Model launches get the headlines; retirements get a one-minute changelog entry. Here is the operational runbook for finding every pinned model ID in your agent stack and migrating it before the shutdown date turns into an outage.
Last Updated: November 17, 2026
GitHub deprecated three Copilot models yesterday and has six more scheduled for December 14. If your agents call models through APIs rather than a picker, a retirement is a hard failure, not a downgrade. Inventory every pinned model string across orchestration, MCP servers, routers and eval suites first; migrate per-step, not per-system; and decide deliberately which steps pin a snapshot and which float to a newer default. Floating does not remove the risk, it converts a scheduled outage into an unscheduled behaviour change.
What GitHub actually retired, and what is queued for December 14
On November 16, GitHub removed three models from most Copilot experiences — Claude Opus 4.8, Claude Opus 5 and GPT-5.6 Terra — across Copilot Chat, inline edits, ask and agent modes, and code completions. The suggested alternatives are Claude Opus 5.5 and, for Terra, GPT-6.1 Sol or GPT-6 Luna. Opus 4.8 survives for individual subscribers on annual plans, and Opus 5 stays reachable as a runtime refusal fallback for already-authorised access — two exceptions that make a tidy "everything moved" assumption wrong in at least two directions.
That retirement was announced on October 15. Thirty-two days of notice. Three days before it landed, GitHub published the next wave: six models deprecated on December 14 — Claude Fable 5, Claude Haiku 4.5, Gemini 3.8 Flash, Gemini 4 Argon, Grok 4.6 and Kimi K3. Thirty-one days of notice. One of those six, Kimi K3, has its suggested alternative listed as "To be announced," which is a politely worded way of saying the migration target does not exist yet.
The same changelog that is removing models is adding them faster: Claude Fable 5.2 on November 9, Gemini 4 Carbon on November 10, Gemini 4 Hydrogen on November 13. Hydrogen is the one that matters most for agent builders — Google positions it for multi-step agentic execution, and reports put it at 57% on Terminal-Bench 4.1. It is also, per GitHub's own December table, the replacement for Gemini 3.8 Flash. The model you are being told to adopt this week is the model that replaces the one being switched off next month. That is the cycle, and it has a back half nobody writes runbooks for.
A retirement notice is an operations event
For a human in a model picker, a deprecation is a non-event. GitHub auto-enables the suggested alternatives for Enterprise and Business customers under default model enablement, and the changelog says plainly that no action is required to remove the deprecated models. The dropdown reshuffles, people carry on.
For a program, it is the opposite. A hardcoded model string that no longer resolves does not degrade — it throws. And it throws the same way on every retry, because a shutdown surfaces as a 400/404-class model error, not a 429. Exponential backoff cannot save you from a model that ceased to exist; it just burns through your retry budget more expensively before the job dead-letters. In a queue-driven automation that processes overnight, the difference between a transient error and a permanent one is the difference between a delayed batch and a morning spent replaying several thousand failed runs by hand.
The second failure mode is worse because it is quiet. Gateways and routers that fall back to a configured default on an unknown model will keep returning 200s. Nothing pages. The step still produces output, just from a different model, with different formatting habits and a different refusal threshold, and you discover it three weeks later in a downstream data quality review. Given the scale enterprise agent estates have reached — AT&T's reported fleet of close to 500 production workflows is now an unremarkable number for a large operator — the blast radius of a silent substitution is no longer a single prompt.
Notice periods are also not uniform, and the published policy is not always what the table shows.
| Surface | Stated policy | What this cycle actually shows |
|---|---|---|
| OpenAI API | At least 6 months for GA models, 3 months for specialised variants, as little as 2 weeks for preview models | gpt-5: noticed Jun 11 2026, shutdown Dec 11 2026 (~183 days). GPT-4o: noticed Oct 14 2026, shutdown Jan 14 2027 — about 92 days, under the stated GA minimum, which the policy allows for safety or compliance reasons |
| Claude API | At least 60 days for publicly released models, with email notification to active deployments | Sonnet 4.5: deprecated Sep 30 2026, retirement Nov 30 2026 — 61 days. Active models carry forward-looking "not sooner than" dates, which is the most useful planning artefact any vendor publishes |
| GitHub Copilot | No minimum notice period stated in either deprecation changelog entry | 32 days (Oct 15 → Nov 16) and 31 days (Nov 13 → Dec 14) |
Plan against the shortest number in your stack, not the longest. If any step in a chain depends on a Copilot-surfaced model, your effective migration window for that chain is about a month, regardless of what the underlying model vendor promises on its own API.
