Claude users report silent fallback and 30-day retention after Fable 5 launch

• Anthropic shipped claudeai's launch post as a general release of Fable 5, while claudeai's Mythos 5 post said the restricted Mythos tier uses the same underlying model with some safeguards lifted.
• The buried launch detail was a hidden downgrade path: Hangsiin's system card excerpt said frontier-LLM-development requests could be limited with prompt modification, steering vectors, and PEFT, while Official launch notes said some risky queries would fall back to Opus 4.8.
• After complaints from users like GergelyOrosz's launch-day thread and haider1's API bill post, ClaudeDevs' apology said flagged requests will now visibly fall back to Opus 4.8 and API refusals will include a reason.
• Anthropic also changed the data terms: Official launch notes require 30-day retention for Mythos-class traffic, and the AWS Bedrock model card says Fable 5 requires provider_data_share with no launch-day console UI.
• The docs expose a second billing wrinkle: the AWS fallback credit page exists because a refused Fable 5 request retried on Opus 4.8 would otherwise re-pay cache-write costs.

You can read Anthropic's launch post, inspect the AWS Bedrock model card, and then compare that paperwork with Hangsiin's system card excerpt, ClaudeDevs' reversal, and the Hacker News launch thread. The weird part was not that Fable had safeguards. It was that Anthropic launched a flagship model with hidden fallback behavior, mandatory retention, and a retry-credit document waiting in the wings.

Anthropic framed Fable 5 as the generally available version of its new Mythos-class model. According to claudeai's launch post, it was the most capable model the company had made broadly available.

The split mattered immediately:

• claudeai's availability post said Fable 5 was available everywhere on day one.
• claudeai's Mythos 5 description said Mythos 5 shared the same underlying model, but with some safeguards lifted.
• claudeai's availability post limited Mythos 5 to Glasswing partners at launch.
• claudeai's access-program note said Anthropic planned a broader trusted-access program for defensive cyber and biomedical work.

Anthropic's own setup guide added a few model-level changes that help explain why users noticed the release so fast. ClaudeDevs' getting-started thread said thinking is always on, effort controls are still available, and older Claude prompts or skills may now be too prescriptive.

The launch post disclosed fallback for cyber, bio, chemistry, and distillation topics. The sharper detail came from the system card excerpt that Hangsiin pulled out: frontier-LLM-development requests could be silently limited through prompt modification, steering vectors, and PEFT, and Anthropic estimated that path would touch about 0.03% of traffic.

That 0.03% estimate collided with a much broader-looking field reaction. GergelyOrosz's SemiAnalysis example claimed SemiAnalysis had already triggered the safeguard, while eliebakouch's critique argued the public evidence for the classifier threshold was thin and the scope of affected work was unclear.

The credibility problem was simple. Users were paying for Fable 5 without an obvious way to tell whether a given response came from the full model, a steered version, or Opus 4.8 instead.

Anthropic reversed course less than two days later. In ClaudeDevs' apology, the company said invisible safeguards were the wrong tradeoff and that flagged requests would visibly fall back to Opus 4.8, matching its existing cyber and bio behavior.

That change landed after a pile of concrete complaints:

• GergelyOrosz's thread objected to both silent downgrades and retention.
• theo's coding-session post said Fable degraded three-quarters of the way through a real auth feature pass.
• haider1's API-bill post described burning 430,000 tokens and $27.40 before getting redirected to GPT-5.5.
• dylan522p's SemiAnalysis report said frustrated power users tried Codex after nonsensical refusals.

Anthropic's official line now matches the support flow: ClaudeDevs' apology said API refusals will return a reason, and the appeals article gives separate reporting paths for Claude Code, Claude surfaces, and API requests.

The other launch-day shock was contractual, not model-behavioral. The official launch post says Anthropic now requires 30-day retention for Fable 5, Mythos 5, and future models with similar or higher capability levels across both first- and third-party surfaces, while also saying the data will not be used to train new Claude models.

That clause spilled directly into infrastructure docs. The AWS Bedrock model card says Fable 5 has a 1 million token context window, 128K max output, always-on adaptive thinking, and a required provider_data_share retention mode. It also says there was no console UI for that setting at launch.

Community reaction focused on the compatibility break. scaling01's retention post argued zero-retention customers were effectively locked out, and GergelyOrosz's open-models post said the combination of stricter usage controls, less transparency, and stored prompts pushed him toward open models and local inference.

The cleanest sign that Anthropic expected refusal-and-retry loops may be the AWS paperwork, not the tweets. The AWS fallback credit page documents a beta token, fallback-credit-2026-06-09, that prevents users from paying cache-write charges twice when a refused Fable 5 request is retried on another model such as Opus 4.8.

That page adds mechanics the launch thread never mentioned:

• Fable 5 refusal responses can issue a fallback_credit_token.
• The retry has to match the original system, messages, and tools, or append the refused partial output as a continuation.
• The token expires after five minutes.
• Without the beta flag on both calls, the token is rejected.

AWS also surfaced a related community complaint fast enough to spawn its own discussion thread. The Hacker News Bedrock thread centered on the requirement to share data with Anthropic for Mythos-class models, which put the retention policy in front of infrastructure teams that might never read a 300-page system card.