What retires, and when
Microsoft's Foundry retirement schedule lists the three Cohere v5 previews with three different dates and no replacement for any of them.1
| Family | Foundry model ids | Stage | Retirement |
|---|---|---|---|
| Parse | Cohere-parse-v5 | Preview | 2026-12-15 |
| Embed | Cohere-Embed-V5-Pro, Cohere-Embed-V5-Fast | Preview | 2027-01-23 |
| Rerank | Cohere-rerank-v5-pro, Cohere-rerank-v5-fast | Preview | 2027-02-12 |
From today that is 25 days to parse, 64 to Embed and 84 to Rerank. January 23, 2027 is a Saturday, so the last working day before Embed retires is Friday, January 22.
The lifecycle policy says what retirement means. Once a model is Retired, every inference request returns 410 Gone. A preview is either force-upgraded to a newer preview or the GA model, or retired with no replacement, with at least 30 days' notice. There is no option to stay on a retiring preview.2 With the replacement column empty for all three, plan for termination and treat an upgrade as a bonus.
There is one discrepancy to settle first. Cohere's launch posts describe Embed 5 and Rerank 5 as generally available on Microsoft Foundry.34 The Foundry schedule shows the v5 entries as Preview.1 Your own subscription's answer is what counts, so the first checklist item below is to query it.
Why one sequence, not three tickets
The three families are stages of one pipeline. Parse turns documents into Markdown, with bounding boxes for visual elements.5 Embed turns chunks of that text into vectors. Rerank scores the retrieved candidates against the query. Each stage's output is the next stage's input, so three separate retirement tickets produce three separate quality changes, and nobody can tell which one moved the answers.
Two facts from Cohere shape the order:
- Embed 5 Pro and Fast share one embedding space, so you can index with Pro and query with either tier without re-indexing.3 That is a statement about the two tiers, not about vectors from a preview deployment and from another host. Treat that as untested until you have tested it.
- Rerank 5 scores are described as calibrated, so one threshold can decide what reaches the LLM.4 A threshold you tuned on one reranker version is a number you must re-check on the next.
The rule that follows: decide all three destinations before the first cutover, then move in pipeline order. If you move parse on December 15 and decide about embeddings later, you may re-embed the corpus twice. This post does not repeat the dual-index plan for Embed v3; the Embed v3 post covers it.
The dated sequence
The retirement dates come from Microsoft. The working dates in between are proposed buffers, not vendor commitments.
| Date | Step | Done when |
|---|---|---|
| Nov 20 | Inventory and lifecycle query | Every caller of the three families is listed |
| Nov 25 | Golden set frozen | Queries, expected passages and sample documents saved |
| Dec 1 | Destinations chosen for parse, embed and rerank | Written decision with host, model id and dimensions |
| Dec 8 | New parser serving all new ingestion | Old parser handles nothing new |
| Dec 15 | Parse v5 retires | No traffic to Cohere-parse-v5 |
| Dec 16 to Jan 9 | One re-embed of the corpus from the new parser output | Index stamped and evaluated on the golden set |
| Jan 12 to Jan 16 | Query path moves to the new embedding model | Canary at 100%, old path left in place |
| Jan 22 | Last working day before Embed retires | No traffic to Cohere-Embed-V5-* |
| Jan 26 to Feb 3 | Reranker moves, threshold set again | Shadow run reviewed, then cutover |
| Feb 12 | Rerank v5 retires | No traffic to Cohere-rerank-v5-* |
The Dec 16 to Jan 9 window is the re-embed. It starts after the parser change because the parser decides the text that gets embedded. It finishes two weeks before Embed retires, so a failed run still has time to be repeated.
Step 1: inventory and lifecycle check (today)
- List every service, job and notebook that calls
Cohere-parse-v5,Cohere-Embed-V5-Pro,Cohere-Embed-V5-Fast,Cohere-rerank-v5-proorCohere-rerank-v5-fast. Search deployment names too, not just model ids, since a deployment can carry any name. - Record for each caller who owns it and whether it serves users, batch ingestion or evaluation.
- Query the lifecycle fields in the Models API for your subscription and region, and save the output. The policy page names
lifecycleStatus,deprecation.inferenceand the per-SKUdeprecationDate.2 - Confirm who receives Microsoft's retirement email. Notices go to subscription owners with active deployments, and preview notices are at least 30 days ahead.2
- Check Azure Service Health for advisories. The policy page describes them and an alert rule for email, text message or webhook.2
A small script keeps this check repeatable:
import os, requests
SUB = os.environ["AZURE_SUBSCRIPTION_ID"]
LOC = os.environ.get("AZURE_LOCATION", "eastus")
TOKEN = os.environ["AZURE_ARM_TOKEN"] # e.g. from `az account get-access-token`
url = (
f"https://management.azure.com/subscriptions/{SUB}"
f"/providers/Microsoft.CognitiveServices/locations/{LOC}/models"
"?api-version=2024-10-01"
)
data = requests.get(url, headers={"Authorization": f"Bearer {TOKEN}"}, timeout=30).json()
def walk(node):
# The exact nesting can differ by api-version; find dicts that carry lifecycleStatus.
if isinstance(node, dict):
if "lifecycleStatus" in node:
yield node
for v in node.values():
yield from walk(v)
elif isinstance(node, list):
for v in node:
yield from walk(v)
for item in walk(data):
text = str(item)
if "Cohere" in text or "cohere" in text:
print(item.get("name"), item.get("lifecycleStatus"), item.get("deprecation"))Run it weekly until February 12 and diff the output. A date that moves is information you want before the vendor email.
