Model Aliases and Retirements in 2026: The Full Ledger

Model aliases and model retirements are the two ways an API model ID stops meaning what it meant when a system was built: a floating alias silently changes what it points to, and a retired ID stops answering at all. This ledger records both for 2026, from the vendors’ own pages, in two complete tables.

Table A censuses 19 floating aliases: twelve tilde-prefixed aliases read from OpenRouter’s live catalog, Google’s native gemini-flash-latest, OpenAI’s two Daybreak aliases and its *-chat-latest naming pattern, Anthropic’s pre-4.6 convenience-alias pattern, and two rows recording that Anthropic’s current dateless IDs and DeepSeek’s API run no floating alias at all. Table B transcribes 89 retirement rows dated in 2026 across ten vendor groups, with the notice period computed wherever the vendor publishes both the announcement and the shutdown date, and marked not computable or inferred wherever it does not. A handful of those rows name several model IDs on one line, exactly as the vendor’s own table does.

The page is organised so that the tables can be cited cell by cell. A methodology block states what was fetched and when, what was excluded and why, and which cells were inferred rather than read. The as-of date for the dataset is 22 August 2026. The one exception is the live alias-target column in Table A, a single point-in-time reading whose limits that table’s caption states.

A floating alias, for Table A, is any API-callable model name that the vendor or router documents as resolving to a different underlying model over time without the caller changing the string. That includes OpenRouter’s ~-prefixed catalog entries, Google’s documented -latest hot-swap, OpenAI’s Daybreak and *-chat-latest names, and Anthropic’s older within-minor-version convenience aliases. Two further rows record, with a source, that a vendor runs no such mechanism: those are kept because the absence is the citable fact.

A retirement, for Table B, is the shutdown or decommission of a specific API-callable model ID: the point at which a request to that literal string starts failing, or, in xAI’s May 15 wave, starts silently redirecting, which the table marks separately. Only rows with a shutdown date in 2026 are included; two Anthropic rows announced in late 2025 are kept because their shutdown fell in 2026 and are labelled as such.

Three categories were deliberately excluded. Consumer-app model switches, such as ChatGPT retiring GPT-4o, GPT-4.1 and o4-mini from the app on February 13, 2026 with no API change, are not model-ID retirements; the underlying API IDs appear in Table B under their own, different API dates. Product or tool retirements that are not model IDs are out, which is why Claude Workbench’s August 17 retirement, a product rather than a model, is absent. Subscription-tier migrations, such as Z.ai’s Legacy Plan phase-out on April 30 and Legacy Plan V2 discontinuation on July 30, are billing changes and are excluded even though they were the only deprecation-adjacent content found on Z.ai’s docs.

Every number on this page comes from one of three operations: a programmatic pull of a public API, a direct fetch of a vendor’s own documentation page, or subtraction between two dates found on the same page. Where a cell rests on anything else, the cell says so in capitals. The full method, including what was not done, follows.

Rows 1 to 12 are OpenRouter’s tilde aliases, listed in the order the catalog returns them, with the listing date taken from the API’s created field. Rows 13 to 17 are vendor-native aliases or alias patterns. Rows 18 and 19 are recorded absences. The resolves-to column is the one live-state reading in the dataset, and the caption below records what that means for citing it.

Two cells in Table A deserve a reading note. The 2098-12-31 value on row 1 is not a lifespan: OpenRouter returns the identical value for other catalog entries with no expiry set, so it is the platform’s placeholder, and the other eleven tilde aliases simply return null for the same field. And the resolved-ID cell on row 13 rests on a single unresolved GitHub issue against the Vertex AI Python SDK, which reported that models.get on gemini-flash-latest returned a version of default and that the response echoed the alias string rather than the serving model; no maintainer response was visible, it was not reproduced against the direct Gemini endpoint, and Google has not acknowledged it. It is recorded as reported, not as fact.

The first thing the census changes is the framing of the glm-latest floating alias, which this site examined on August 19 as a single example. The live catalog holds twelve such entries across seven vendors (Z.ai, DeepSeek, xAI, Anthropic, OpenAI, Google, Moonshot), seven of them listed within two minutes of each other on 2026-04-27, all on the same ~vendor/family-latest naming pattern and the same description template. That is a platform feature with a rollout date, not a property of one vendor’s model.

OpenRouter’s own documentation is the clearest statement of the trade-off. Its routing guide says versions can change at any time and tells callers who need reproducibility to use the concrete slug instead; its June 15 tutorial on keeping agents running when models disappear recommends the opposite, routing fallback chains through a self-updating alias as insurance against retirement. Both are linked in the sources list in section 08; the ledger records the mechanism, not a verdict on which advice to follow. The guide also documents a second silent-change vector: unsupported reasoning parameters are remapped to the nearest supported value on a ~latest slug where a concrete slug would return a 400.

The second finding is the least obvious one. A name ending in -latest is not exempt from a retirement schedule. gpt-5.2-chat-latest and gpt-5.3-chat-latest sit on OpenAI’s deprecations page as ordinary entries with an announcement date, a shutdown date and a replacement, and chatgpt-image-latest joins them with a December 1 shutdown. A floating alias can float for a while and then stop answering on a clock, exactly like a dated snapshot.

