The one-screen answer: August 2026 has six dated AI-deprecation cutoffs, and they're spaced across the month so that missing one is the default outcome, not the exception. Atlas dies today. Anthropic's prompt-tools API goes on the 17th. OpenAI's Assistants API — the big one — sunsets on the 26th, with no grace period. And the 31st carries two at once: Moonshot retires a pair of model IDs, and Sonnet 5's introductory price ends. Here is the whole month on one screen, each with the one-line fix and a pointer to the deep dive.
The calendar, first (because this is the part you came for)#
| Date | Vendor | What ends | The one-line fix |
|---|---|---|---|
| Aug 5 ✓ | Anthropic | claude-opus-4-1 retired; sampling params 400 on Opus 4.7+ | ID → claude-opus-4-8; strip temperature/top_p/top_k |
| Aug 9 (today) | OpenAI | Atlas browser stops working | Export now; move to ChatGPT desktop / Comet / Claude in Chrome |
| Aug 17 | Anthropic | Prompt-Tools API + legacy Workbench | Export prompts/evals; rebuild in one Messages call |
| Aug 26 | OpenAI | Assistants API sunset (no grace period) | Migrate to the Responses API |
| Aug 31 | Moonshot | kimi-k2.5 and moonshot-v1 IDs retired | Repin the model string |
| Aug 31 | Anthropic | Sonnet 5 intro pricing ends | Re-budget: $2/$10 → $3/$15 per MTok |
A note on trust, because this is a facts piece: each date below is anchored to the vendor's own deprecations page or changelog (linked in Sources). The vendors move these, and the exact wording drifts — re-read the linked page before you hard-code a plan around a specific date. That caveat is itself the lesson: the deprecations page is the source of truth, and you should be watching it.
Aug 5 — Anthropic retired claude-opus-4-1 (and a sampling-param trap)#
Already fired, so this one is a check-you-caught-it. On August 5, claude-opus-4-1 stopped resolving — requests to it now error, with no grace period. The drop-in replacement is claude-opus-4-8. Bundled with the same wave: on Opus 4.7 and later, setting temperature, top_p, or top_k to any non-default value now returns a 400, so an agent that pinned a sampling param to be "deterministic" is throwing errors it didn't a month ago. Both are one-line fixes — the full before/after is here.
Aug 9 — Atlas stops working, today#
OpenAI's Atlas agentic browser stops working on August 9, and there is no automatic data migration — if you wired an agent to it, whatever lives only in Atlas leaves with it. The export takes about ten minutes and the honest replacement decision (ChatGPT's desktop app, Comet, Claude in Chrome, or the open-source escape hatch) is laid out here. If Atlas is in your stack, do the export before you finish this piece.
Aug 17 — Anthropic's Prompt-Tools API and legacy Workbench#
Three experimental endpoints — generate, improve, and templatize a prompt — return an error after August 17, and the legacy Workbench that held your saved prompts and evals goes dark with them. There's no one-to-one replacement, and there doesn't need to be: these were convenience wrappers. Export your saved prompts, variables, and evals today, then rebuild the behavior in a single Messages call — a prompt that asks Claude to draft or critique another prompt is a plain request nobody can deprecate. Manual prompt work moves to the updated Workbench (Playground).
Aug 26 — the OpenAI Assistants API sunset#
This is the one to circle. On August 26, every call to /v1/assistants, /v1/threads, and /v1/threads/runs returns an error — no grace period, no degraded mode. The migration target is the Responses API, which folds threads, runs, and tools into one stateful primitive; you carry conversation state with previous_response_id (or the Conversations API) instead of thread objects. The mapping is mechanical but not trivial if you leaned on the Assistants file-search or code-interpreter tools — the exact endpoint-by-endpoint migration, with code, is here. If you're still on Assistants, this is your highest-priority migration of the month.
Aug 31 — a model retirement and a price cliff, same day#
The 31st does double duty. Moonshot retires the kimi-k2.5 and moonshot-v1 model IDs — names that live in older Kimi and Moonshot integrations. The catch is that the like-for-like replacement isn't K3; it's the model you probably overlooked, and the fix is one string per call. The same day, Claude Sonnet 5's introductory pricing ends: from September 1 it moves from $2/$10 to $3/$15 per million tokens — a flat 50% rise that hits base input, output, every cache tier, and the batch rate identically, so caching softens it but won't cancel it. If Sonnet 5 sits in a high-volume loop, re-run the math before the cliff.
Two that already fired in July — worth a compliance check#
Not August, but close enough that a stale integration might still be pointed at them:
- Jul 24 — DeepSeek hard-retired the
deepseek-chatanddeepseek-reasoneraliases with no soft redirect; only the explicitdeepseek-v4-flash/deepseek-v4-proIDs resolve now. Details. - Jul 31 — GitHub Copilot removed Gemini 2.5 Pro and Gemini 3 Flash from every surface — chat, agent mode, inline edits, completions. Move pins to Gemini 3.6 Flash, and check the admin model toggle that decides whether the replacement even shows up. Checklist.
What to actually do with this#
The month's lesson is the same one every heavy deprecation cycle teaches, so build for it once: pin dated model IDs, never a floating -latest in production; keep every ID in one config file so a migration is a single edit; subscribe to each vendor's deprecations page; and hold a golden eval set so qualifying a replacement is a morning, not a fire drill. That discipline — plus a repeatable retirement runbook — turns a month like this from six separate scares into one calendar entry and a grep. The dates will keep coming; the only thing you control is whether you hear about them from a changelog or from your users.
