Pick a model name once and, sooner or later, the vendor puts a date on it. Then it's a calendar problem. This page keeps those dates in one place: API and model retirements, price cliffs, service closures, limit changes and breaking changes, each with the action it asks of you. It's built from the news we've covered, then checked again against the vendors' own pages. If you run systems that call AI APIs or sit behind someone else's service, start with the Coming up table.
Last verified: 7 October 2026. Every future date in the first table was read on the vendor's own page that day. Where a vendor hasn't published a date, the row says so.
How to read this page
Each row gives a date, what happens, who gets hit, what to do, a status, a source and our own write-up when we have one. The "What to do" column is the reason the page exists. It's a task, not a summary: change this parameter, swap that alias, check this limit, budget for that price.
The status pill tells you who said it. Confirmed means the vendor's own page states the date. Reported means only the press or a third party does, and the cell names them. No date yet means the vendor announced a change without a day. "At least" and "not sooner than" are floors, not dates, and we write them that way.
Vendors contradict themselves more than you'd expect. Google's pricing page still says gemini-2.5-flash-image shuts down on 2 October 2026, while its deprecations page, updated the day we checked, says 15 March 2027. We follow the deprecations page. A handful of rows, mostly OpenAI, Google and Anthropic shutdowns, come straight from those vendors' deprecation pages. Our articles cover the replacement models, not the shutdowns, so the last column points at the closest piece and says "Related". And where the source has overtaken one of our older articles, we link the source only.
Coming up
Nearest first, from 7 October 2026. Rows with no published day sit at the end.
| Date | What happens | Who is affected | What to do | Status | Source | Our write-up |
|---|---|---|---|---|---|---|
| The certificate behind the Windows boot loader expires. Devices without its 2023 replacement keep booting but get no new early-boot protections. Windows signing moves to RSA-3072 and SHA-384 by year end. | Windows fleets, dual-boot machines, vendors whose software pins Microsoft certificates. | Check the Windows UEFI CA 2023 is in each device's Secure Boot DB (Windows Update covers most; managed fleets use registry keys, Group Policy or Intune). Linux: fwupdmgr update once your vendor ships the key. Ask vendors if their software pins certificates. | Confirmed | Microsoft: Secure Boot certificate expiration Microsoft: next-generation code signing | Secure Boot expiry, the Linux fix | |
Google shuts down veo-3.1-generate-preview, veo-3.1-fast-generate-preview, veo-3.1-lite-generate-preview and gemini-omni-flash-preview. | Gemini API users who generate video. | Move to gemini-omni-1.1-flash, the replacement Google names. Re-price first: our write-up of Google's own table has Veo 3.1 Lite at half the Omni 1.1 Flash rate at 720p. | Confirmed | Gemini API deprecations | Related: Gemini Omni 1.1 Flash | |
OpenAI shuts down older snapshots, including gpt-4, gpt-4-turbo, gpt-3.5-turbo, gpt-4o-2024-05-13, o1, o1-pro, o3-mini, o4-mini, gpt-4.1-nano, gpt-image-1, plus fine-tunes built on several of them. | API users still pinned to those names, and owners of the listed fine-tunes. | Grep code and config for the names. OpenAI's replacements are gpt-5.6-sol, gpt-5.6-terra and gpt-5.6-luna (images: gpt-image-2.5-sunburst or gpt-image-2.5-flare). Rerun evals instead of swapping strings. | Confirmed | OpenAI deprecations | Related: o3 retirement | |
| Last day to approve, reject or bonus work submitted before the closure. Anything left untouched is auto-approved. Prepaid balances are refunded within 30 days of the 30 September closure. | MTurk requesters. | Review every pending HIT before the 30th, because untouched work gets paid. Export qualification and assignment data now. Move any Ground Truth job off the MTurk worker type. | Confirmed | MTurk closure notice and FAQ | MTurk wind-down | |
