AI model retirements, 19 to 30 November 2026: Azure o3, Agent Builder, Sonnet 4.5

Summary. Three vendors retire production AI services in the second half of November 2026, and the dates do not line up. On 19 November, Microsoft Foundry retires seven Azure OpenAI reasoning models: o1, o1-pro, o3, o3-deep-research and o3-pro (replacement gpt-5.6-sol) and o3-mini and o4-mini (replacement gpt-5.6-terra). Retired deployments return 410 Gone, Provisioned deployments are never upgraded automatically, and Microsoft does not extend retirement dates. An automatic upgrade also raises the rate: Azure lists gpt-5.6-sol at $4 input and $20 output per million tokens, against $2 and $8 for o3. On 30 November, OpenAI shuts Agent Builder, the Evals dashboard and API, and /v1/prompts. The same day Anthropic retires Claude Sonnet 4.5 on the Claude API, Claude Platform on AWS, Google Cloud and Microsoft Foundry, and removes fast mode from Claude Opus 4.8. Amazon Bedrock keeps Sonnet 4.5 until 8 April 2027. OpenAI's own API is no fallback for most of the Azure models. Sonnet 5.5 costs $2 and $10 per million tokens against $3 and $15, but counts about 30% more tokens and thinks by default. Start with the 19 November items now: from 15 November, that leaves four days.

About this article. This is for platform, ML and engineering leads who run models on Azure OpenAI, the OpenAI API, the Claude API, Google Cloud or Amazon Bedrock. We checked every date, replacement and price on 15 November 2026 against the vendors' own retirement schedules, deprecation tables, migration guides, release notes and pricing pages from Microsoft, OpenAI, Anthropic, Google, AWS and GitHub. Vendors move these dates (Microsoft moved two Claude dates this week), so re-check before you commit; no vendor paid for inclusion. Researched with AI assistance and reviewed by our editorial team.

The calendar: what stops between 15 and 30 November 2026

Dates come from Microsoft's Foundry retirement schedule, OpenAI's deprecations page, Anthropic's model deprecations page and release notes, Google's partner model deprecations and Agent Platform model versions pages, and the Bedrock legacy model lifecycle page.

The codex-mini date is today. Two other dates have moved. Google's Gemini API deprecations page now gives 18 December 2026 as the earliest shutdown for its preview text-to-speech models and no date for gemini-3.1-flash-live-preview, where a copy archived on 29 October listed 17 November. Foundry's Claude Opus 4.5 and Haiku 4.5 dates also moved, as covered below.

Just after the window, the promotional GPT-5.6 Sol price on OpenAI's API ($4 input and $20 output per million tokens) ends on 30 November. On 11 December OpenAI shuts gpt-5, gpt-5-mini, gpt-5-nano, gpt-5-pro, o3 and o3-pro on its own API. Its deprecations page now names gpt-6.1-sol as the replacement for gpt-5 and o3 (and, in pro reasoning mode through the Responses API, for gpt-5-pro and o3-pro), gpt-6.1-terra for gpt-5-mini and gpt-6-luna for gpt-5-nano. Our GPT-5 and o3 December shutdown plan covers that move.

19 November: seven Azure OpenAI reasoning models retire

Microsoft's lifecycle policy defines retired as removed from service: every inference request returns 410 Gone, and neither new nor existing deployments work. Its FAQ says retirement dates cannot be extended.

What happens to your deployment depends on its type.

• Global Standard, Data Zone Standard and Standard. Microsoft upgrades these automatically on a rolling, region-by-region basis, and the schedule's Replacement column is the upgrade target where supported. The deployment's versionUpgradeOption decides the outcome: OnceCurrentVersionExpired upgrades at retirement, NoAutoUpgrade simply stops working, and an unset value counts as OnceCurrentVersionExpired.

• Provisioned. Never upgraded automatically. You either migrate in place, with Azure moving traffic over 20 to 30 minutes without downtime, or create a side-by-side deployment, test it and switch. Quota for the target model must exist first.