Inventory first: find every pinned model string
You cannot migrate what you cannot enumerate, and model IDs leak into more places than anyone's architecture diagram admits. The ones that bite, in rough order of how often they get missed:
Orchestration node parameters. In n8n, the model name lives inside each node's parameter JSON, and it gets copied every time somebody duplicates a workflow. Twenty workflows cloned from one template means twenty independent copies of the same string, none of which read from an environment variable. Export the workflows and grep the JSON; do not trust the UI.
MCP server configuration. Sampling requests carry model preference hints, and tool servers frequently embed a preferred model for their own internal summarisation or extraction calls. These are invisible from the agent side until they fail.
Router and gateway configs. Primary, fallback, and the fallback's fallback. The last one is almost always stale, because nobody ever sees it fire.
Eval suites and golden datasets. Baselines pinned to a retired model become unreproducible the moment it is switched off. Anthropic is unusually candid about this cost, listing it as an acknowledged downside of retirement: researchers lose access to models for ongoing and comparative studies.
Prompt templates and fine-tune references. Model names embedded in system prompts ("You are GPT-5…"), and fine-tuned derivatives that follow their base model's shutdown schedule.
# 1. Every model-ish string across code, configs and exported workflow JSON
rg -n --no-heading -g '!node_modules' -g '!*.lock' \
-e 'gpt-[0-9][^"'\'' ]*' \
-e 'claude-[a-z0-9.-]+' \
-e 'gemini-[0-9][^"'\'' ]*' \
-e 'grok-[0-9][^"'\'' ]*' . | sort -u
# 2. Ask the vendor what is already condemned
curl -s https://api.openai.com/v1/models \
-H "Authorization: Bearer $OPENAI_API_KEY" \
| jq -r '.data[] | select(.shutdown_date != null)
| "\(.id)\t\(.shutdown_date)"'That second command is the piece most teams miss. OpenAI exposes an optional shutdown_date field on the list-models and retrieve-model endpoints specifically so you can monitor shutdowns programmatically rather than by reading blog posts. Anthropic's equivalent is the Usage export in the Console: export the CSV and you get usage broken down by API key and model, which tells you not just which deprecated models you reference but which ones are still taking live traffic. Diff those two outputs against your grep results. The gap between "strings in the repo" and "models receiving traffic" is your real migration scope — usually a third of what the grep returned, because half the hits are in dead branches and demo scripts.
Treating a dateless model ID as an evergreen pointer
The naming convention changed and the mental model did not. Anthropic's docs address this directly: for the 4.6 generation and later, a dateless ID such as claude-sonnet-4-6 is the canonical ID for a single fixed snapshot, not an alias that resolves to the latest version. Pre-4.6 IDs are different — there, a short form like claude-sonnet-4-5 is a convenience pointer to the most recent dated snapshot. So the same-looking string has opposite semantics depending on generation, and every one of them has its own retirement schedule.
What actually drifts when you swap a version
Swapping a model ID is a one-line diff with a surface area of about six things, none of which show up in a benchmark table.
Tool-call formatting and parallelism. The most common breakage is not a wrong answer, it is a differently-shaped one: a model that previously emitted one tool call per turn now emits three in parallel, and an orchestrator written to read tool_calls[0] silently drops two. Argument serialisation changes too — numbers arriving as strings, optional fields now always present and null.
Structured output adherence. Newer models are generally stricter about declared JSON schemas, which sounds like good news until you realise your prompt was compensating for the old model's looseness. Prompts carry years of scar tissue: "respond ONLY with JSON, no markdown fences." Against a model with native strict mode, that instruction is inert at best and occasionally counterproductive.
Refusal thresholds. Safety calibration moves between versions in both directions. A document-extraction step processing medical or financial records that ran clean for eight months can start refusing a small percentage of inputs with no code change. Measure refusal rate explicitly in your regression suite; it is the metric nobody tracks until it costs them a batch.
Token accounting and cost. Reasoning-capable replacements bill thinking tokens as output. A step costed at one price per thousand runs can move materially even when the headline per-token price fell. Prompt-cache behaviour also resets — your cached prefixes are cold against the new ID, so the first hours after a cutover look worse than steady state.