| Mullvad shuts down its public encrypted DNS: 194.242.2.2, .3, .4, .5, .6 and .9 and the matching dns.mullvad.net hostnames. Mullvad's iOS and macOS DoH profiles stop working. | Anyone who hard-coded those addresses in routers, resolvers or browsers. Mullvad VPN users aren't affected. | Run grep -rn 194.242.2 /etc/, then switch to Quad9: 9.9.9.9 or dns.quad9.net. Quad9 blocks malicious domains, not ads. If you used the adblock endpoints, run your own blocklist forwarder. | Confirmed | Mullvad announcement Mullvad DoH and DoT page | Mullvad DNS and Quad9 | |
Google shuts down five audio previews: gemini-3.1-flash-live-preview, gemini-2.5-flash-native-audio-preview-12-2025, gemini-3.1-flash-tts-preview, gemini-2.5-flash-preview-tts and gemini-2.5-pro-preview-tts. | Voice agents and text-to-speech pipelines on the Gemini API. | Move Live sessions to gemini-3.8-live and speech to gemini-3.8-flash-tts or gemini-3.8-flash-lite-tts, as Google names them. If you take the Extended Thinking variant, every function must be declared NON_BLOCKING. | Confirmed | Gemini API deprecations | Related: Gemini 3.8 Live | |
| 21 Nov 2026 (at least) | OpenAI's pricing page says GPT-5.6 Sol's promotional rate ($4 in, $20 out per million tokens) holds at least through 21 November. It doesn't say what applies after. | Anyone budgeting on gpt-5.6-sol. | Don't budget past 21 November on $4 and $20. Test gpt-6-sol and gpt-6.1-sol, which list at $2 in and $10 out. | Confirmed OpenAI; the date is a floor and the later price is unpublished | OpenAI API pricing | GPT-6 Astra pricing |
Anthropic retires claude-sonnet-4-5-20250929 (deprecated 30 September) and names claude-sonnet-5-5 as the replacement. Bedrock and Google Cloud keep their own schedules. | Anyone pinned to Sonnet 4.5 on the Claude API, Claude Platform on AWS or Microsoft Foundry. | Plan a staged move, not a string swap. Sonnet 5.5 returns a 400 for thinking: disabled and for forced tool_choice. Test between_tools, strict tools and re-tuned effort. | Confirmed | Anthropic model deprecations | Related: Sonnet 5.5 breaking changes | |
| The grace period for machine-readable marking under Article 50(2) ends for generative AI systems already on the EU market before 2 August 2026. Newer systems had to comply from launch. | Providers of generative AI systems used in the EU, wherever they're based. | Settle per product whether you're the provider or only the deployer. Providers: add machine-readable marking and a detection route before 2 December. The Commission's voluntary Code of Practice is the nearest yardstick. Fines reach 15 million euros or 3% of turnover. | Confirmed | Commission FAQ on Article 50 Regulation (EU) 2026/1744 | EU AI Act Article 50 | |
OpenAI shuts down o3-2025-04-16, o3-pro-2025-06-10, gpt-5-2025-08-07, gpt-5-mini-2025-08-07, gpt-5-nano-2025-08-07 and gpt-5-pro-2025-10-06 in the API. | API users on those pinned snapshots. ChatGPT is a separate clock: o3 already left it on 26 August. | Replace with gpt-5.6-sol (add reasoning.mode: pro for the two pro snapshots), gpt-5.6-terra for mini and gpt-5.6-luna for nano. Budget an eval rerun, and test in the API: the ChatGPT Sol is tuned differently. | Confirmed | OpenAI deprecations | o3 retirement | |
| The introductory price of Gemini 3.7 Flash and 3.8 Flash ends: $0.75 in and $3.75 out per million tokens through 31 December, then $1.50 and $7.50 from 1 January 2027. | Teams standardising agents on Flash. | Budget at $1.50 and $7.50 now; anything cheaper until then is a bonus. 3.8 Flash can spend more tokens per task, so measure on your own traces. thinking_level: minimal errors on 3.8. | Confirmed | Gemini API pricing Google: Introducing Gemini 3.7 Flash | Gemini 3.7 Flash Gemini 3.8 Flash | |
| End of 2026 | Ramp says Router routing is free through 2026 and gives no price for after. Ramp doesn't name an exact day. New accounts got $26 of model credit. | Teams that put Router in front of OpenAI, Anthropic and xAI calls. | Mirror a slice of traffic in shadow mode before committing. Set retention yourself before the first production call (TechCrunch reports a one year default). Line up a fallback gateway before January. | Confirmed Ramp; no exact day, later price unpublished | Ramp: Router launch router.com | Ramp Router |
