CertSafari
    Anthropic Claude Certified Architect – Professional (CCAR-P) - December 7 onwards· Lessons

    Domain 4 · Lesson 27/39

    Model version migration: deprecation timelines, regressions and drift

    Model Version Migration

    17 min read
    2.43% of exam
    8 sources
    Published 18 Nov 2026
    Docs as of 18 Nov 2026

    What you will be able to do

    • Read a model's lifecycle status and retirement date and treat that date as the point after which requests fail
    • Explain what pinning a model ID guarantees and why it cannot hold a workload on a model once that model retires
    • Interpret regression results for a candidate model against fixed success criteria rather than the candidate's own scores
    • Choose between pinning, re-tuning prompts, re-baselining evaluations, or deferring a workload so that drift is managed before the retirement date

    Key concept

    Retirement date as a hard deadline — Every migration plan works backwards from the retirement date of the current model ID. Until then, pinning keeps behaviour stable. After it, requests fail, so re-tuning, re-baselining and any fallback all have to be finished before that date.

    1.The model lifecycle and why retirement is a hard deadline

    A model version migration starts with a date you don't control. Anthropic retires older models as newer ones launch, and an application that calls a retired model stops working. Behavioural drift, meaning the ways a new model answers your prompts differently from the old one, has to be measured and fixed inside the window between the deprecation notice and the retirement date. Every model passes through four named states. Only the last one breaks your application.

    Model lifecycle states and what each means for a running application
    StateMeaningWhat it means for your application
    ActiveFully supported and recommended for useNo action needed. Track the tentative retirement date.
    LegacyNo longer receives updates and might be deprecated in the futureStart planning, because a deprecation notice may follow.
    DeprecatedStill functional but no longer recommended; a replacement and retirement date are assignedThe migration clock is running. Test the replacement now.
    RetiredNo longer available for useRequests fail.

    Retirement is a failure, not a slowdown or a silent swap. Nothing maps a retired ID to its replacement, so an unmigrated integration starts erroring on that date. The planning window is predictable, though. Anthropic notifies customers with active deployments, giving at least 60 days' notice before retiring a publicly released model. Notice arrives directly, by email and in the documentation, so you don't have to discover it by polling a web page.

    Anthropic's published model deprecations page carries a table of deprecation and retirement dates for each model, and it puts real dates on this. claude-sonnet-4-5-20250929 was deprecated on September 30, 2026, with a retirement date of November 30, 2026 and claude-sonnet-5-5 as the recommended replacement. That is 61 days from notice to failure. Everything in this lesson has to fit inside that window: the regression run, prompt changes, re-measurement and rollout.

    Two details change which date applies and which code is affected. First, the published dates cover Anthropic-operated platforms only. Amazon Bedrock and Claude Models on Google Cloud (Google managed) set their own retirement schedules, so the same model can retire on a different day there. Second, you can't migrate usage you haven't found. Exporting from the Console Usage page gives a CSV of usage by API key and model, which shows every place that still calls the deprecated model.

    Checkpoint 1 of 9· Match them up

    Match each model lifecycle state to what it means for an application that still calls that model.

    Tap a term, then the definition that fits it.

    Checkpoint 2 of 9· Exam question

    Which statement correctly describes how model retirement works for Anthropic-operated platforms, and is therefore the right planning assumption for a migration timeline?

    Sources1

    2.What pinning a model ID guarantees, and when it stops working

    Pinning means calling a specific model ID so that the model doesn't change underneath you. Each Claude model ID identifies a fixed snapshot. From the 4.6 generation on, the dateless ID, such as claude-sonnet-4-6 or claude-opus-4-8, is the snapshot itself, not a pointer to the newest version. Updated weights always ship under a new ID. Earlier models have dated IDs like claude-sonnet-4-5-20250929, plus convenience aliases on the Claude API, such as claude-sonnet-4-5, that resolve to the latest dated snapshot. The aliases are the one place where the model behind a name can move.

