top of page
Search

Claude, Gemini and DeepSeek API model names in 2026: aliases, pinned versions and retired IDs

45 minutes ago
8 min read
Cover showing the title Claude, Gemini and DeepSeek API model names in 2026: aliases, pinned versions and retired IDs. Data Studios original cover

Last verified: November 3, 2026


Claude, Gemini and DeepSeek name their API models in three different ways, so a name that looks the same can point to a model that never changes or to one that is replaced without any edit to your code. On the Claude API, every model from the 4.6 generation onward is a fixed snapshot, even though its name carries no date. On the Gemini API, the stable codes are fixed, while the "latest" aliases move to a new model after two weeks of email notice. On the DeepSeek API, the recommended names are moving pointers by design: on October 17, 2026, DeepSeek moved the name deepseek-flash to a new model version without changing how it is called or what it costs.


Some older names also stop working, or start answering with a different model, within weeks. A Claude Sonnet 4.5 snapshot retires on November 30, 2026; on December 1, 2026 Google will reroute calls to Gemini 2.5 Pro and Gemini 2.5 Flash to newer models; and DeepSeek retired its older chat and reasoner names on July 24, 2026. The tables below give the exact IDs each provider documents today, the retirement dates, and the one request per provider that shows which names your API key can call.


··········


CLAUDE API MODEL IDS IN NOVEMBER 2026.


Anthropic's models overview lists four current models, all with a 1M-token context window, 128K tokens of maximum synchronous output and a June 2026 reliable knowledge cutoff. The Claude API ID and the Claude API alias are identical for these models, because the dateless string is itself the snapshot.


........


Model

Claude API ID

Amazon Bedrock ID

Retirement (Anthropic platforms)

Claude Fable 5.1

claude-fable-5-1

anthropic.claude-fable-5-1

Not sooner than Sep 1, 2027

Claude Opus 5.5

claude-opus-5-5

anthropic.claude-opus-5-5

Not sooner than Sep 22, 2027

Claude Sonnet 5.5

claude-sonnet-5-5

anthropic.claude-sonnet-5-5

Not sooner than Sep 28, 2027

Claude Haiku 5.5

claude-haiku-5-5

anthropic.claude-haiku-5-5

Not sooner than Oct 7, 2027


........


Anthropic's documentation on model IDs addresses a common misreading directly. A dateless ID such as claude-sonnet-4-6 or claude-opus-5-5 does not route to the newest or best version of a model line: from the Claude 4.6 generation onward, the dateless string is the canonical ID of one release and maps to a single fixed snapshot, and Anthropic ships any updated version under a new ID. The format is claude-{name}-{major}[-{minor}], so a major release such as Claude Opus 5 drops the minor segment and becomes claude-opus-5.


Older models follow the earlier convention, claude-{name}-{major}-{minor}-{YYYYMMDD}, for example claude-sonnet-4-5-20250929 and claude-haiku-4-5-20251001. For these models the Claude API also accepts a shorter alias such as claude-sonnet-4-5, which resolves to the most recent dated snapshot of that minor version. Code written in 2025 against claude-sonnet-4-5 is therefore calling claude-sonnet-4-5-20250929, the snapshot that retires on November 30, 2026.


The same model carries a different string on partner clouds. Amazon Bedrock prefixes the ID with anthropic.; Claude Opus 4.6 was the last Bedrock ID with the -v1 suffix (anthropic.claude-opus-4-6-v1), and dated Bedrock IDs end in -v1:0. Google Cloud's Google-managed offering matches the Claude API strings for current models and writes older dated snapshots with an @, as in claude-haiku-4-5@20251001. Microsoft Foundry sends the deployment name, which defaults to the Claude API ID. Anthropic notes that Bedrock and the Google-managed offering set their own lifecycle dates, so the retirement column above applies only to Anthropic-operated platforms. Anthropic also documents that the weights behind an ID never change while the serving infrastructure around them (routing, safety classifiers, sampling logic) can, which explains small behavior shifts on an unchanged ID. Pricing, context and tool details for each model are in our guide to the Claude API models, pricing and limits.


··········


CLAUDE MODEL IDS RETIRING OR ALREADY RETIRED.


Anthropic uses four lifecycle states: active, legacy, deprecated and retired. Requests to a retired model fail, and Anthropic commits to at least 60 days of notice before retiring a publicly released model. The rows below are the ones that affect code still running in late 2026.


........


