Claude Fable Access, Pricing and Limits

Fable did not end when its July subscription promotion ended. The model, the included allowance and the promotional credit were separate things, with separate conditions. As of October 4, 2026, Fable 5.1 is generally available, Fable 5 remains documented as a legacy option, and Mythos access still requires approval. The useful question is which surface you can use, how that surface bills you, and what happens when you reach a limit.

This ledger brings those answers together without turning old deadlines into current offers. It separates vendor announcements from mutable documentation and from gaps in the historical record. For an evaluation shortlist across vendors, use our maintained frontier-model selection guide. Here, the focus is the access and cost contract that a Fable decision depends on.

The current Fable and Mythos model pages identify the following API models. “Generally available” describes the product’s release status; it does not certify your account, region, organization policy or entitlement. A model ID in a price table is particularly weak evidence of Mythos access: its documentation expressly limits who may use it.

The API overview describes Mythos 5.1 through Project Glasswing. Anthropic’s September launch announcement also describes trusted access for cybersecurity and life sciences. Its September 17 life-sciences announcement then opens a beta for verified teams and institutions through the first-party API, Team and Enterprise surfaces. That announcement does not extend access to individual plans or third-party platforms; Mythos high-risk grants remain more narrowly vetted. Its September 30 correction says the linked form records interest, rather than serving as an application.

The Mythos product page still describes US-organization restrictions and future Mythos inclusion in the Cyber Verification Program. Do not infer worldwide or individual access from the existence of the newer life-sciences route. These documents describe different programmes at different dates. Obtain confirmation for the programme, geography and model you intend to use.

A deadline is not automatically an outage date, and a later help article is not an archived copy of an earlier offer. The ledger labels the distinction in each row. “UNVERIFIED” means the historical announcement was not recovered from a primary source in this check. It does not mean that the event did not happen. The linked site chapters preserve what was published at the time, without serving as independent proof.

The June 30 redeployment statement retains the original July 7 end of the included-use window. The later plan guide confirms a July 19 ending. Neither establishes the exact wording, publication time or scope of the intermediate extension announcement. That gap stays open instead of being filled with a reconstruction from our own articles or a social-platform summary.

The July chapter’s specific August 19 announcement terms remain unverified. A later official guide confirms the boost continued through September 13, followed by a different weekly baseline. That later record supports the final dates, without reconstructing each intermediate announcement.

The table records standard first-party token rates in US dollars per million tokens. Cache writes, cache reads and ordinary input are different billing categories. They should not be added as if every input token belonged to every category. Batch rates refer to asynchronous API processing, not a discount on a Claude subscription.

The cache-read rate falls by 75% between Fable 5 and 5.1. That percentage applies to one line item. Anthropic’s launch estimates of about 25% lower typical cost and up to approximately 45% for highly agentic work came from its own August usage at default effort, across Enterprise, Code and API workloads. They are vendor workload estimates, not a cut to input/output list prices or a forecast for your application. The launch methodology explains that distinction.

An example that keeps the token categories separate

Assume a workload bills 100,000 ordinary input tokens, 100,000 tokens as a five-minute cache write, 1,000,000 cache-read tokens and 20,000 output tokens. At the table’s rates, Fable 5 costs $1 + $1.25 + $1 + $1 = $4.25. Fable 5.1 costs $1 + $1.25 + $0.25 + $1 = $3.50. The $0.75 saving is about 17.65%, despite a 75% reduction in the cache-read rate.

This is worked billing arithmetic, not a measured agent run. It assumes the same billable token counts and successful cache hits for both models. With no cache reads, this rate change produces no saving. With more cached reuse, it matters more. A cache miss, extra reasoning or additional attempts changes the result; tool fees and human review sit outside this example.

The caching guide specifies a 512-token minimum cacheable prompt for these Fable models. Eligibility to cache does not prove a hit occurred. Inspect actual usage records and the write duration. The pricing guide also applies a 1.1× US-only inference multiplier to all token categories, including cache writes and reads, on the relevant first-party surfaces. The illustrated $3.50 becomes $3.85 under that modifier alone. Partner-operated cloud prices must be checked separately. For wider comparisons, see the API price index.

The current plan guide applies the following treatment to both Fable 5 and 5.1. The shared weekly percentage is a ceiling within an existing allowance. It is not additional usage on top of the plan, and it is not a fixed number of prompts.

Anthropic’s usage-credit guide says credits are billed separately from the subscription at API rates and combine Claude conversation and Code usage within that billing route. Spend caps and auto-reload affect what happens when funds run low. Purchased-credit terms can also depend on jurisdiction; the old promotional expiry is not a general rule for every credit balance.

The one-time Fable 5 promotion offered eligible Pro users $100 and eligible Team organizations $100 per purchased standard seat, capped at $2,500 and pooled across the organization. It began July 20, its claim window closed August 2, and the credits expired September 17, 2026, regardless of claim date. Fable 5.1 had no equivalent offer. Those dates are retained for interpreting old invoices and articles, not as an invitation to claim an expired credit.

Subscription allowance, context capacity and API rate limits answer different questions. A context window describes how much a model can process in a request. A weekly allowance constrains subscription use. API limits constrain an organization’s request and token traffic or spend. Increasing one does not establish that the others increased.