    Pinning gives you a stable baseline to evaluate against, with one caveat. Weights are fixed for a given ID, but the serving infrastructure around the model can change: the request router, safety classifiers and sampling logic. That can cause minor behavioural differences on an ID that has not changed. If a pinned workload drifts slightly when no migration is under way, suspect infrastructure, not a model swap.

    The limit of pinning follows from that. A pin defends against unplanned change. It does nothing against a planned retirement. Pinning a workload to the current model is a sound way to buy time while you fix it, but only up to the retirement date. Deferring a workload "to the next release cycle" by pinning it fails whenever the retirement date comes first.

    Checkpoint 3 of 9· Check yourself

    A team is setting a fixed baseline for a regression run on the Claude API. They want the model behind their configured name to stay the same. Which of these model names could start resolving to a different model snapshot without the team changing anything?

    Sources2

    3.Reading regression results against fixed success criteria

    Migration readiness is decided by running the candidate model through the same regression suite as the current model and comparing the two. The suite has to reflect your own use case: benchmark tests built from your actual prompts and data, compared on accuracy, response quality and edge-case handling. Anthropic advises doing this well before the retirement date, because a failed run with only a week left leaves no room to respond. How to build the comparison itself is covered under controlled configuration comparison.

    Pass thresholds come from what the application needs, not from what the candidate scored. A good success criterion is specific, measurable and aligned with the application's purpose and its users. A JSON workflow that feeds a downstream parser needs schema-valid output whichever model produces it. Lowering that workflow's threshold to the candidate's observed score turns a failing result into a passing one without changing what the parser will receive in production.

    Where the failures fall matters as much as how many there are. A 97 percent overall pass rate with nearly every failure in one workflow is a localized regression: the rest of the application is ready and one piece is not. A fifteen percent drop on a compliance-sensitive metric, such as the correctness of rejection reasons in loan decisions, is a different case. Here the size of the drop and the cost of each error both rule out shipping the candidate as it stands, whatever the aggregate score says.

    Checkpoint 4 of 9· Check yourself

    Five weeks before the current model retires, the candidate passes 97 percent of the regression suite. Almost every failure is in one invoice-extraction workflow. Its JSON goes to a downstream parser, and only 88 percent of the candidate's output is schema-valid against the 99 percent bar. What should the team do?

    Sources314

    4.Re-tuning prompts and re-baselining evaluations

    Prompts that work on one model are not guaranteed to work on the next, because each model release changes behaviour in documented ways. Anthropic publishes a per-model prompting page listing differences such as effort calibration, literal instruction following, response length and tool-use triggering. Anthropic's guidance says a technique measured on one model should be treated as measured on that model and checked against your own evals before you rely on it elsewhere.

    Effort is a request setting (output_config with an effort level such as medium or high) that controls how deeply the model thinks before answering. Each model allocates tokens differently at each level, so the same level can cost more or less on a new model. An effort sweep means running your own evals at each effort level and comparing quality, cost and latency, so you can pick the lowest level that still meets your success criteria.

    The code change in a migration is often a single line, which is why behaviour changes are the real workpython
    model = "claude-fable-5-1"  # Before
    model = "claude-fable-5-2"  # After

    Claude Fable 5.2 shows how a carried-over prompt can drift. Claude Fable 5.1 usually follows a line such as "answer directly without deliberating", but in Anthropic's testing that line did not stop Claude Fable 5.2 from thinking at medium effort and above. Thinking tokens count toward max_tokens. A request that depended on that line to fit a small max_tokens can therefore stop with stop_reason: "max_tokens" before producing any text. The fix is to re-tune: remove the line, lower the effort level, or raise max_tokens. Some settings fail outright instead of drifting. From API version 2026-10-21, temperature, top_p and top_k return a 400 error, and the documented replacement is to omit them and guide behaviour through prompting.

