December 2026 model changes: migration checklist

Last verified: 2026-11-03 against the official pages: Google AI Studio changelog, deprecations and pricing; Anthropic's model deprecations, API release notes and fast mode docs; OpenAI's deprecations; xAI's release notes. All dollar figures are the vendors' own list prices and can change; check the source before acting.

Short answer

Four hard dates between late November and mid-December. 11-17: Gemini 2.5 / 3.1 TTS and Live preview models shut down. 11-30: claude-sonnet-4-5-20250929 retires, and speed: "fast" on claude-opus-4-8 is removed. 12-01: gemini-2.5-pro / gemini-2.5-flash start being served by 3.8 Flash / 3.5 Flash-Lite, and older GPT Image models shut down. 12-11: OpenAI's older gpt-5 and o3 snapshots shut down. The easy one to miss is Gemini 2.5. It does not error. The model name stays the same, a different model answers, billing follows the new model, and old settings such as thinking_budget, temperature and prefilled turns may stop working.

Timeline

Date	Vendor	Change	What happens after	Official replacement
11-02 (in effect)	xAI	grok-imagine-image-quality retired	Served by grok-imagine-image-2.0 with quality: low	grok-imagine-image-2.0
11-17	Google	gemini-2.5-flash-preview-tts, gemini-2.5-pro-preview-tts, gemini-3.1-flash-tts-preview shut down	Shut down	gemini-3.8-flash-tts / gemini-3.8-flash-lite-tts
11-17	Google	gemini-2.5-flash-native-audio-preview-12-2025, gemini-3.1-flash-live-preview shut down	Shut down	gemini-3.8-live
11-30	Anthropic	claude-sonnet-4-5-20250929 retires	Requests fail	claude-sonnet-5-5
11-30	Anthropic	Fast mode on claude-opus-4-8 removed	Requests with speed: "fast" error; standard requests unaffected	Fast mode on claude-opus-5-5
11-30	OpenAI	v1/prompts (reusable prompt objects), Evals, Agent Builder shut down	Shut down	Prompts in code / Promptfoo / Agents SDK
12-01	Google	gemini-2.5-pro → gemini-3.8-flash	No error, model silently replaced, billed at the new model's rates	gemini-3.8-flash
12-01	Google	gemini-2.5-flash → gemini-3.5-flash-lite	No error, model silently replaced, billed at the new model's rates	gemini-3.5-flash-lite
12-01	OpenAI	gpt-image-1-mini, gpt-image-1.5, chatgpt-image-latest shut down	Shut down	gpt-image-2.5-sunburst / gpt-image-2.5-flare
12-11	OpenAI	gpt-5, gpt-5-mini, gpt-5-nano, gpt-5-pro, o3, o3-pro and their dated snapshots shut down	Shut down	gpt-5.6-sol / -terra / -luna

As of 11-03, xAI has announced no text-model changes taking effect in December.

1. Gemini 2.5 Pro / Flash: silently replaced from 12-01

The Google AI Studio changelog of 2026-11-02 says gemini-2.5-pro and gemini-2.5-flash are deprecated, and that from 2026-12-01 every API request to those names is automatically routed to a newer model and billed at that model's rates. The changelog mentions no opt-out.

Model name you send	Model that answers from 12-01	Official advice
gemini-2.5-pro	gemini-3.8-flash	Change the model string to gemini-3.8-flash before 12-01
gemini-2.5-flash	gemini-3.5-flash-lite	Change the model string to gemini-3.5-flash-lite before 12-01
gemini-2.5-flash-lite	Unchanged	Not deprecated; no shutdown date announced

The second row is a tier change, from Flash down to Flash-Lite, and Pro to Flash is not like-for-like either. Code that calls these models by name keeps running untouched, while answer quality, reasoning depth and output style may already have changed.

What happens to the bill (Google list prices, USD per million tokens)

gemini-2.5-pro → gemini-3.8-flash. 3.8 Flash has a lower price that ends on 12-31 and doubles from 2027-01-01:

Item	2.5 Pro (≤200k)	2.5 Pro (>200k)	3.8 Flash (to 12-31)	3.8 Flash (from 2027-01-01)
Input	1.25	2.50	0.75	1.50
Output (incl. thinking)	10.00	15.00	3.75	7.50
Cache read	0.125	0.25	0.075	0.15
Cache storage (per hour)	4.50	4.50	0.50	1.00

• December: input and output are both cheaper than 2.5 Pro.
• From January: for prompts up to 200k, input costs 20% more than 2.5 Pro, and cache reads also cost 20% more. Output is still 25% cheaper. Input-heavy workloads such as RAG and long-document Q&A can see the January bill go up, not down.
• 3.8 Flash has no 200k tier, so very long prompts are cheaper than on 2.5 Pro in both months.

gemini-2.5-flash → gemini-3.5-flash-lite. Text, image and video input is 0.30, output 2.50 and cache reads 0.03, the same as 2.5 Flash. Audio input drops from 1.00 to 0.30.