| MTurk transaction history stops being available. | Requesters who need an audit trail for studies or finance. | Export assignment-level records and worker qualification data well before the date. Don't treat a closed portal as an archive. | Confirmed | MTurk closure notice and FAQ | MTurk wind-down | |
| 28 Jul 2027 (earliest) | Earliest date MCP can remove Roots, Sampling, Logging and OAuth Dynamic Client Registration, all deprecated in 2026-07-28. Removal lands in the first spec revision released on or after that date, at the maintainers' discretion. | MCP servers and clients that use those features. | Sampling: call provider APIs directly. Roots: pass paths in tool parameters or resource URIs. Logging: stderr or OpenTelemetry. DCR: Client ID Metadata Documents. Move HTTP+SSE first: its clock is three months after the lifecycle SEP reaches Final. | Confirmed MCP; the date is a floor | MCP deprecated features registry MCP feature lifecycle policy | MCP 2026-07-28 |
| 22 Sep 2027 (earliest) | Anthropic's retirement floors: claude-opus-5-5 not sooner than 22 September 2027, claude-sonnet-5-5 28 September 2027, claude-haiku-5-5 7 October 2027. | Teams that pin these models. | You can pin them for a year without a retirement forcing a move. A floor isn't a retirement date: Anthropic promises at least 60 days' notice before it retires a public model, so watch the deprecations page. | Confirmed Anthropic; the dates are floors | Anthropic model deprecations | Opus 5.5 breaking changes Sonnet 5.5 breaking changes |
| No date yet | Gemini 4 Argon lists at $2 in and $10 out per million tokens, rising to $4 and $20 after an introductory period of unstated length. No API access date, and the pricing page doesn't list Argon yet. | Teams planning around Argon. | Plan on $4 and $20. Write down your current cost per finished task now, so you have a number to beat when you can finally call it. | No date yet | Google: Gemini 4 Argon Gemini API pricing | Gemini 4 Argon pricing |
| No date yet | The old deepseek-v4-flash names route to V4.1-Flash temporarily, and V4 Pro stays with further notice of any change. No dates for either. The launch post still says Pro reroutes from 14 September. | Anyone on the old Flash names or on V4 Pro. | Move to deepseek-flash now and rerun evals, since the model behind the old names already changed on 10 September. For Pro workloads, watch the change log and keep a Flash fallback tested. | No date yet | DeepSeek API change log DeepSeek V4.1-Flash launch post | DeepSeek V4.1-Flash |
| No date yet | No major host has published when it stops offering the MCP 2025-11-25 revision. Anthropic says support for the new revision is rolling out across Claude products and names no sunset for the old one. | MCP server maintainers. | Serve both revisions until a host prints a date. Meanwhile add the Mcp-Method and Mcp-Name headers on Streamable HTTP POSTs and fix code that matches the old not-found code -32002. | No date yet | MCP 2026-07-28 changelog Anthropic: MCP 2026-07-28 in Claude | MCP 2026-07-28 |
| Later this fall | Anthropic's Enterprise Frontier Safeguards, which keeps monitoring data in your own cloud, rolls out in phases later this fall. Until then, eligible customers can get zero data retention on Fable 5 and 5.1 by grant. | Security and procurement teams that stalled on 30 day retention. | Ask Anthropic whether you qualify for the interim zero data retention grant. Don't plan around a rollout date, because there isn't one. | No date yet | Anthropic: Enterprise Frontier Safeguards | Enterprise Frontier Safeguards |
We'd start with the two that hurt in different ways. On 23 October the OpenAI snapshots start returning errors, and you'll hear about it. On 2 November Mullvad's resolvers go dark and plenty of machines fall back to whatever DHCP handed them, usually your internet provider's resolver. Nobody hears anything. The 31 December price cliff can wait until you write next year's budget. Write it at double. The date I trust least is 21 November: OpenAI printed it as a floor and published no price for after, so I might be wrong that anything changes that day.
Already happened
Most recent first. This isn't history for its own sake: if you haven't made the change yet, the fix is the same fix.