The Fable 5.1 and Mythos 5.1 API references document a 1,000,000-token context window and 128,000 maximum output tokens. These are model specifications, not a promise that every consumer conversation or Code session can use the entire window without product controls or compaction. The API limits guide ties limits to organization tiers and account history, distinguishes spend from rate limits, and warns that listed maxima are not guaranteed minimum capacity. Confirm the actual organization limit rather than copying a generic requests-per-minute figure.

The plan guide documents Claude Code version 2.1.170 or later for Fable 5 and 2.1.255 or later for Fable 5.1. Those are minimum client versions, not proof that an organization has enabled a model. The Code cost guide distinguishes allowance exhaustion, credit spend limits and context warnings. Match the remedy to the specific condition; a longer context setting cannot replenish credits.

Data handling is another access condition. The current Fable 5.1 documentation specifies 30-day retention for Fable and Mythos unless Anthropic expressly authorizes zero data retention. The EFS announcement describes a phased future rollout and interim ZDR for eligible customers. The captured evidence does not establish general EFS availability. A contract or explicit authorization must resolve your organization’s case; a general statement about enterprise privacy is insufficient.

Safeguards can also change which model answers. The September launch expands permitted Fable uses, but restricted work remains subject to fallback or refusal. API integrations must handle that outcome deliberately. Since September 24, pre-output refusals in the bio, frontier_llm and reasoning_extraction categories are billed at the model’s rates; other pre-output categories remain unbilled. Mid-stream refusals were already billed, and fallback credit is unchanged. The dated API release entry supersedes older blanket statements that pre-output refusals are free.

The official Code promotion record now says the temporary weekly boost ended September 13. From September 14, eligible Pro, Max, Team and seat-based Enterprise plans have weekly limits 25% above the pre-promotion baseline. The earlier boost was 50%; five-hour limits were unchanged. These are relative allowance levels, not published token quotas.

With the old baseline indexed to 100, the temporary level was 150 and the later level is 125. The change from 150 to 125 is a 16.67% reduction, rounded, while 125 is still 25% above 100. This arithmetic explains the apparently conflicting descriptions. It does not measure the amount of work a user can finish, and it is separate from Fable’s share of an included plan allowance.

The September release preserves base rates while changing API behavior. The what’s-new page and migration guide identify three changes that deserve explicit testing before a Fable 5 integration moves to 5.1:

1. Forced tool use: tool_choice values any and tool return a 400 error. The auto and none modes remain. A valid JSON schema does not by itself force a particular tool call.
2. Model-bound thinking: earlier models cannot read Fable 5.1 thinking blocks. A fallback to an older model can lose that reasoning; this is not equivalent to losing every visible message.
3. Conversation-bound thinking: editing the prefix before a Fable 5.1 thinking block can reject the request or drop the block under configured controls. Enforcement applies to accounts created on or after August 31, 2026; older accounts record mismatches and act when configured. The current guide says Mythos 5.1 does not run this prefix check.

These are documented integration constraints, not an implementation test performed for this article. Preserve a representative conversation when evaluating the migration: tool results, reasoning blocks, retries and the intended fallback. A successful first turn cannot establish that a later history edit or model switch works. The September 14 reporting addition helps distinguish a mismatch that was recorded from one that was enforced.

Fable 5.1’s API default effort is high; the launch announcement also identifies high in Claude Code and medium in Cowork and Claude chat. Keep the surface and effort in any cost comparison. Reusing the same label across products does not establish the same billable work. The dated September migration chapter remains useful for release context, while current documentation should decide the implementation details.

Use the current workload selection guide for candidates including Opus 5.5 and Sonnet 5.5. Their September arrivals make the July routing chapter historical. This ledger does not declare a quality winner or convert vendor benchmarks into a claim about your codebase.

Fable belongs in an evaluation when the task is difficult enough that better completion could justify its access conditions and total cost. Compare accepted outcomes, including retries and review, rather than a single answer or token rate. For repetitive work, test a lower-cost candidate first; for unresolved reasoning or long-running work, include Fable as an escalation candidate. Those are evaluation choices, not measured results.

These articles keep their original URLs and publication dates. They show the questions readers faced as access and competing models changed. Publication dates below come from each article’s metadata; they are not additional verification of its vendor claims. For current availability, return to the status and plan tables above.

A maintained page earns its place by absorbing new status information without erasing the earlier record. When a plan changes again, the current table should change and a new dated ledger entry should explain why. A correction to an old claim should retain its source boundary, rather than making an old article appear to have known a later decision.

This is a public-document reference. We collected first-party model pages, API and app release notes, pricing and support documentation, and dated Anthropic announcements. We retained the HTTP response bytes, decoded content and searchable text, then compared the claims used here with the detailed source. Publication labels, event dates and capture dates were recorded separately.

Mutable sources can disagree with their own summary pages. The biology announcement is dated August 7 although the Fable product index labels it August 6. The life-sciences announcement has an explicit September 30 correction. The June launch announcement now carries a July restoration update. The ledger preserves those distinctions; a current capture cannot reproduce the exact state a reader saw months earlier.

For applying this kind of record to a business workflow, our AI implementation services cover requirements, evaluation and operating-cost planning. The decision itself should remain traceable to the chosen surface, the relevant terms and evidence from the work you need completed.