Search grounding is billed in a different unit. The 2.5 series gets 1,500 free requests per day, then $35 per 1,000 grounded prompts. Gemini 3 and later get 5,000 free per month (shared across all Gemini 3+ models), then $14 per 1,000 search queries, capped at 1,500 per day. Google notes that one request can trigger several search queries and each is billed, so heavy grounding users should recompute in the new unit.

Parameters that break

The changelog names four kinds of breaking change: function calling and thought_signature validation, thinking_level replacing thinking_budget, removed sampling parameters, and removed prefilled model turns. Here is how they line up against the two target models' migration pages:

Old usage	3.8 Flash	3.5 Flash-Lite
thinking_budget	Use thinking_level: low / medium / high, default medium; minimal returns an error	Use thinking_level; default minimal
temperature / top_p / top_k	Migration page says remove; behaviour if sent not documented	Currently ignored; Google says future models will return an error
Last non-empty turn is model (prefill)	No longer supported	Returns 400
candidate_count	Unsupported from Gemini 3	Same
Function calling	On generateContent, every FunctionResponse needs call_id and name; put multimodal assets inside the response payload	Same
Thought signatures	Pass thought_signature back unchanged in multi-turn tool use	Same

Google does not say what happens after the reroute to a gemini-2.5-pro request that still carries thinking_budget or other old parameters: an error, silent ignore, or automatic conversion. Plan for the worst case: change the model string and the parameters yourself before 12-01, rather than relying on the reroute to handle it.

Also, gemini-3.5-flash-lite defaults to almost no thinking (minimal). If an agent relied on 2.5 Flash's default thinking, Google recommends setting thinking_level to medium or high so it does not stop tool use too early.

2. Claude Sonnet 4.5: retires 11-30

Per Anthropic's model deprecations page, claude-sonnet-4-5-20250929 was deprecated on 2026-09-30 and retires on 2026-11-30. The recommended replacement is claude-sonnet-5-5, and requests fail after retirement. The date applies to the Claude API and the other Anthropic-operated platforms. Amazon Bedrock and Google-managed Vertex AI set their own schedules.

Moving from 4.5 to 5.5 takes more than a model-name change. The Sonnet 5.5 migration guide lists these changes:

• Thinking is on by default. Requests with no thinking field still think on 5.5, and thinking tokens are billed as output. Revisit max_tokens. To keep thinking to a minimum, send the lowest thinking setting and keep effort at high or below.
• Forced tool use is rejected. tool_choice of any or tool returns 400: tool_choice: type "tool" and "any" are not supported for this model. Use auto with strict tools instead.
• Remove sampling parameters: temperature, top_p, top_k.
• No assistant prefill. Use structured outputs or system-prompt instructions instead.
• Set output_config.effort explicitly. Move output_format to output_config.format.
• Clean up beta headers. Remove any context-window beta header and interleaved-thinking-2025-05-14. Replace fine-grained-tool-streaming-2025-05-14 with eager_input_streaming.
• Read content blocks by type, and pass thinking blocks back unchanged in tool loops.

3. Opus 4.8 fast mode: removed 11-30

On 2026-10-30 Anthropic deprecated fast mode on Opus 4.8, with removal on 2026-11-30. After that, requests to claude-opus-4-8 that set speed: "fast" return an error.

Only fast mode is affected. claude-opus-4-8 itself keeps working, and the deprecations page puts its retirement at "not sooner than" 2027-05-28. That leaves two options:

1. Stay on Opus 4.8. Drop speed: "fast" and the fast-mode-2026-02-01 beta header and run at standard speed.
2. Move to claude-opus-5-5 and keep fast mode. This is the path Anthropic recommends. Opus 5.5 fast mode left research preview on 10-15 and needs no beta header. Official fast mode pricing is $8 input / $40 output per million tokens on Opus 5.5, versus $10 / $50 on Opus 4.8 and Opus 5.

If you take option 2, migrate Opus 5.5's behaviour changes along with it:

• Thinking cannot be disabled. thinking: {"type": "disabled"} and manual budget_tokens budgets return 400; control depth with effort instead. The default effort is medium, whereas Opus 4.8 defaults to high.
• Forced tool use is rejected. tool_choice of any or tool returns 400.
• More output tokens are possible. Requests that omitted thinking did not think on 4.8 but do think on 5.5, so output tokens per request can go up.
• No Priority Tier on Opus 5.5.
• Lower caching minimum. The minimum cacheable prompt drops from 1,024 to 512 tokens.

Either way, two things are worth knowing:

• Fast and standard speed do not share prompt cache. The first request that falls back from fast to standard misses the cache.
• usage.speed reports the speed actually used, "fast" or "standard", so you can monitor on it.

4. OpenAI: older GPT Image and GPT-5 / o3 snapshots

• 12-01: gpt-image-1-mini, gpt-image-1.5 (gpt-image-1.5-2025-12-16) and chatgpt-image-latest shut down. The replacements are gpt-image-2.5-sunburst and gpt-image-2.5-flare. gpt-image-1 already shut down on 10-23.
• 12-11: the older snapshots shut down, including their dated versions:
  • gpt-5 → gpt-5.6-sol
  • gpt-5-mini → gpt-5.6-terra
  • gpt-5-nano → gpt-5.6-luna
  • o3 → gpt-5.6-sol
  • gpt-5-pro and o3-pro → gpt-5.6-sol with reasoning.mode: pro
• 11-30: v1/prompts and reusable prompt objects, the Evals platform and Agent Builder shut down.