| Date | What happens | Who is affected | What to do | Status | Source | Our write-up |
|---|---|---|---|---|---|---|
Anthropic launches Claude Haiku 5.5. Code written for Haiku 4.5 can break: manual budget_tokens returns a 400, adaptive thinking is on by default, and the same text counts as more tokens. | Teams on claude-haiku-4-5. | Drop budget_tokens, set effort explicitly and re-measure token use before swapping. Haiku 4.5 isn't deprecated yet; Anthropic's floor for it is 15 October 2026, with 60 days' notice promised. | Confirmed | Claude API release notes Anthropic model deprecations | None yet | |
| Amazon Mechanical Turk closed and HIT submission ended. The MTurk worker type left SageMaker Ground Truth and Augmented AI. | Requesters, and labelling jobs that used the MTurk workforce. | Point Ground Truth jobs at another workforce. Use the window to 30 October to settle pending work (see Coming up). | Confirmed | MTurk closure notice and FAQ | MTurk wind-down | |
Sonnet 5.5 ships at $2 and $10 per million tokens with five breaking changes: thinking can't be disabled, forced tool_choice fails, computer_20251124 is refused (Claude API, Google Cloud), thinking blocks are model-bound, older advisors are rejected. | Anyone moving from Sonnet 5 or earlier. | Use thinking.type = between_tools (low, medium or high effort only). Use auto plus strict tools instead of forced calls. Switch to computer_toolset_20260801. Test your UI: between-tool notes now arrive as empty thinking blocks. | Confirmed | Anthropic: What's new in Sonnet 5.5 | Sonnet 5.5 breaking changes | |
Opus 5.5 ships at $4 and $20 per million tokens with four breaking changes: thinking can't be disabled, forced tool_choice fails, computer_20251124 is refused (Claude API, Google Cloud), thinking blocks are model-bound. Default effort drops from high to medium. | Opus 5 callers. | Remove thinking: disabled and budget_tokens. Move forced tool calls to auto plus strict tools or structured outputs. Set effort explicitly. Check progress-message UIs, which go silent by default. | Confirmed | Anthropic: What's new in Opus 5.5 | Opus 5.5 breaking changes | |
AWS says it can't restore resources and data hosted only in Bahrain (me-south-1) or in one UAE zone (mec1-az2). The damage spanned several Availability Zones. Next Bahrain update: early 2027. | Anyone who kept their only copy in those places. | Confirm a copy lives in another Region. AWS says the damage exceeded what multi-AZ is built to survive, so multi-AZ isn't a disaster recovery plan. Test a restore into a second Region. | Confirmed | AWS status feed The National | AWS Bahrain region | |
| iOS 27, iPadOS 27 and macOS 27 ship. Siri AI is an English beta; French, Japanese, Korean, Portuguese and Spanish are due in October. Not offered on iPhone, iPad or Watch in the EU, or in China. | Managed Apple fleets, EU and non-English users. | Slot the update into your normal ring. Tell EU and non-English users what they won't get on iPhone. Treat Siri AI as a beta in any support promise. | Confirmed | Apple newsroom: software updates available Apple: Siri AI and the DMA | iOS 27 and Siri AI | |
| Claude Code's 50% weekly-limit promotion ended on 13 September at 11:59 PM PT, not on 19 August. From 14 September weekly limits stay 25% above their pre-promotion level. The 5 hour window never changed. | Pro, Max, Team and seat-based Enterprise users. | Read /usage for your real ceiling instead of a published hour range. Size seats on what the meter shows after 14 September. | Confirmed | Anthropic: Claude Code weekly limits promotion | None yet | |
DeepSeek ships V4.1-Flash as deepseek-flash: $0.15 in and $0.60 out per million tokens off-peak, double at peak. Old Flash names route to it temporarily. Peak and off-peak billing began on 16 August. | Anyone on the old names, or on V4 Pro. | Change the model string to deepseek-flash and rerun evals. Budget the peak hours (01:00 to 04:00 and 06:00 to 10:00 UTC on weekdays) at double. | Confirmed | DeepSeek V4.1-Flash launch post DeepSeek models and pricing | DeepSeek V4.1-Flash | |
| The 50% launch discount on GLM-5.3-Flash via Z.ai on OpenRouter ended. Rates doubled from $0.075 in and $0.25 out to $0.15 and $0.50 per million tokens. | Agent loops budgeted on the discounted rate. | Re-budget at $0.15 and $0.50 per million tokens. | Reported Banner date from OpenRouter via our write-up; today's price read on OpenRouter's models API | OpenRouter: GLM-5.3-Flash | GLM-5.3-Flash | |