API model name

State on Nov 3, 2026

Retirement date

Recommended replacement

claude-sonnet-4-5-20250929

Deprecated

November 30, 2026

claude-sonnet-5-5

claude-opus-4-5-20251101

Active

Not sooner than Nov 24, 2026

Not yet assigned

claude-haiku-4-5-20251001

Active

Not sooner than Oct 15, 2026

Not yet assigned

claude-mythos-preview

Deprecated

To be announced

Not listed

claude-opus-4-1-20250805

Retired

August 5, 2026

claude-opus-4-8

claude-sonnet-4-20250514

Retired

June 15, 2026

claude-sonnet-5-5

claude-opus-4-20250514

Retired

June 15, 2026

claude-opus-4-8

claude-3-haiku-20240307

Retired

April 20, 2026

claude-haiku-4-5-20251001


........


The Haiku 4.5 row shows that a "not sooner than" date is only a floor: the commitment passed in October and the model was still listed as active on November 3. Anthropic's deprecation page also suggests a practical audit: in the Claude Console, the Usage page has an Export button that downloads a CSV broken down by API key and model, which shows every place a deprecated ID is still being called. Parameters can be retired too. In API version 2026-10-21 the sampling parameters temperature, top_p and top_k were removed and now return a 400 error, and Opus 4.8 fast mode is on its own schedule, covered in our report on the Opus 4.8 fast mode deprecation.


··········


GEMINI API MODEL CODES: STABLE, PREVIEW, LATEST AND EXPERIMENTAL.


Google's model page describes four kinds of model code, a convention that applies to models released from September 2025 onward. The difference between them decides whether the model behind your string can change.


........


Gemini code type

Example

Can the model behind it change

Notice before change

Stable

gemini-3.6-flash

Usually no

Deprecation schedule

Preview

gemini-3.1-pro-preview

Deprecated, then shut down

At least 2 weeks

Latest

gemini-flash-latest

Yes, hot-swapped at each release

2 weeks by email

Experimental

No example listed

Availability may change

None stated


........


The current stable codes include gemini-4-argon, gemini-3.8-flash, gemini-3.6-flash, gemini-3.5-flash-lite, gemini-3.1-flash-lite and gemini-2.5-flash-lite, plus specialised models such as gemini-3.8-live, gemini-3.8-flash-tts and gemini-embedding-2. Google recommends stable codes for most production applications. Its model page does not say which model gemini-flash-latest currently resolves to. A developer reported on the Google AI Developers Forum in January 2026 that calls made with gemini-flash-latest appeared in the AI Studio dashboard under the specific model name, and a community member replied that billing follows the specific model; Google has not documented this behavior on its model page.


On the REST API the code sits in the URL path, as in models/gemini-3.8-flash:generateContent. The Models resource gives each model a name in the form models/{model} and a baseModelId, and lists the methods each model supports. Google's documentation now labels the generateContent guide as legacy and recommends the Interactions API for new projects, although the generateContent endpoint is still documented and working.


··········


GEMINI 2.5 PRO AND 2.5 FLASH: THE DECEMBER 1 REROUTING AND OTHER SHUTDOWNS.


Google's deprecations page lists the earliest possible shutdown dates. Two upcoming entries behave differently from a normal shutdown: requests to the two Gemini 2.5 models will not fail on December 1, 2026. They will be served by a newer model and billed at that model's rates, as explained in our report on the Gemini 2.5 rerouting.


........


Gemini model code

Shutdown (earliest)

What happens or replacement

gemini-2.5-pro

December 1, 2026

Routed to gemini-3.8-flash, billed at 3.8 Flash rates

gemini-2.5-flash

December 1, 2026

Routed to gemini-3.5-flash-lite, billed at its rates

gemini-2.5-flash-preview-tts, gemini-2.5-pro-preview-tts

November 17, 2026

gemini-3.8-flash-tts or gemini-3.8-flash-lite-tts

gemini-3.1-flash-tts-preview

November 17, 2026

gemini-3.8-flash-tts or gemini-3.8-flash-lite-tts

gemini-3.1-flash-live-preview

November 17, 2026

gemini-3.8-live

gemini-2.5-flash-image

March 15, 2027

gemini-3.1-flash-lite-image

gemini-3.1-flash-lite

May 7, 2027

gemini-3.5-flash-lite


........


Google says Gemini 2.5 Pro and 2.5 Flash are no longer selectable in Google AI Studio, that access to the 2.5 models is limited to past users, and that gemini-2.5-flash-lite is not deprecated. Its page asks developers to review the migration guides for breaking API changes before the rerouting date, since a request that worked on 2.5 Pro can behave differently once a Flash model answers it.


··········


DEEPSEEK API MODEL NAMES: DEEPSEEK-FLASH, DEEPSEEK-V4-PRO AND LEGACY NAMES.


DeepSeek publishes only two model names on its Models and Pricing page. Both offer thinking and non-thinking modes (thinking is the default), a 1M-token context length and up to 384K tokens of output.


........