Step 2: pick the three destinations (by Dec 1)
The destination decision is the step that most often gets skipped. For each family, write down the host, the model id, and what changes for callers. The options the sources support:
- The same models on another host. Cohere says Embed 5 is available on the Cohere API and Model Vault, Microsoft Foundry and Amazon SageMaker. It says Rerank 5 is available on those and to North users, and Parse through the Cohere API and Model Vault.345 OpenRouter also lists Embed 5 and Rerank 5.6
- A Foundry GA listing, if your subscription shows one. The Foundry schedule already points Cohere's older GA
rerank-v4.0entries at the v5 ids as their replacement, which is odd for a preview target.1 Ask Microsoft which listing is meant to carry on. - A different vendor for one stage. Allowed, but it changes that stage's output and every stage after it.
Cohere lists these API rates, which may differ from what your Foundry deployment or another host bills. Check with whoever invoices you.
Before choosing, ask whether the hosts return the same vectors and the same scores. Neither source says preview deployments and other hosts are interchangeable. Test with 200 documents: embed on both, then compare nearest neighbours on the golden set.
Step 3: parse cutover (Dec 1 to Dec 15)
Parse is first because it is the earliest retirement, and because its output sets the text for everything after it.
- Run the destination parser over 200 representative documents. Include tables, scans, multi-column pages, and any language Parse supports. Cohere lists nine major world languages.5
- Diff the Markdown against the old output. Look at table structure, reading order, repeated headers and dropped content, since these are what change chunking.
- Re-chunk the new output with the same chunker. If chunk boundaries move a lot, expect the golden set's expected passages to need re-labelling, so do that now.
- If you store bounding boxes for citations or highlighting, confirm the new parser returns them in the same shape. Cohere describes Parse as returning bounding boxes for visual elements.5
- By Dec 8, send all new ingestion through the new parser, and keep the old parse outputs read-only.
- On Dec 15, confirm no traffic still reaches
Cohere-parse-v5.
A parser swap leaves the old index intact, and old chunks remain searchable. They are now inconsistent with new documents, which is why the re-embed follows rather than being skipped.
Step 4: one re-embed, then the query path (Dec 16 to Jan 22)
Re-embed everything from the new parser output, using the embedding destination you chose on Dec 1. Doing this once, after the parser settles, is the point of the sequence.
- Stamp every chunk with parser id, embedding model id, dimensions, vector format and ingest date. Without a stamp you cannot tell which chunks are stale later.
- Pick dimensions deliberately. Embed 5 offers 2048, 1536, 1024, 768, 512 and 256, and float, int8 and binary formats.3 Cohere's own OpenRouter walkthrough uses 1,024-dimension vectors.6 Changing dimensions changes the index, so choose it here, not mid-flight.
- Keep the document and query input types matched to what your host expects. The OpenRouter example passes
search_documentfor indexing andsearch_queryfor queries.6 - Evaluate the new index on the golden set before any traffic moves: recall at your candidate depth, and the rank of the expected passage.
- From Jan 12, route the query path to the new embedding model behind a canary, and keep the old query path untouched until Jan 16.
- On Jan 22, confirm no traffic still reaches
Cohere-Embed-V5-ProorCohere-Embed-V5-Fast.
An embedding model and the index it built are a pair. Never let a query vector from one model search an index built by another.
{
"index": "docs-2027-01",
"parser": "<destination-parser-id>",
"embedding_model": "<destination-embed-id>",
"dimensions": 1024,
"vector_format": "float",
"built_at": "2027-01-08",
"rerank_model": "<destination-rerank-id>",
"rerank_threshold": null
}Keep the manifest next to the index and make the query service refuse to start if its configured model does not match.
Step 5: rerank and its threshold (Jan 26 to Feb 12)
Rerank goes last because it reads whatever the first two stages produce. Its inputs are the new chunks and the new candidate lists. Run it against the new retrieval, not the old.
- Replay the golden set through the new retrieval and the destination reranker, and compare the order of the top results with the old pipeline.
- Set the relevance threshold again from this run. Cohere describes Rerank 5 scores as calibrated across queries and domains.4 That is a property within one model. It does not tell you a threshold from another version will hold.
- Check chunk size against input limits. Cohere's OpenRouter post says Rerank 5 reads up to 32k tokens of query and chunk combined and splits longer input into several passes.6 Splits change both latency and billed tokens.
- If you adopted Context Focus, which Cohere describes as new in Rerank 5 and returning the most relevant passages inside each document,4 list which downstream code reads those spans, because a fallback to an older reranker will not return them.
- Choose Pro or Fast on measured results. The Pro vs Fast post covers that test.
- Decide what the pipeline does when the reranker is unreachable: skip reranking and use the embedding order, or fail the request. Write it down and test it.
- On Feb 12, confirm no traffic still reaches
Cohere-rerank-v5-proorCohere-rerank-v5-fast.
Catching the dates from the gateway side
If your calls pass through INFRO, its model retirement records warn organizations still sending a retiring model and, from the shutdown date, refuse it before any provider call, naming the replacement. Deprecations also appear under the console bell and are emailed to owners and admins. That helps the inventory step, but it does not replace the Models API check above, because INFRO can only see calls that pass through it.
INFRO's POST /v1/rerank uses Cohere's request shape, but it is served by Jev or another decision model you connect, which judges each passage against the query. It is not Cohere Rerank.7 Treat it as a candidate to evaluate on your golden set, not as a drop-in for Rerank v5 scores or thresholds.
What to leave open
Three things are not confirmed by the pages read on 2026-11-20: what Foundry bills for the v5 models, whether vectors from a retiring preview match another host's vectors, and whether Microsoft will name a replacement or extend any date. Preview retirement dates are described as "not sooner than" dates that are sometimes extended.2 Plan to the dates in the table, and treat an extension as a reason to keep your buffer, not to skip the work.