| Google starts removing Assistant from Android phones and tablets, and from paired Wear OS watches, headphones and projected Android Auto. The rollout takes weeks and can't be reversed. Cars with Google built in keep it. | Android users, shared-handset fleets, anyone with Assistant routines. | Test your daily voice commands and routines on Gemini now. Move anything that matters to a non-assistant backup: a Clock alarm, a Home automation, an app. No fleet deferral has been published. | Reported Google's email, reproduced by Droid Life and 9to5Google; no Google support page states the date | Droid Life 9to5Google | Google Assistant shutdown | |
Gemini 3.8 Flash ships at the same introductory price, so the 31 December cliff doesn't move. thinking_level: minimal isn't supported and returns an error. | Services that pin minimal thinking for latency. | Pick low, medium or high before you change the model string, and test on your own traces: Google says 3.8 Flash can use more tokens at higher effort. | Confirmed | Google: What's new in Gemini 3.8 Flash | Gemini 3.8 Flash | |
Claude Fable 5.1 breaks forced tool use (tool_choice any or tool returns 400) and errors when you edit history before a thinking block, enforced for accounts created from 31 August. Cache reads drop to $0.25 per million tokens. | Callers of claude-fable-5 who swap the model ID. | Move structured output to auto plus strict tool use or structured outputs. Keep message history append-only. Re-tune effort. Zero data retention still needs Anthropic's authorisation on Fable-class models. | Confirmed | Anthropic: What's new in Fable 5.1 | Fable 5.1 breaking changes | |
| Google details Android 17 network changes. Encrypted Client Hello hides the SNI for apps targeting API 37 when their networking library supports it. Certificate Transparency is on by default. Local network access needs permission. | Network teams that filter or bill on SNI, and Android app developers. | Enforce hostname policy at the resolver, and check what your rules do when the name is missing. App teams: only OkHttp 5.5.0 with explicit DNS config does ECH today; HttpEngine waits for Android 17 QPR2 and WebView has no date. | Confirmed | Google: Android 17 network security Android Developers: Encrypted Client Hello | Android 17 and ECH | |
| o3 left the ChatGPT model picker after a 90 day window. The API runs on a separate clock and shuts down o3 on 11 December (see Coming up). | Plus, Pro and Team users, and API teams that read ChatGPT news as API news. | Export long o3 threads you rely on. Don't prototype o3 replacements in ChatGPT: its Sol is tuned differently from the API's gpt-5.6-sol. | Reported ChatGPT date via Android Authority and gHacks; OpenAI's help centre returned 403 to us | Android Authority OpenAI deprecations | o3 retirement | |
| Broadcom's public downloads for the VMware Virtual Disk Development Kit (VDDK 8.0 and 9.0) started returning 404 with no announcement. Broadcom tells customers it's no longer available and points to technology alliance partners. | Migration and backup tooling that links the VDDK. | Keep and checksum the copies you already hold; the licence doesn't let you redistribute them. Check your tool's fallbacks: storage-assisted migration, an agent in the guest, an OVF export, or an importer that doesn't need the VDDK. | Reported ShapeBlue and Platform9 checked the 404s; Broadcom's words via Platform9 | ShapeBlue Platform9 | Broadcom VDDK | |
Bun 1.4 ships, the first stable Rust build. Bun's own upgrade notes list five changes: Node.js 26 compatibility (module version 147, res.writeHeader() removed), an isolated linker for new monorepos, no .env loading when run as node, YAML 1.2 parsing, strict TOML. | Projects moving to Bun 1.4, and CI that reads workflow YAML. | Grep tooling that reads on: from workflow files: it now parses as the string on and fails quietly. Check native addons ship a build for module version 147. Pass --env-file where you relied on implicit .env. | Confirmed | Bun 1.4 release post | Bun 1.4 | |
| Microsoft retires Group Chat, Podcasts and consumer Deep Research in the Copilot app, starting 18 August. Group chat content isn't carried forward and podcasts become unavailable. | Consumer Copilot users, and admins fielding questions about the merged app. | Download podcasts from the podcast library and any group-chat media you want before your account updates; Microsoft says accounts update gradually. | Confirmed | Microsoft: changes to the Copilot app | None yet | |
Z.ai's GLM-5.3 no longer supports thinking.type: disabled; the request fails. Announced 14 August, listed in the docs on 18 August. The weights reached Hugging Face by 28 August. | GLM-5.2 users who turned thinking off for latency. | Set thinking to enabled and reasoning_effort to low, then change the model ID, in that order. Retest latency: low-effort thinking isn't the old non-thinking mode. | Confirmed | Z.ai docs: GLM-5.3 Hugging Face: zai-org/GLM-5.3 | GLM-5.3 thinking change | |
