What breaks next

New Cowork tasks on Pro and Max plans run on Anthropic’s servers, and the option to keep them on the user’s own computer is gone. Local folders stay on the machine, but each file a cloud task needs is copied up for that session, and scheduled tasks that touch local files need the desktop app open. Work that must not leave one machine belongs in Claude Code instead.

Business and Enterprise organizations get the replacements switched on automatically unless an admin turned off the global default or disabled the model; those admins have to enable the replacements in Copilot model policy themselves. Workflows and integrations that name one of the retiring models need to move before the cutoff.

Only Agent Registry moves: endpoints, IAM actions, ARNs, SDK clients, CLI commands, EventBridge sources and CloudWatch metrics all change, and the request and response schemas break. Data is not migrated automatically; AWS ships migration tooling that has to be run by hand. The shutdown was originally set for September 17, 2026; that date passed without the namespace going off, and AWS’s migration guide now puts the close of the window at October 30, 2026, after which remaining data in the old namespace is lost. Workload identity and OAuth credential provider resources stay under bedrock-agentcore.

Anthropic deprecated Claude Sonnet 4.5 on September 30, 2026, with retirement on November 30, 2026 across the Claude API, Claude Platform on AWS and Microsoft Foundry. Requests to the model string fail after that date; Amazon Bedrock and Google Cloud set their own schedules. Move pinned configs to claude-sonnet-5-5 and use the Console usage export to find remaining callers.

OpenAI removes four text-to-speech models from the API on January 6, 2027, and names gpt-realtime-2.1-mini as the replacement for all of them. Speech output pinned to one of these identifiers needs moving and re-checking for voice and latency before the cutoff.

Requests naming gpt-5.1, gpt-5.3-codex or gpt-5.4-nano stop working when the models are removed from the API on April 1, 2027. OpenAI recommends gpt-6-sol in place of gpt-5.1 and gpt-5.3-codex, and gpt-6-luna in place of gpt-5.4-nano; re-run evaluations on the replacement before switching pinned workloads.

Part of the 2026-07-28 authorization hardening. DCR keeps working for now; no removal date has been set.

The 2026-07-28 revision made the protocol stateless and deprecated four long- standing features with a twelve-month floor before removal. No firm removal date has been published, so the cutoff here is open.

Business and Enterprise admins have to enable the replacements in org model policy first, or their developers lose the option rather than getting it swapped.

The four models were deprecated across all Copilot experiences on the day the change was announced. Workflows and integrations that name one of them need to switch to the listed replacement, and enterprise admins have to enable the replacements in Copilot model policy.

Applications close October 1; anything approved by September 30 carries student coupons good for six months, and every bootcamp coupon plus the graduation discount expires April 1, 2027. The alternatives JetBrains points to key off accredited institutions, which is the thing a bootcamp is not.

Requests naming gpt-5.4-cyber stop working once the model is removed, so any security tooling pinned to it has to move to whichever cyber model the account can access. The notice was far shorter than the three months OpenAI’s policy sets for specialized variants, which allows less only for safety or compliance reasons; OpenAI did not say which applied.

Removed across chat, inline edits, ask and agent modes, and completions. Sonnet 4.6 survives for individual subscribers on annual plans. If an agent config pins a version rather than a family, that pin is now dead.

Spark stopped accepting new users and new apps immediately, with existing accounts losing access on August 31. Deployed apps keep running, but exporting the code meant opening the workbench and creating a repo before that date, and anything calling llm() needs its own inference provider afterwards.

Disabling a connector permanently deletes its indexed data, and the deprecation notice did not say whether the shutdown triggers that automatically. Nothing migrates itself: admins install the replacements, then every user reconnects their account.

Decommissioned three months after sign-ups closed. Accounts, repositories and organisations survive intact; classroom and assignment names, tests defined outside repositories, and LTI-integrated rosters fall under the retention policy. What replaces it is a list of partner solutions in the educator community discussion, not a GitHub product.

Prompts replace Assistants, Conversations replace Threads, and Responses replace Runs. OpenAI stated plainly that there would be no automated tool for migrating Threads to Conversations — the guidance is to route new conversations to the new API and carry old ones across as they come up.

Announced alongside the GPT-4.5 retirement, o3 carried a longer 90-day sunset before leaving ChatGPT.

Removed from chat, inline edits, ask and agent modes, and completions. Enterprise admins have to enable the successors in Copilot policy before anyone sees them in the model picker.

Playground, model catalog, inference API and bring-your-own-key all went away, including for existing customers with live usage. Wind-down began July 1 and new signups closed June 16. Free inference attached to a token you already had was a convenient CI default and now needs a budget line.

Cognition retired the Windsurf brand and relaunched the editor as Devin Desktop, shipped as an over-the-air update with plans, pricing, settings and extensions carried over. The local agent Cascade was replaced by Devin Local, rewritten in Rust, and stayed available only through July 1.

Noted in the same deprecation docs as the Claude 4 retirements.

OpenAI confirmed a 30-day sunset for GPT-4.5 in ChatGPT as GPT-5.5 Instant settled in as the default.

Deprecated since April 14 and retired across the Claude API, Claude on AWS and Microsoft Foundry. The dated 4.0 strings now fail rather than falling back silently, so anything pinning them in CI or an agent config breaks outright.

GPT-5.2 stopped being selectable and existing conversations continued automatically on the matching GPT-5.5 model. The cut followed OpenAI’s policy of keeping a ChatGPT model available for roughly 90 days after its successor ships. The silent in-place migration means prompts or evals tuned against GPT-5.2 now run on a different model without anything being changed.

V4 Pro requests were due to start landing on V4.1-Flash at Flash billing from 04:00 UTC. The changelog now says the API keeps serving V4 Pro past September 14 with billing unchanged, after customers asked for it. The older deepseek-v4-flash and deepseek-v4-flash-vision-exp names stay retired and still route to V4.1 Flash.