    Re-baselining means measuring the candidate's operating profile again, not adopting its quality scores as the new standard. Anthropic's migration guide for each model includes a checklist of the changes to make when moving a workload onto it, and the Claude Fable 5.2 guide's checklist asks you to re-run your effort sweep on your own evals and re-baseline cost and latency, because effort levels, thinking behaviour and pricing all change between models. Quality thresholds stay where your success criteria put them. Re-tuning should also be scoped. Revising the prompt for the one workflow that failed, then re-running the whole suite to confirm the other workflows didn't regress, keeps the change contained. Moving every workflow to a new shared template at once would put workflows that already passed at risk.

    Checkpoint 5 of 9· Check yourself

    A request moved from claude-fable-5-1 to claude-fable-5-2 now stops with stop_reason: "max_tokens" and returns no text. The system prompt still says "answer directly without deliberating", effort is medium, and max_tokens is small. Which change deals with the cause?

    Checkpoint 6 of 9· Check yourself

    Which of these is re-baselining as Anthropic's migration checklist means it?

    Sources56178

    5.Choosing pin, re-tune, re-baseline or defer within the timeline

    No single response handles every regression. The right one depends on where the regression sits, how serious it is, and how many weeks remain before the retirement date. Each response has a situation where it wins and a limit it cannot get past.

    Migration responses: when each wins and what it cannot do
    ResponseWhen it winsWhat it cannot do
    Pin the current model IDHolding a failing workload stable while its fix is developedServe requests after the current model's retirement date
    Re-tune promptsThe regression traces to model-specific behaviour, such as instruction following or effort calibrationMake a fix count without a full regression re-run afterwards
    Re-baseline evaluationsEffort, cost and latency profiles change with the new modelJustify lowering quality thresholds to the candidate's scores
    Defer cutover of a workloadThe candidate still misses target scores and weeks remain for a gated decisionRun past the retirement date; deferral needs a go/no-go gate before it

    Localized failure, comfortable timeline. Take a support bot with eight weeks to retirement and failures concentrated in one strict-JSON workflow. That case calls for splitting the migration. Re-tune the prompt for the JSON workflow only and re-run the full suite on the candidate. Meanwhile, move the workflows that pass onto the candidate gradually, watching production as each share grows, while the JSON workflow stays on the current model until its fix passes. Check the candidate's rate limits before you move production traffic, since a new model can have its own limit. The two tracks run in parallel and both finish before retirement. Pinning the JSON workflow and pushing it out to a later release cycle fails, because the retirement date arrives first.

    Large regression on a high-stakes metric, no tuning yet. Now take a loan pre-approval assistant with nine weeks left, a fifteen percent drop in rejection-reason correctness, and no prompt work attempted. Here the candidate should not touch production, not even a small share behind a feature flag. Routing real applicants to a model with a known compliance regression and relying on reviewers to catch errors puts that harm in front of customers. Keep production on the current model until retirement. Time-box prompt tuning of the candidate against fixed target scores. Set a go/no-go gate early enough to leave weeks for a fallback, such as another replacement model or a heavier review process, if tuning falls short.

    Checkpoint 7 of 9· Match them up

    A migration plan has four responses to choose from. Match each one to the situation it fits and the limit it cannot get past.

    Tap a term, then the definition that fits it.

    Checkpoint 8 of 9· Exam question

    A support-bot team evaluated a candidate model. Overall pass rate is 97 percent, but nearly all failures sit in one workflow whose responses must match a strict JSON schema. Retirement of the current model is eight weeks away. Choose two actions that best manage the drift within the timeline.(Select 2)

    0 of 2 selected

    Checkpoint 9 of 9· Exam question

    A bank's loan-pre-approval assistant runs on a model scheduled for retirement in nine weeks. The candidate model shows a fifteen percent drop in accuracy on rejection-reason correctness, a compliance-sensitive metric. Prompt re-tuning has not yet been attempted. Which response best fits the timeline?

    Sources61

    Sources

    (8)

    Every claim above is drawn from one of these pages, quoted as it was written on the date shown.