• Batch. Side by side only: deploy the new model, resubmit jobs, then retire the old deployment.

• Fine-tuned o4-mini. A separate schedule. The retirement schedule lists deployment retirement on 19 November 2027, while the lifecycle policy page, last updated in July, still shows 16 October 2027; plan for the earlier date.

An automatic upgrade swaps in a different model with different reasoning behaviour, output length and price. The Azure pricing page's rate data, read on 15 November, gives these Global deployment rates in US dollars per million input and output tokens (short-context tier for the replacements).

Cached input moves the other way: $0.50 for o3 against $0.40 for gpt-5.6-sol, and $0.275 for o4-mini against $0.20 for gpt-5.6-terra. Data Zone rates are higher. The Azure page does not mark the gpt-5.6-sol rate as promotional, while on OpenAI's own API the same $4 and $20 is a promotion that ends on 30 November. Confirm the rate on your own price sheet before the upgrade lands.

There is no OpenAI fallback for most of these. OpenAI shut o1, o1-pro, o3-mini and o4-mini on its own API on 23 October 2026, and o3-deep-research on 23 July. o3 and o3-pro survive there only until 11 December. A failover that sends Azure traffic to OpenAI buys about three weeks for o3 and nothing for o4-mini.

Teams on Azure GPT-5 have the next deadline on 9 February 2027; our Azure OpenAI GPT-5 retirement checklist covers it.

30 November: OpenAI Agent Builder, Evals and reusable prompts

OpenAI's feature deprecation table lists three surfaces with the same shutdown date and a notice date of 3 June 2026.

Evals are already read-only, so no new run can start on OpenAI's platform. OpenAI's migration cookbook describes a manual path, not an export: recreate each eval's prompts, provider, test cases and assertions in a promptfooconfig.yaml file, rebuild graders as Promptfoo assertions, and validate any LLM-as-judge grader before it gates a release. OpenAI's pages do not say whether run history survives the shutdown, so copy the results you cite in release gates now.

The Agent Builder migration guide exports a workflow as Agents SDK code in TypeScript or Python from the Code menu. It does not convert the workflow graph, and connected apps, authentication and permissions need separate work. ChatKit stays, but an embed that points at a hosted Agent Builder workflow ID works only during the transition window; OpenAI's ChatKit guide sends new and migrating apps to its advanced integration. Our earlier guide to the Agent Builder, Evals and prompts shutdown covers the background.

One cost trap: when prompts move from /v1/prompts into code, keep static content first and variables last. OpenAI's migration guide notes that cache hits depend on exact prefix matches, so a customer name near the top of a long system prompt turns every request into a cache miss.

30 November: Claude Sonnet 4.5 leaves four platforms, but not Bedrock

Anthropic notified developers on 30 September 2026 that claude-sonnet-4-5-20250929 retires on 30 November, with claude-sonnet-5-5 as the replacement. Its dates apply to Anthropic-operated platforms: the Claude API, Claude Platform on AWS, Claude Models on Google Cloud (Anthropic managed) and Microsoft Foundry. Partner-operated platforms, Google-managed Claude on Google Cloud and Amazon Bedrock, set their own. The table draws on Anthropic's page, Google's partner model deprecations page, the Foundry schedule and the Bedrock lifecycle page.

Bedrock's longer runway has conditions. New customers cannot use a Legacy model, existing customers may lose access after 15 days of inactivity, no new Provisioned Throughput can be created, and during public extended access the provider sets higher prices. Our Bedrock Sonnet 4.5 and Opus 4.1 retirement plan covers that path.

Two smaller Anthropic items share the window. The 30 October release note deprecates fast mode for Claude Opus 4.8, with removal on 30 November: after that, requests to claude-opus-4-8 with speed set to fast return an error, and the migration target is Claude Opus 5.5. And Microsoft has moved two Claude dates. On 15 November its schedule lists claude-opus-4-5 and claude-haiku-4-5 retiring on Foundry on 22 February 2027, with claude-opus-5-5 and claude-haiku-5-5 as replacements; a copy archived on 13 November still showed 24 and 15 November. Anthropic's own table lists both models as Active.

