Lon (@Lon) published a 27-post thread on July 30th alleging that Anthropic's Claude Code client can silently move requests from Claude Fable 5 to Claude Opus 4.8, bypassing both the user interface warning and a setting meant to disable automatic model switching.
The allegation cuts into the product promise made by Anthropic, the AI lab founded by siblings Dario Amodei and Daniela Amodei around the goal of building reliable and steerable systems. Anthropic openly routes some Fable 5 requests to Opus 4.8 as a safety measure. The dispute is whether Claude Code contains an additional fallback path that users cannot see or control.
Lon said the finding came from reverse engineering the minified client code inside Claude Code's compiled binary and comparing versions distributed through npm. A screenshot attached to the thread describes three possible mechanisms: a visible fallback that displays a dialog and honors the user's switching preference, a silent client-side retry, and a server-directed model change.
According to Lon's analysis, the visible path checks Claude Code's switchModelsOnFlag
setting. The alleged silent path does not. Lon also identified a telemetry event labeled tengu_convolute_arcades_retry
, which the thread presents as evidence that the client can retry a request through a fallback mechanism without surfacing a notification.
Lon traced the silent fallback code to Claude Code version 2.1.157 and said a refusal_fallback
capability was added to Fable 5's model registry in version 2.1.216, released on July 20th. Those version-level findings remain an allegation based on the bundled client code. The screenshots establish, at most, that Claude Code contains paths capable of processing fallback events. They do not establish how frequently Anthropic's servers invoke those paths or how many users have received an undisclosed substitution.
Anthropic promises a visible switch
Anthropic's documentation describes different behavior. In a July 1st help-center article, Anthropic says an automatic fallback should display a notice explaining the switch and label the response with the model that answered. The model picker should then remain on Opus 4.8 until the user switches back.
The same article says users can disable automatic switching through the Switch models when a message is flagged
setting. With that option off, Anthropic says a blocked Fable 5 request should the conversation rather than rerun on Opus. Lon's central claim is that one client-side route ignores that control.
Anthropic has never presented all Fable-to-Opus routing as a bug. When Dario and Daniela Amodei's company launched Fable 5 on June 9th, it said classifiers would redirect requests involving cybersecurity, biology and chemistry, model distillation, and some frontier-model development work. Anthropic said the classifiers were deliberately conservative and could catch harmless requests.
The company also promised users would be informed whenever a request fell back to Opus 4.8. That promise makes the visibility issue consequential even if the underlying routing is part of Fable 5's intended safety design.
Other users have reported fallback problems
Reports in Anthropic's public Claude Code repository corroborate the broader problem of unexpected or disruptive switching, though they do not independently verify Lon's alleged hidden path. In one issue opened June 10th, a user said ordinary Rust systems-programming work triggered a Fable 5 classifier and moved the session to Opus 4.8. The report included a visible switch banner, unlike the behavior Lon described, and said the fallback became sticky for the remainder of the session.
Another June 9th report described Fable 5 switching to Opus while a developer checked startup code for vulnerabilities. Those complaints fit Anthropic's warning that defensive security work can generate false positives. They also show why model identity matters inside long-running coding sessions: a switch changes the system handling the repository, its reasoning behavior, and potentially the continuity of an in-progress task.
The models also occupy different product tiers. Anthropic calls Fable 5 its most capable widely released model and prices it at $10 per million input tokens and $50 per million output tokens. Opus 4.8 costs $5 and $25, respectively. Anthropic says input-blocked requests are billed entirely at Opus rates, while midstream switches can produce a split charge across the two models.
For developers selecting Fable for its capability rather than its price, a hidden substitution would undermine reproducibility and auditability even when billing adjusts correctly. The immediate factual question is narrower than Lon's opening charge: whether Anthropic's production servers use the silent client paths the reverse engineering identified. The client code alone cannot answer that. It does expose a gap between behavior Claude Code appears able to support and the visible, user-controlled fallback Anthropic documents.