild using standard Chat Completions modules. Do not assume a drop-in replacement module will appear before 26 August.

Bubble, WeWeb, and custom plugins: If your plugin or API connector calls the Assistants API directly, you're doing a manual migration. Check with your plugin provider. If the plugin hasn't been updated, you need a new integration path before 26 August.

Which deadlines hit first, and in what order?

The five Mistral models retiring 30 and 31 July are the most urgent. If your app uses Mistral Small 3.2, Magistral Medium 1.2, Magistral Small 1.2, Devstral 2, or Mistral Nemo 12B, you effectively have days. Check your model IDs and swap to the replacements before the end of the month.

The Assistants API audit should start now regardless. Waiting guarantees breakage, and the work isn't a one-afternoon model swap. It's a redesign of how your app structures conversations, stores thread state, and manages tool calls. The shape of the Responses API is different: you manage message history yourself rather than leaning on server-side Threads, and tool calls return inline rather than requiring polling.

The Claude Opus 4.1 and Azure 5 May preview retirements on 5 August give you roughly two weeks. After that: embedding-2-preview and the OpenAI chat-latest aliases on 10 August, Imagen 4.0 on 17 August, Claude 3 Haiku on Google Cloud on 23 August, the Assistants API shutdown on 26 August, the remaining Azure previews on 28 August, and the two Mistral Medium models on 30 August.

Does model-agnostic infrastructure actually help?

Platforms that let you choose your model provider or bring your own model can reduce the blast radius of a single provider's retirement. Stacker supports model choice and BYOL, so you can swap from one provider to another without rebuilding your application logic. That doesn't mean you never need to track deprecations. Model behaviour still changes and output quality still needs verification after any swap. But you aren't locked to a single provider's retirement calendar, and you can route around a deprecation rather than being forced through it.

For teams building directly against provider APIs, model lifecycle management is now a recurring maintenance task. Subscribe to provider deprecation feeds, check `endoflife.date/claude`, and add a quarterly dependency audit to your calendar. The providers aren't slowing down.

Open your AI integrations today. List every model ID and API endpoint your apps depend on. Check the dates above. The work isn't optional. It's finite, and you can finish it before the 30 July Mistral deadline if you start now. The cost of doing nothing isn't a performance dip. It's a hard error.