Next after that is the GPT-4o / GPT-4.1 family on 2027-01-14. See GPT-4o shutdown and migration for the full calendar.

5. Extra work if you run or use a gateway

Teams that run their own gateway, or call models through a third-party one, face three extra risks this round:

1. Per-model price tables will drift from the upstream bill. From 12-01, upstream bills gemini-2.5-pro at 3.8 Flash rates, and 3.8 Flash doubles on 2027-01-01. A gateway that bills by model name needs two price updates, on 12-01 and on 01-01, or it will over- or under-charge.
2. Passing the model name straight through means accepting the silent swap on the user's behalf. For gemini-2.5-*, either map explicitly to the new model name and convert parameters on the way through (thinking_budget → thinking_level, strip sampling parameters and prefilled turns), or return a clear error so the caller can decide. Doing nothing and letting the upstream swap the model quietly is the worst option.
3. For Claude, the opposite: do not map silently. Quietly mapping claude-sonnet-4-5 to claude-sonnet-5-5 keeps requests flowing, but thinking turns on by default, forced tool use returns 400 and output tokens grow. That is harder to debug than a plain error. The safer approach is to notify callers before 11-30 and return an explicit "model retired" error afterwards. If your gateway injects speed or the fast-mode beta header on Opus 4.8 requests, remove that before 11-30 too.

Grep checklist

Run these across your code and config (requires ripgrep):

# Gemini: exclude flash-lite and flash-image, which stay available
rg -n 'gemini-2\.5-(pro|flash)\b' | rg -v 'flash-(lite|image)'
rg -n 'thinking_budget|thinkingBudget'

# Claude
rg -n 'claude-sonnet-4-5'
rg -n -i 'speed.{0,4}fast|fast-mode-2026-02-01'

# OpenAI
rg -n 'gpt-image-1\.5|gpt-image-1-mini|chatgpt-image-latest'
rg -n -w -e 'gpt-5' -e 'gpt-5-mini' -e 'gpt-5-nano' -e 'gpt-5-pro' -e 'o3' -e 'o3-pro'

The last command also matches newer names such as gpt-5.6-sol, so review each hit. Anthropic customers can also open the Usage page in Claude Console and click Export to get a CSV broken down by API key and model, which shows which keys still call old models.

Not changing in December

None of these change in December, so no action is needed yet:

• gemini-2.5-flash-lite: Google states it is not deprecated.
• gemini-2.5-flash-image: shuts down on 2027-03-15.
• claude-opus-4-8 itself: only fast mode goes away.
• claude-opus-4-5-20251101 and claude-haiku-4-5-20251001: still Active on the deprecations page. Anthropic commits to at least 60 days' notice before retirement, and nothing had been announced as of 11-03, so the earliest possible date is January 2027.
• GPT-4o / GPT-4.1: 2027-01-14.

Sources

• Google AI Studio: changelog (2026-11-02), deprecations, pricing, Gemini 3.8 migration, Gemini 3.6 / 3.5 Flash-Lite migration
• Anthropic: model deprecations, API release notes (2026-10-30), Fast mode, Opus 5.5 migration guide, Sonnet 5.5 migration guide
• OpenAI: Deprecations
• xAI: Release notes

FAQ

Can I still call gemini-2.5-pro after December 1, 2026?

Yes, but Gemini 2.5 Pro will not be the model answering. From 12-01, every request to gemini-2.5-pro is served by gemini-3.8-flash and billed at 3.8 Flash rates, and gemini-2.5-flash is served by gemini-3.5-flash-lite. The model name stays the same and no error is returned.

Will the Gemini 2.5 reroute cost more?

Not in December; from January it depends on your mix. At list prices, 3.8 Flash is $0.75 input / $3.75 output through 12-31 and $1.50 / $7.50 from 2027-01-01, against $1.25 / $10 for 2.5 Pro on prompts up to 200k. So from January input costs 20% more than 2.5 Pro and output 25% less. 2.5 Flash moving to 3.5 Flash-Lite keeps the same text prices.

When does Claude Sonnet 4.5 retire?

claude-sonnet-4-5-20250929 retires on 2026-11-30, with claude-sonnet-5-5 as the recommended replacement, and requests fail after that date. Bedrock and Google-managed Vertex AI follow their own schedules.

What happens when Opus 4.8 fast mode is removed?

After 11-30, sending speed: "fast" to claude-opus-4-8 returns an error, and requests without speed are unaffected. To keep fast mode, move to claude-opus-5-5, where fast mode is generally available and needs no beta header.

How do I find out whether I use any of these models?

Run the grep checklist above. Anthropic customers can also export a CSV by key and model from the Usage page in Claude Console.