Detailed Analysis
A Reddit user on Anthropic's Max plan reported an alarming pattern of rapid usage-limit exhaustion, claiming to hit caps within two or three prompts even in brand-new conversations. The poster attributes this to anomalous cache-token consumption tied to "Fable 5" (an apparent internal or codenamed model reference distinct from "Opus 4.8"), citing a stark discrepancy of 54 million cache tokens consumed versus essentially zero for comparable usage on the other model. When the user contacted Anthropic support, they say representatives declined to investigate the root cause, refused escalation to a human agent, and instead repeatedly suggested purchasing additional credits — credits that would presumably be consumed by the same unexplained caching behavior. Support also reportedly stated unequivocally that no compensation or limit resets would be offered.
The core grievance here is less about pricing and more about accountability and transparency in metered AI consumption. Prompt caching is a legitimate and normally cost-saving mechanism — it lets models reuse previously processed context (like system prompts or long conversation histories) instead of reprocessing everything from scratch, which should reduce token costs for the user. If a bug or misconfiguration causes cache reads to balloon unnecessarily, users can be billed or rate-limited for token activity they neither initiated nor can observe directly, since the internals of caching are opaque to the end user. Because Claude Code and the Claude apps meter usage against subscription limits, a backend caching regression can effectively make a Max plan (which costs significantly more than base tiers) feel like it delivers a fraction of its advertised value overnight.
This complaint fits into a broader pattern of friction between AI vendors and power users as usage-based and hybrid subscription-metering models mature. Anthropic, like OpenAI and Google, has increasingly tied product access to consumption limits calculated via opaque internal accounting (tokens, cache reads/writes, context windows), and when something goes wrong on the backend, customers often have no visibility into diagnostics and no clear recourse beyond canned responses from tiered support systems optimized for upsells rather than troubleshooting. This mirrors recurring criticism across the AI industry that customer support for infrastructure-level products has not kept pace with the complexity of the systems being sold, leaving developers and heavy users to crowdsource incident detection on forums like Reddit rather than through vendor status pages or transparent postmortems.
The specific mention of "Fable 5" is notable, as it does not correspond to any publicly announced Anthropic model name as of mid-2026, suggesting either an internal codename surfaced through logs/UI, a beta/experimental model being tested with select users, or possibly a mislabeling artifact within Claude Code's telemetry. If real, it would imply Anthropic was routing some traffic through a new or experimental model whose caching behavior was not properly optimized before wider rollout, a scenario reminiscent of past incidents in the industry where silent model swaps or A/B testing led to inconsistent performance and cost characteristics without adequate user disclosure. Whether this reflects a genuine bug, a rollout of new infrastructure, or friction from evolving pricing/caching mechanics, the underlying issue — support systems unable or unwilling to explain unusual billing/usage behavior — represents a recurring pain point that could erode trust among high-usage subscribers if left unaddressed.
Read original article →