Latency distribution. Fast models are fast at the median. Agent loops care about p95, because a ten-step chain multiplies the tail. A model that is 20% faster on average but has a heavier tail will make your timeouts fire more often, not less.
Context window arithmetic. If a step was tuned to pack a context window to 85%, a replacement with different tokenisation changes what fits. Truncation then happens silently at the top of the message list, which is where your system prompt lives in several popular frameworks.
One underrated detail from Anthropic's versioning docs: even a pinned ID is not a fully frozen system. Weights are fixed for a given model ID, but the serving infrastructure around it — request router, safety classifiers, sampling logic — can change, and the docs note that infrastructure updates are the most likely cause when a previously stable model ID starts behaving differently. Pinning buys you weight stability. It does not buy you behavioural determinism, and anyone selling internal stakeholders on "we pinned it, so nothing can change" is going to have an uncomfortable conversation eventually.
Pin or float: the shelf-life test
The standard advice — just use the latest model — is a cost transfer, not a risk reduction. A floating alias means you never schedule a migration; it means the migration happens on the vendor's calendar, at an hour you did not choose, to a model you did not test, with no change in your repo to point at during the post-mortem. For a summarisation step with a human reading every output, that is a fine trade. For a confidence-gated step that decides whether a transaction proceeds without review, it is not.
Run each step through four questions. If any answer is yes, pin the snapshot.
| Question about the step | Pin a snapshot | Float to a current alias/default |
|---|---|---|
| Does a deterministic parser consume the output with no human in the loop? | Yes — formatting drift becomes data corruption | Only if the parser is schema-validated and fails closed |
| Is the decision audited, regulated, or subject to model-change disclosure? | Yes — you need to name the model version that produced a given decision | No |
| Does autonomy depend on a measured accuracy threshold? | Yes — the threshold is a property of the model version, not the workflow | No |
| Was the prompt tuned against this model's specific quirks? | Yes — the compensations become liabilities | If the prompt is generic and schema-driven |
| Is the step high-volume, low-stakes, human-reviewed? | Over-engineering | Yes — take the price and capability improvements automatically |
The practical outcome in most estates is a roughly 70/30 split: the majority of steps float, a minority pin — and the pinned minority is where your entire migration budget goes. That is the right shape. It also tells you where to spend eval effort, which is the same place. Our notes on model selection by workload go deeper on matching step type to model class.
The migration runbook
Migrate per step, not per system. A "move everything to Gemini 4 Hydrogen" ticket is unreviewable and untestable; forty small diffs, each with its own regression evidence, are both.
Deprecation response, in order
Grep plus vendor usage export. Every hit gets a tier: Tier 1 money-or-compliance-touching, Tier 2 customer-visible, Tier 3 internal and reversible.
Replay 200–500 real production inputs per Tier 1 step against the outgoing model and store outputs, tool-call traces, token counts and refusals. After the shutdown date this evidence is unobtainable. This is the step teams skip and regret.
Same inputs, new model ID, no side effects. Diff on schema validity, tool-call shape, refusal rate, p95 latency and cost per run — not on string equality, which will always differ and tells you nothing.
Where the diff is formatting, prefer tightening the schema contract over adding another defensive regex. Remove the old compensating instructions rather than layering new ones.
Tier 3 at 100%, Tier 2 at 10% for 48 hours, Tier 1 last and with human review restored until the accuracy threshold is re-established on the new version.
Keep the old ID reachable by config flag until the shutdown date, then delete it. A rollback to a retired model is not a rollback; it is a second outage.
Thirty days is enough for this if the inventory exists on day one, and nowhere near enough if it does not. That asymmetry is the whole argument for maintaining a model register continuously — a small table of call site, model ID, pinned-or-floating, vendor retirement date, owner, and tier — rather than reconstructing it under deadline every time a changelog entry lands. Teams we talk to through implementation work typically build this as a generated artefact from the grep, not a hand-maintained wiki page, because hand-maintained ones are wrong within a sprint.
When the honest answer is "do not migrate that step yet"
Three situations where moving fast is the wrong call.
The replacement is listed as "to be announced." Kimi K3's December 14 entry has no named alternative. There is no migration to execute; there is only a decision about whether to pre-emptively move to a different family now and absorb a second migration later, or wait and compress the work into whatever window remains after the announcement. If the step is Tier 1, move now to something stable and eat the cost. If it is Tier 3, wait.
