Ahab

Posted on Sep 2 Originally published at indieseek.co

Claude Fable 5.1 GitHub Copilot Data Retention and Pricing Checklist

Claude Fable 5.1 in GitHub Copilot: verify retention, policy, and token cost before rollout

Quick answer

GitHub announced Claude Fable 5.1 as generally available inside GitHub Copilot on September 1, 2026. That does not establish broad availability through Anthropic's API, Claude apps, or every third-party model router. GitHub says rollout is gradual, the model is available to Copilot Pro+, Max, Business, and Enterprise users, and Business or Enterprise administrators must enable a policy that is off by default.

The unusual gate is data handling. GitHub's current model documentation says Anthropic retains prompts and outputs for Claude Fable 5 and 5.1 by default to operate safety classifiers. Certain eligible enterprise customers can request a time-limited zero-data-retention path through the end of 2026, but approval does not enable the model and it carries an internal-use boundary. Decide the retention lane before the model policy, then canary cost, availability, fallback, and audit evidence.

Who this is for

This checklist is for individual developers comparing premium Copilot models and for Business or Enterprise administrators deciding whether Fable 5.1 belongs in coding agent, CLI, IDE, mobile, or github.com workflows. It is especially relevant when repositories contain customer code, regulated data, confidential prompts, or long-running agent tasks that can generate substantial output.

If your first problem is inherited model access, start with the Copilot model-availability reconciliation checklist. For interrupted CLI sessions, use the Copilot CLI permission and restore checklist. Those controls do not replace this model-specific retention decision.

What changed—and what did not

GitHub describes Fable 5.1 as an Anthropic Mythos-class model designed for long-horizon autonomous coding and knowledge work. In Copilot, the model picker spans Visual Studio Code, Visual Studio, Copilot CLI, coding agent, the GitHub Copilot app, github.com, GitHub Mobile, JetBrains IDEs, Xcode, and Eclipse. Availability can still lag during gradual rollout.

The announcement is not a promise that Copilot Free or Pro includes the model. It is also not evidence that an organization policy, ZDR approval, or account-team approval has already taken effect. Treat plan entitlement, model visibility, admin policy, retention route, and billing as five separate states.

Choose the retention lane first

Lane	Entry gate	Data boundary	Safe use
Pro+ or Max	Eligible individual plan and completed rollout	Default Anthropic retention for prompts and outputs	Non-sensitive canary, then normal work only after accepting the policy
Business or Enterprise, default	Admin enables Fable 5.1 after organizational review	Default retention applies	Repositories approved for this data path
Eligible enterprise ZDR exemption	GitHub account team confirms eligibility and configuration; admin still enables the model	ZDR through the end of 2026 under the documented exemption	Internal operations, including internal product development and evaluation
Free, Pro, or not-yet-rolled-out account	No documented entitlement or model not visible	Not applicable	Keep the current approved model; do not infer an outage

GitHub says the time-limited ZDR terms prohibit making the model endpoints or outputs externally available. GitHub Support cannot determine eligibility or turn this access on. After the exemption, the documented direction is Enterprise Frontier Safeguards; that future state needs a fresh policy review rather than an automatic rollover assumption.

Price the workload, not just the subscription

GitHub's current pricing table is per one million tokens. Fable 5.1 is listed at $10 input, $0.25 cached input, $12.50 cache write, and $50 output. Anthropic models charge cache write in addition to cached input. Copilot plans include different AI-credit allowances; usage beyond allowances is billed through GitHub AI Credits, where one credit equals $0.01.

Use a reference estimate before rollout:

estimated_usd =
  input_tokens / 1_000_000 * 10
  + cached_input_tokens / 1_000_000 * 0.25
  + cache_write_tokens / 1_000_000 * 12.50
  + output_tokens / 1_000_000 * 50

A long task with 200,000 input tokens, 100,000 cached-input tokens, 20,000 cache-write tokens, and 40,000 output tokens has a $4.275 reference cost before plan allowances. Output and cache writes dominate quickly. Set a task budget and stop condition before comparing quality.

A six-stage rollout workflow

1. Classify the account and workload

Record plan, billing owner, organization and enterprise, repository, surface, data classification, expected token shape, and whether the output stays internal. Reject the rollout if nobody owns retention acceptance or overage spend.

2. Capture the effective policy

For Business and Enterprise, record the enterprise and organization settings, repository access, and the final Fable 5.1 policy state. “Allowed by enterprise” is not the same as “enabled by organization.” Preserve a screenshot or exported setting with timestamp and owner.

3. Resolve the retention path

Choose default retention or a confirmed ZDR exemption. Do not label a request, support ticket, or account-team conversation as approved configuration. For ZDR, retain the eligibility evidence, expiry date, internal-use restriction, enabled policy, and a re-review owner for the end of 2026.

4. Run a non-sensitive canary

Use a synthetic repository with no secrets or customer code. Verify the exact model label, surface, policy, request completion, attribution, audit log, and fallback behavior. If the model is absent, confirm gradual rollout and policy inheritance before retrying; do not widen repository access.

5. Measure cost and quality together

Run the same bounded task on the current approved model and Fable 5.1. Compare accepted changes, tests, review corrections, input, cached input, cache writes, output, wall time, and estimated credits. A stronger result is not economical if the output loop remains unbounded.

6. Expand gradually and keep rollback boring

Start with one low-risk repository and one surface. Roll back by disabling the policy or returning users to the prior approved model. Stop on unexplained model routing, missing audit evidence, retention mismatch, unexpected external output, or spend beyond the declared threshold.

Copyable acceptance record

model: Claude Fable 5.1
product_boundary: GitHub Copilot
plan: Business
surface: Copilot CLI
policy_owner: platform-admin@example.com
policy_state: enabled
retention_lane: default-retention
repository_class: internal-non-sensitive
canary_repo: synthetic-fable-canary
estimated_usd_limit: 5
fallback_model: prior-approved-model
audit_evidence: captured
rollback_owner: platform-admin@example.com
review_date: 2026-10-02

Building something? Take a 60-sec game break. Score to rank your product or profile on tapto.top and get more exposure—free, no signup.

Common mistakes

Writing “Claude Fable 5.1 is available everywhere” when the verified release is GitHub Copilot-specific.
Enabling the model before deciding whether default retention is acceptable.
Treating ZDR eligibility approval as automatic model enablement.
Ignoring cache-write and output prices when estimating long agent tasks.
Calling a missing picker entry an outage before checking gradual rollout, plan, and policy.
Sending customer code through a synthetic canary or exposing ZDR outputs externally.

FAQ

Is Claude Fable 5.1 generally available outside GitHub Copilot?

This source set proves GA in GitHub Copilot only. Check the official product or API you intend to use before claiming wider availability.

Does Copilot Business or Enterprise enable it automatically?

No. GitHub says the policy is off by default and an administrator must enable it. Rollout is also gradual.

Is zero data retention automatic for enterprise customers?

No. Certain eligible enterprises can request the time-limited path through their GitHub account team. Approval and policy enablement remain separate.

Is the token price the same as the subscription price?

No. Plans include different allowances, while the model table supplies per-token reference rates used for additional AI-credit billing.

Sources

Originally published at IndieSeek.

DEV Community