| Auto mode becomes the default for new Claude Code sessions on Pro, Max and Team. Since v2.1.283 (25 September) it's the starting mode for interactive terminal and VS Code sessions on every plan and provider. | Teams on the API, Bedrock, Google Cloud, Foundry and Enterprise. | Decide on purpose: set permissions.defaultMode, or block auto mode with permissions.disableAutoMode in managed settings. Classifier calls still count toward token usage on Enterprise and API accounts. | Confirmed | Anthropic: auto mode default Claude Code docs: permission modes | None yet | |
xAI's grok-voice-latest alias moves to Grok Voice Think Fast 2.0 at $0.08 a minute, about 60% above the 1.0 rate our write-up reports. | Voice agents that connect with the alias. | Pin grok-voice-think-fast-1.0 if you need the old model, and compare call volume against the new rate. xAI's docs no longer list 1.0 and give no retirement date, so test that the pin still resolves. | Confirmed xAI confirms the alias move and the $0.08; the old $0.05 comes from a third party | xAI: Grok Voice Think Fast 2 xAI docs: models | grok-voice-latest alias | |
| Article 50 of the EU AI Act applies. Providers must tell users they're talking to an AI and machine-mark generative output. Deployers must label deepfakes and AI text on public-interest matters. | Anyone shipping chatbots or generative features to EU users. | Settle provider or deployer status per product. Show the chatbot disclosure before the first message. Systems on the market before 2 August have until 2 December for machine marking (see Coming up). | Confirmed | Commission FAQ on Article 50 Regulation (EU) 2026/1744 | EU AI Act Article 50 | |
MCP 2026-07-28 removes sessions and Mcp-Session-Id, the initialize handshake, ping and SSE resumability, and moves not-found errors to -32602. Roots, Sampling and Logging are deprecated, not removed. | MCP server and client maintainers. | Nothing switches off yet: servers on 2025-11-25 keep negotiating. Plan cross-call state as server-minted handles, implement server/discover, send Mcp-Method and Mcp-Name on Streamable HTTP POSTs, and fix code that matches -32002. | Confirmed | MCP 2026-07-28 changelog MCP release post | MCP 2026-07-28 | |
deepseek-chat and deepseek-reasoner were retired after 15:59 UTC. Both already routed to deepseek-v4-flash. | Anyone still sending the old names. | Use deepseek-flash, the name since 10 September. deepseek-reasoner never mapped to Pro: if hard problems leaned on it, test deepseek-v4-pro on purpose. | Confirmed | DeepSeek V4 preview and API notice DeepSeek API change log | DeepSeek alias retirement | |
| TypeScript 7.0 ships with the compiler rewritten in Go. Microsoft measures full builds 8x to 12x faster on its codebases. There is no stable programmatic API yet. | Projects on Astro, Svelte, Vue or other tools that call the compiler as a library. | Use 7.0 for plain tsc in CI. Keep editor and framework tooling on 6.x with typescript@npm:@typescript/typescript6, which gives you tsc6. Microsoft expects 7.1 to bring a new API, with no date. | Confirmed | Announcing TypeScript 7.0 | TypeScript 7.0 | |
| 24 and 27 Jun 2026 | Microsoft's KEK CA 2011 expired on 24 June, and the UEFI CA 2011, which signs third-party boot loaders such as the Linux shim, on 27 June. Nothing that already booted stopped booting. | Linux and dual-boot machines, and new installs. | Run sudo fwupdmgr refresh then sudo fwupdmgr update and reboot, once your vendor has published the key update. Don't delete the 2011 certificate, and don't switch Secure Boot off. | Confirmed | Microsoft: Secure Boot certificate expiration | Secure Boot expiry, the Linux fix |
A2A v1.0.0 ships and breaks v0.3.0: one unified Part type, enums renamed (TASK_STATE_COMPLETED, ROLE_USER), Agent Card url and protocolVersion replaced by supportedInterfaces, OAuth implicit and password flows removed. | Anyone still emitting or parsing v0.3.0 messages or Agent Cards. | Grep Agent Card generation for a top-level url: that's v0.3.0. Sweep the enum strings. Emit both card formats while you migrate. A2A publishes no v0.3 sunset date, so your counterparties set the deadline. | Confirmed | A2A: what's new in v1.0 A2A releases on GitHub | A2A v1.0 migration |
One pattern repeats in the Anthropic rows. Forced tool_choice broke on Fable 5.1, then again on Opus 5.5 and Sonnet 5.5 six days apart, and switching thinking off broke on both of those. Haiku 5.5 now rejects budget_tokens. If you maintain a wrapper around these APIs, fix it once in the wrapper instead of per model, and keep the model-specific rules in one table you can diff.