DeepSeek model name

Served by on Nov 3, 2026

Status

deepseek-flash

DeepSeek-V4.1-Flash-1017

Current, vision supported

deepseek-v4-pro

DeepSeek-V4-Pro-0813

Current, no vision

deepseek-v4-flash

DeepSeek-V4.1-Flash-1017

Legacy name, still accepted

deepseek-v4-flash-vision-exp

DeepSeek-V4.1-Flash-1017

Legacy name, still accepted

deepseek-chat, deepseek-reasoner

None

Retired July 24, 2026, 15:59 UTC


........


The history explains the table. At the V4 launch on April 24, 2026, DeepSeek kept deepseek-chat and deepseek-reasoner as legacy names pointing to the non-thinking and thinking modes of deepseek-v4-flash, and announced that both would be "fully retired and inaccessible" after July 24, 2026. On September 10 it introduced deepseek-flash for V4.1 Flash, retired the V4 Flash and V4 Flash Vision Exp models and routed their old names to V4.1 Flash for compatibility, billed at the Flash price. On October 17 deepseek-flash moved to DeepSeek-V4.1-Flash-1017. DeepSeek's documentation gives no end date for the two legacy V4 Flash names, so code that still sends them works today on borrowed time. Older background on the model line is in our DeepSeek API models guide.


··········


HOW TO CHECK WHICH MODEL IDS YOUR CLAUDE, GEMINI AND DEEPSEEK KEY CAN CALL.


The most reliable source for a model string is the provider's own list endpoint, queried with your own key, because it reflects what your account can call today. Anthropic's List Models request, taken from its API reference, returns for each model the id, display_name, a lifecycle value (active, deprecated or retired), deprecated_at, retires_at, max_input_tokens and max_tokens:


curl https://api.anthropic.com/v1/models \
    -H 'anthropic-version: 2026-10-21' \
    -H "Authorization: Bearer $ANTHROPIC_API_KEY"

Google documents GET /v1beta/models for the list and GET /v1beta/models/{model} for a single model. Its getting-started guide shows the model code inside the request path:


curl "https://generativelanguage.googleapis.com/v1beta/models/gemini-3.8-flash:generateContent" \
  -H "x-goog-api-key: $GEMINI_API_KEY" \
  -H 'Content-Type: application/json' \
  -X POST \
  -d '{
    "contents": [
      {
        "parts": [
          {
            "text": "Explain how AI works in a few words"
          }
        ]
      }
    ]
  }'

DeepSeek documents GET /models, which returns each model's id with optional fields such as context_window and max_output_tokens. Its first-call example, against the OpenAI-compatible base URL https://api.deepseek.com, uses the recommended name:


curl https://api.deepseek.com/chat/completions \
  -H "Content-Type: application/json" \
  -H "Authorization: Bearer ${DEEPSEEK_API_KEY}" \
  -d '{
        "model": "deepseek-flash",
        "messages": [
          {"role": "system", "content": "You are a helpful assistant."},
          {"role": "user", "content": "Hello!"}
        ],
        "thinking": {"type": "enabled"},
        "reasoning_effort": "high",
        "stream": false
      }'

When the Claude API cannot find what a request asks for, it answers with HTTP 404 and an error object whose type is not_found_error; a request to a retired model fails. The Gemini rerouting and DeepSeek's legacy routing produce no error at all, so the only sign of a change is the model name in billing and usage reports or a difference in the answers.


··········


API MODEL NAMES IN NOVEMBER 2026: DOCUMENTED FACTS AND OPEN POINTS.


The documented facts are firm. Claude IDs from the 4.6 generation onward are pinned snapshots with their own retirement schedules, the dated claude-sonnet-4-5-20250929 retires on November 30, 2026, and the older claude-sonnet-4-5 alias resolves to that snapshot. Gemini stable codes are fixed while -latest aliases move with two weeks of email notice, and requests to gemini-2.5-pro and gemini-2.5-flash keep working after December 1 because Google reroutes them to newer models. DeepSeek's deepseek-flash is a pointer that has served two different models since September, and deepseek-chat and deepseek-reasoner have been retired since July 24.


Three points remain undocumented. Google does not publish which model gemini-flash-latest resolves to at a given moment; the forum report that usage is billed under the specific model comes from developers, not from Google. DeepSeek has set no end date for the legacy deepseek-v4-flash names. Anthropic's commitments for Claude Opus 4.5 and Claude Haiku 4.5 are floors that have been reached or are close, with no deprecation notice or replacement published as of November 3. For production code, the documented path is the same with all three providers: send an explicit current name, read the provider's list endpoint and deprecation page before each release, and treat any alias as a value that can change under you.


··········


FOLLOW US FOR MORE.


·····


DATA STUDIOS


·····


datastudios.org


bottom of page