The replacement is brand new. Gemini 4 Hydrogen shipped on November 12 and appeared in Copilot on November 13. Adopting a four-day-old model for a compliance-gated step because it is the designated successor to one being retired a month from now is trading a known risk for an unmeasured one. New models accumulate their own early-life quirks — rate-limit behaviour under burst, tool-call edge cases, provider-side capacity throttling — that surface in weeks two through six, not in launch benchmarks.
Your eval suite cannot tell the difference. If your regression tests pass identically against both the old and new model, that is not reassurance, it is a broken test suite. A real suite catches something on every version swap, even if the something is benign. Fix the evals before you trust them to approve a cutover. If a step has no eval at all and it is Tier 1, the migration task is actually two tasks, and the first one is not a migration.
None of this is exotic engineering. It is dependency management applied to a dependency that happens to be probabilistic, ships every three weeks, and gets removed with a month's notice by a vendor whose changelog entry takes one minute to read. The teams that handle the December 14 wave without incident will be the ones that already know, today, which of their steps hold a string that stops resolving in four weeks. If you want a second pair of eyes on that inventory before the window closes, start here — and the build notes cover how these registers get wired into existing orchestration.
Frequently Asked Questions
Which GitHub Copilot models were retired on November 16, 2026?
Claude Opus 4.8, Claude Opus 5 and GPT-5.6 Terra were deprecated across most Copilot experiences, including Chat, inline edits, ask and agent modes, and code completions. GitHub suggests Claude Opus 5.5 for both Opus models and GPT-6.1 Sol or GPT-6 Luna for Terra. Opus 4.8 remains available to individual subscribers on annual plans, and Opus 5 stays reachable as a runtime refusal fallback for existing authorised access.
How much notice do AI vendors actually give before retiring a model?
It varies by an order of magnitude. OpenAI states at least six months for generally available models, three months for specialised variants, and as little as two weeks for preview models. Anthropic states at least 60 days for publicly released models. GitHub's two most recent Copilot deprecation notices gave 32 and 31 days respectively, with no stated minimum. Plan against the shortest window in your stack.
Should I pin model versions or use floating aliases in production agents?
Both, chosen per step. Pin where a deterministic parser consumes the output, where decisions are audited, where autonomy depends on a measured accuracy threshold, or where the prompt was tuned to a specific model's quirks. Float where a human reviews every output and the step tolerates variance. Floating does not eliminate migration risk — it converts a scheduled, tested cutover into an unscheduled behaviour change you did not choose the timing of.
How do I find every hardcoded model ID across n8n, MCP servers and routers?
Combine three sources. Grep exported workflow JSON, server configs and prompt templates for vendor-specific model patterns. Query the vendor APIs — OpenAI exposes an optional shutdown_date field on the list-models and retrieve-model endpoints. Export actual usage, such as Anthropic's Console usage CSV broken down by API key and model, to see which deprecated models still take live traffic. The intersection of references and real traffic is your migration scope.
Does a pinned model ID guarantee identical behaviour over time?
No. Anthropic's versioning documentation notes that weights are fixed for a given model ID, but the surrounding serving infrastructure — request router, safety classifiers, sampling logic — can change, and infrastructure updates are the most likely explanation when a stable model ID starts behaving differently. Pinning gives you weight stability and a defensible audit record, not behavioural determinism. Keep monitoring output quality on pinned steps.
What should an agent regression test check after a model swap?
Not string equality — outputs always differ. Check schema validity rate, tool-call shape including how many calls arrive per turn and how arguments are serialised, refusal rate on real inputs, p95 latency rather than the median, token counts including reasoning tokens, and cost per run. Capture these against the outgoing model before its shutdown date, because once it is retired the baseline cannot be reproduced.
What do I do when the suggested replacement model has not been announced yet?
GitHub's December 14 list includes Kimi K3 with its alternative marked "To be announced." For high-tier steps, migrate now to a stable model in another family and accept that you may migrate again later; an unknown replacement arriving late compresses your testing window to near zero. For low-stakes, reversible steps, wait for the announcement and keep a dated reminder well before the shutdown date.
Research digest
AI Research Briefing
Honest insights on AI agents, Small Language Models, and local RAG. No hype. Only when we have something worth sending.
- No hype, just measurable outcomes
- Read by 2,400+ engineers
- Unsubscribe anytime