The two absence rows are the third finding, and they are different kinds of absence. Anthropic’s is a stated design position: its documentation says a dateless ID in the 4.6 generation and later maps to a single fixed snapshot, that weights and configuration are never updated in place, and that an updated model ships under a new ID. The older pre-4.6 pattern, where claude-sonnet-4-5 resolved to the newest dated snapshot within that minor version, is kept as row 17 because it is a genuine if narrow floating mechanism and because the response named the resolved snapshot. DeepSeek’s absence is an outcome rather than a policy: when the legacy deepseek-chat and deepseek-reasoner names were discontinued on July 24, no floating replacement was offered and callers had to hard-code the new literal IDs. The two rows are the same column entry, no floating alias, reached by opposite routes.

Rows are grouped by vendor, each group headed by the page it was transcribed from and the vendor’s own notice-period language where any exists. Within a group, rows follow the order of the vendor’s page. The notice column is a plain number only where it was computed from two dates on that page; otherwise it reads not computable, or INFERRED with the basis stated in the announcement cell. Every shutdown date later than 22 August 2026 is scheduled rather than observed. That holds for the whole table, so read it as a blanket condition on the shutdown column rather than looking for a marker on each such row.

Two rows sit close to the dataset’s own as-of date. Google’s Imagen 4.0 generate, ultra and fast IDs shut down on 2026-08-17, five days before the as-of date and inside the window this batch was researched in. Moonshot’s Kimi K2.5 and moonshot-v1 rows carry a scheduled 2026-08-31 shutdown, nine days after it, and an announcement date this ledger could only infer. Both are in the table as they stand; neither has been observed to happen.

The notice column is the one number in the ledger that is comparable across vendors, computed rather than quoted, and the first thing a reader building a pinning policy needs. The chart below plots it for the four vendor groups where it can be computed at all, as a minimum-to-maximum range with the row count, and keeps a fifth row for the 38 rows where it cannot, because a vendor that publishes no announcement date is a finding, not an empty space.

Read against the vendors’ own policy language, the computed numbers behave. Anthropic’s deprecations page commits to at least 60 days’ notice for publicly released models, and its six rows land at 60, 61, 62, 62, 62 and 114. OpenAI’s page states at least six months for GA models, at least three months for chat, codex and deep-research variants, and possibly as little as two weeks for previews; its 31 rows split into 182 to 184 days for the GA and image waves and 94 days for the two -chat-latest aliases, which matches the three-month tier rather than the six-month one. DeepSeek’s single 2026 entry said three months and delivered 91 days. Mistral publishes no general sentence, and its per-model spread, from Leanstral’s 39 days to Medium 3.1’s 101, is the widest in the computable set.

The rows where the thesis of a tidy retirement clock fails are the most useful ones to a reader checking whether an integration will break, and they fall into four distinct stances. Each is a row group in Table B or a row in Table A, kept as found.

Projecting forward from the table rather than from any single row, two pressures point in opposite directions. Routers have a product reason to add floating aliases, and OpenRouter’s seven-at-once rollout on April 27 suggests they are added in batches, so the alias count is more likely to rise than fall. Model vendors, by contrast, have been converging on dated or fixed IDs with a published floor: the Anthropic clarification, OpenAI’s tiers and DeepSeek’s literal-ID migration all move the same way. A reasonable expectation is that the alias layer keeps thickening at the router while the vendor layer keeps hardening underneath it, which makes the resolves-to column in Table A the cell most likely to be stale by the time it is read, and the one this ledger will refresh first.

This page is the census; three earlier pages on this site are its narrative companions, each built on rows that now sit in these tables. The glm-latest alias post examines Table A row 1 in depth, as the failure mode of a name that repoints with no signal at the call. The DeepSeek July 24 retirement post is the migration story behind Table B’s two DeepSeek rows and their 91-day notice. The one-page model-pinning policy is the policy companion to this census, turning the notice-period column into rules for production automations.

Two further pages give the wider frame. A floating model alias is a specific instance of the general versioning-contract problem covered in the API-versioning decision matrix, and the created caveat on Table A’s listing dates is the same lesson as listing dates versus launch dates in the catalog-literacy guide. Teams that want to apply the ledger to their own model inventory, pinning concrete IDs and scheduling migrations against the notice column, can look at our AI transformation engagements.

When citing a cell, quote the as-of date with it. Table B figures are as of 22 August 2026; Table A’s resolves-to values are a point-in-time reading, flagged as such in that table’s caption, and may have changed since, which is the point of the column. If a vendor page you re-fetch later disagrees with a cell here, the change is itself the finding, and the as-of date is what makes it one.

Primary sources, as linked in the tables

• OpenRouter: models API, Latest Model Resolution guide, June 15 tutorial on model disappearance.
• Anthropic: Model deprecations, Model IDs and versions.
• OpenAI: Deprecations, Models index, daybreak-red-latest.
• Google: Gemini API deprecations, Gemini API models; googleapis/python-genai issue #2271 (unresolved, reported only).
• DeepSeek: API change log. Moonshot AI: Kimi model list. Alibaba Cloud: Model decommissioning policy.
• Cohere: Deprecations. xAI: May 15 retirement migration guide. Mistral AI: Models overview.
• Z.ai, checked and found to hold no lifecycle content: pricing docs, GLM-4.6 docs.