What breaks moving from Sonnet 4.5 to Sonnet 5.5

Anthropic's migration guide lists these breaking changes from 4.5, each a 400 error or a change in behaviour:

• Prefill. A prefilled final assistant turn returns 400. Replace it with structured outputs, tools with enum fields, or system-prompt instructions.

• Sampling parameters. Remove temperature, top_p and top_k.

• Thinking on by default. A request with no thinking field runs adaptive thinking, billed as output tokens. To turn up-front thinking off, set thinking.type to between_tools at high effort or below; disabled returns 400.

• Thinking budgets. budget_tokens returns 400. Set output_config.effort instead and test two or three levels.

• Forced tool use. A tool_choice of any or tool returns 400. Use auto with strict tools.

• Response parsing. Responses can start with thinking blocks, so code that reads content[0].text breaks.

• Structured outputs. output_format returns 400 without its old beta header; move it to output_config.format.

• Tokens. The same text produces about 30% more tokens than on Sonnet 4.5. Recount with token counting and revisit max_tokens.

Worked cost example: Sonnet 4.5 to Sonnet 5.5

Assume a software company runs a customer-facing assistant on the Claude API. Each month it sends 600 million input tokens, half of them prompt cache hits, and receives 60 million output tokens. Prices are Anthropic's list prices in US dollars per million tokens: Sonnet 4.5 at $3 input, $0.30 cache hits and $15 output; Sonnet 5.5 at $2, $0.10 and $10. Cache writes are left out to keep the arithmetic readable.

The Sonnet 5.5 columns add 30% to every token count for the tokenizer change (300M × 1.3 = 390M; 60M × 1.3 = 78M; 78M × 1.5 = 117M). With thinking held down through between_tools or a low effort level, the bill falls by $291, or 15.4%. If the team swaps the model string and lets adaptive thinking run, it rises by $99, or 5.2%. The 50% thinking overhead is our assumption; measure yours on a day of shadow traffic. The migration guide also notes that the minimum cacheable prompt drops from 1,024 tokens on Sonnet 4.5 to 512 on Sonnet 5.5, which can raise the hit rate on short prompts.

Date-by-date migration checklist

15 and 16 November

• List every Azure OpenAI deployment (command below) and flag o-series, codex-mini, gpt-chat-latest preview and Llama-4-Maverick deployments.

• Read each deployment's versionUpgradeOption and decide whether you want Microsoft's upgrade or a manual move.

• Search code and config for claude-sonnet-4-5, pmpt_, /v1/prompts and Evals API calls, and open one ticket per hit.

17 and 18 November

• Deploy gpt-5.6-sol or gpt-5.6-terra side by side for every Provisioned and Batch o-series deployment, confirm quota and run your evaluation set.

• Set reasoning effort and output limits explicitly on the new deployments, then compare cost per task.

• Cut Provisioned traffic over by the end of 18 November. Nothing upgrades Provisioned deployments, and time zones make the 19th too late.

• Confirm no failover rule points o-series traffic at OpenAI's API, where most of these models are already gone.

• Move any gemini-3.6-flash traffic on Google's Agent Platform to gemini-3.8-flash.

19 to 23 November

• Alert on 410 Gone in every region; rollout is not simultaneous.

• Check usage and invoices after any auto-upgrade; the replacement bills at its own rate.

• Copy every OpenAI eval result you rely on and store it with the release it gated.

24 to 27 November

• Bedrock users in us-east-1 or ca-central-1: move Jamba 1.5 and Claude 3 Haiku or Sonnet traffic before 26 and 27 November.

• Finish Sonnet 5.5 testing: prefill, sampling parameters, thinking defaults, forced tool use and response parsing.

By 30 November

• Switch production to claude-sonnet-5-5 on every Anthropic-operated platform and on Google-managed Claude. Keep Bedrock on its own plan.