    1. 1.
      “Deprecated: The model is still functional but no longer recommended. Anthropic provides a recommended replacement and assigns a retirement date.”
      ↩︎ The model lifecycle and why retirement is a hard deadline
      “providing at least 60 days’ notice before model retirement for publicly released models”
      ↩︎ The model lifecycle and why retirement is a hard deadline
      “Impacted customers will always be notified by email and in the documentation.”
      ↩︎ The model lifecycle and why retirement is a hard deadline
      “Partner-operated platforms, Claude Models on Google Cloud (Google managed) and Amazon Bedrock, set their own retirement schedules”
      ↩︎ The model lifecycle and why retirement is a hard deadline
      “Review the downloaded CSV to see usage broken down by API key and model.”
      ↩︎ The model lifecycle and why retirement is a hard deadline
      “Current and recently retired models are listed in the following table with their status:”
      ↩︎ The model lifecycle and why retirement is a hard deadline
      “Test your applications with newer models well before the retirement date of your current model.”
      ↩︎ Reading regression results against fixed success criteria
      “Omit and use prompting to guide model behavior.”
      ↩︎ Re-tuning prompts and re-baselining evaluations
      “consider thorough testing of your applications with the new models well before the retirement date.”
      ↩︎ Choosing pin, re-tune, re-baseline or defer within the timeline
      “Once a model is deprecated, migrate all usage to a suitable replacement before the retirement date. Requests to models past the retirement date will fail.”
      ↩︎ Key concept
      “Retired: The model is no longer available for use. Requests to retired models fail.”
      ↩︎ Exam trap 1
      “Retired: The model is no longer available for use. Requests to retired models fail.”
      ↩︎ Prediction
      “Once a model is deprecated, migrate all usage to a suitable replacement before the retirement date. Requests to models past the retirement date will fail.”
      ↩︎ Checkpoint
    2. 2.
      “When you use a model ID in an API request, the underlying model remains constant for the lifetime of that ID.”
      ↩︎ What pinning a model ID guarantees, and when it stops working
      “An alias such as claude-sonnet-4-5 is a convenience pointer that resolves to the most recent dated snapshot for that minor version.”
      ↩︎ What pinning a model ID guarantees, and when it stops working
      “Occasionally, infrastructure updates produce minor differences in observable behavior even when the model ID and weights have not changed.”
      ↩︎ What pinning a model ID guarantees, and when it stops working
      “Every model ID, whether dated or dateless, has its own distinct deprecation and retirement schedule.”
      ↩︎ Exam trap 2
    3. 3.
      “Create benchmark tests specific to your use case. Having a good evaluation set is the most important step in the process.”
      ↩︎ Reading regression results against fixed success criteria
    4. 4.
      “A/B testing: Compare performance against a baseline model or earlier version.”
      ↩︎ Reading regression results against fixed success criteria
      “Align your criteria with your application’s purpose and user needs.”
      ↩︎ Exam trap 3
      “Relevant: Align your criteria with your application’s purpose and user needs.”
      ↩︎ Checkpoint
    5. 5.
      “treat it as measured on that model and re-check it against your own evals before applying it to another.”
      ↩︎ Re-tuning prompts and re-baselining evaluations
    6. 6.
      “Re-run your effort sweep on your own evals, and re-baseline cost and latency.”
      ↩︎ Re-tuning prompts and re-baselining evaluations
      “Check your rate limits before you move production traffic.”
      ↩︎ Choosing pin, re-tune, re-baseline or defer within the timeline
      “In Anthropic’s testing, such a line didn’t stop Claude Fable 5.2 from thinking at medium and above.”
      ↩︎ Exam trap 4
      “such a line didn’t stop Claude Fable 5.2 from thinking at medium and above. Remove these lines and lower the effort level.”
      ↩︎ Checkpoint
    7. 7.
      “Effort is the only request parameter that controls thinking depth on Claude Opus 5.5.”
      ↩︎ Re-tuning prompts and re-baselining evaluations
    8. 8.

    Ready to test yourself?

    Practise Anthropic Claude Certified Architect – Professional (CCAR-P) - December 7 onwards in quiz mode.

    Spotted a mistake, or was something unclear? Tell us.