What changed since
- 7 October 2026. First version of this page. Two things moved the same day: Anthropic launched Haiku 5.5, which rejects manual
budget_tokens, and cut Sonnet 5.5 cache reads from $0.20 to $0.10 per million tokens. Google's deprecations page also now lists 17 November for the Gemini Live preview, a date our 16 September article on Gemini 3.8 Live couldn't find. - 30 September 2026. Anthropic deprecated
claude-sonnet-4-5-20250929, retiring on 30 November with Sonnet 5.5 as the named replacement. MTurk closed the same day. - 28 September 2026. Sonnet 5.5 shipped six days after Opus 5.5, with the same forced
tool_choicebreak and a renamed way to switch thinking off. - 25 September 2026. Claude Code v2.1.283 made auto mode the starting permission mode on every plan and provider for interactive terminal and VS Code sessions. It had been Pro, Max and Team only.
- 22 September 2026. OpenAI filled in the GPT-6 family. GPT-6 Sol lists at $2 and $10, half the $4 and $20 promotional rate on GPT-5.6 Sol, whose pricing footnote guarantees that rate through 21 November and says nothing after.
- 15 September 2026. AWS said it can't restore data hosted only in Bahrain, or in one UAE zone. It promised the next Bahrain update for early 2027.
- 14 September 2026. iOS 27 shipped on the date Apple had named. Claude Code's weekly-limit promotion had ended the day before, on 13 September rather than 19 August, and left weekly limits 25% above their pre-promotion level.
- 11 September 2026 (reported). DeepSeek reversed the V4 Pro retirement a day after announcing it. Outlets date the reversal to the 11th; the change log folds it into the 10 September entry, and the launch post still carries the old routing sentence.
- 3 September 2026. Mullvad announced it would shut its public encrypted DNS on 2 November and sponsor Quad9 instead.
- 25 August 2026 (reported). CNBC reported Amazon's decision to close MTurk on 30 September, five weeks out, with approvals running to 30 October. The closure notice on mturk.com carries the dates themselves.
- 24 July 2026. The EU's Digital Omnibus on AI, Regulation (EU) 2026/1744, was published in the Official Journal and took effect on 27 July. That makes 2 December 2026 settled law, not a draft, for older generative systems that still owe machine marking.
Questions, and where this comes from
Which dates on this page are firm?
The ones marked Confirmed come from the vendor's own page, read on 7 October. Two kinds need care. Floors, like OpenAI's 21 November pricing note and Anthropic's "not sooner than" dates, promise nothing on the day itself. Reported rows rest on the press or a third party, and the cell says who. If a date moves, the vendor's page wins and the change goes in the log above.
What should I fix first?
Start with the two that fail differently: the OpenAI snapshots on 23 October and the Mullvad addresses on 2 November. Before you cut DNS over, check what a resolver actually answers with our DNS Lookup tool. Then grep for pinned model strings, because the 11 December list is longer than the o3 headline suggests: four gpt-5 snapshots go with it.
Why don't all rows link to one of your articles?
Some rows come from vendor deprecation pages we haven't written up yet. Others rest on a source that has moved on since we published, so we link the source and nothing else. A calendar that sends you to a stale number is worse than a calendar with a gap.
Do the big vendors give the same notice?
No. Anthropic promises at least 60 days' notice before it retires a public model, and publishes "not sooner than" floors. OpenAI lists a shutdown date and a replacement per entry, and gave o3 in the API 183 days. Google lists a date for some models and "No shutdown date announced" for others, gemini-3.8-flash among them. Silence isn't a promise, so don't read it as safety.
Sources
Vendor pages read on 7 October 2026:
- Microsoft, Secure Boot certificate expiration and next-generation code signing
- Amazon, MTurk closure notice and FAQ
- Mullvad, shutdown announcement and Quad9 FAQ
- OpenAI, deprecations and API pricing
- Anthropic, model deprecations, API release notes and Claude Code permission modes
- Google, Gemini API deprecations and pricing
- European Commission, Article 50 FAQ and Regulation (EU) 2026/1744
- MCP, deprecated features registry and lifecycle policy
- DeepSeek, API change log
- Ramp, Router launch
- AWS, status feed
Every other item links its exact source in the tables above.