• Remove fast mode from Claude Opus 4.8 calls or move them to Claude Opus 5.5.

• Ship Agent Builder replacements on the Agents SDK or Workspace Agents, and move any ChatKit embed off its hosted workflow ID.

• Move prompt objects into source control with static text first, and point CI at Promptfoo.

• Move TwelveLabs Marengo Embed v2.7 traffic on Bedrock to an Active model.

Right after

• From 1 December, Gemini API requests to gemini-2.5-pro route to gemini-3.8-flash and are billed at its rates. Our Gemini 4 Hydrogen pricing and migration guide covers the alternatives.

• On 11 December OpenAI shuts gpt-5, its mini, nano and pro variants, o3 and o3-pro; on 14 December GitHub Copilot drops Gemini 3.8 Flash and Gemini 4 Argon, among others.

Find what you are actually running

            # Azure: list deployments with model, version, SKU and upgrade setting
az cognitiveservices account deployment list -g <resource-group> -n <account> \
  --query "[].{name:name, model:properties.model.name, version:properties.model.version, sku:sku.name, upgrade:properties.versionUpgradeOption}" -o table

# OpenAI: each model object can carry an optional shutdown_date field
curl -s https://api.openai.com/v1/models -H "Authorization: Bearer $OPENAI_API_KEY"

# Code: find the strings that stop working this month
grep -rnE "o1-pro|o3-mini|o4-mini|o3-pro|o3-deep-research|codex-mini|claude-sonnet-4-5|pmpt_|/v1/prompts" .
          

The Azure CLI command is documented on Microsoft Learn, and the CLI shows null when versionUpgradeOption was never set, which Microsoft treats as OnceCurrentVersionExpired. Microsoft's lifecycle page also documents a Models API with lifecycleStatus and per-SKU deprecationDate fields. Mind the naming trap it flags: in that API, Deprecating means still serving and Deprecated means retired. OpenAI's list models endpoint carries the optional shutdown_date field. The Usage page in Claude Console exports a CSV broken down by API key and model. On Bedrock, ListFoundationModels returns each model's modelLifecycle per region.

Decision table

How eCorpIT can help

eCorpIT is a Gurugram-based technology organisation, founded in 2021, assessed at CMMI Level 5 and MSME certified, with senior-led engineering teams working across AWS, Microsoft and Google platforms. Our AI and machine learning services team can inventory your Azure, OpenAI, Anthropic, Google Cloud and Bedrock deployments, test each replacement against your own evaluation set and report cost per task before cut-over. Talk to us at /contact-us/.

Last updated: 15 November 2026.

References

1. Model retirement schedule, Microsoft Foundry — Microsoft Learn

1. Foundry Models lifecycle and support policy — Microsoft Learn

1. Azure OpenAI Service pricing — Microsoft Azure

1. Deprecations — OpenAI API docs

1. Model deprecations — Anthropic, Claude Platform docs

1. Claude Platform release notes (28 September, 30 September and 30 October 2026 entries) — Anthropic

1. Model lifecycle (Legacy) — Amazon Bedrock User Guide, AWS

1. Pricing — Anthropic, Claude Platform docs

1. Claude Sonnet 5.5 migration guide — Anthropic, Claude Platform docs

1. Partner model deprecations, Gemini Enterprise Agent Platform — Google Cloud Documentation

1. Model versions and lifecycle, Gemini Enterprise Agent Platform — Google Cloud Documentation

1. Gemini deprecations — Google AI for Developers

1. GPT-5.6 Sol model page — OpenAI API docs

1. Azure OpenAI in Microsoft Foundry Models: working with models — Microsoft Learn

1. Moving from OpenAI Evals to Promptfoo — OpenAI Cookbook

1. Migrate from Agent Builder — OpenAI API docs

1. ChatKit — OpenAI API docs

1. Migrate from prompt objects — OpenAI API docs

1. Upcoming deprecation of selected GitHub Copilot models in December, 13 November 2026 — GitHub Changelog

1. az cognitiveservices account deployment — Microsoft Learn