Detailed Analysis
A Reddit post from a Claude Max 20 subscriber highlights an emerging friction point in Anthropic's model lineup: the gap between "Fable" (apparently a newer, more advanced coding-capable model or mode) and the existing Opus 4.5/5 tier has grown large enough that some power users find Opus effectively unusable for their most demanding work. The poster describes attempting to have Opus review, refactor, and rewrite a codebase spanning tens of thousands of lines, providing extensive scaffolding — inline documentation, multiple architecture documents, and continuous guidance — only to find that Opus repeatedly misreads the code, loses critical context, and requires constant intervention. By contrast, Fable is described as capable of dropping into the same codebase cold, with zero prior context, and immediately grasping its structure and intent. The user frames this not as a stylistic mismatch but as a raw capability gap: the architectural complexity of the work exceeds what Opus can reliably handle, regardless of how much hand-holding is provided.
This experience report is notable because it inverts Anthropic's stated product positioning. The company has apparently marketed Fable as suited for "demanding tasks" while positioning Opus and Sonnet as adequate for everything else — implying a graceful degradation across the model tier, where users pick the right tool for the job. The poster's real-world usage suggests something closer to a cliff edge rather than a gradient: for complex, large-scale codebases, Opus doesn't just perform somewhat worse, it becomes nearly non-functional as an autonomous agent, forcing the user into a "junior apprentice" management style rather than genuine delegation. That distinction matters a great deal for how subscribers value and use their plans. An "agent" in the sense Anthropic and competitors are selling implies a system that can be trusted with meaningful autonomy; a system requiring repeated manual intervention and steering isn't functioning as an agent at all, no matter how sophisticated its outputs might be in isolated, smaller tasks.
The practical consequence raised is an economic one specific to Anthropic's usage-based Max plan structure. Because the poster's real workload requires Fable specifically, and Fable has its own separate usage cap within the same subscription, they exhaust their entire Fable allowance while leaving their Opus/Sonnet allowance almost completely untouched — landing around 50% total utilization of their Max 20 plan by the end of the weekly window. The user's proposed fix — a "Fable-only" plan with a higher cap on the higher-capability model rather than a blended allocation across tiers — reflects a broader tension in how AI companies price differentiated model capability. When a lower tier isn't merely "cheaper" but qualitatively unusable for a customer's core use case, blended plans effectively force that customer to subsidize other users' access to lower-tier models they don't personally use, which is precisely the complaint the poster raises.
This dynamic sits within a larger industry-wide pattern: as frontier labs like Anthropic, OpenAI, and Google roll out increasingly differentiated model tiers (flagship vs. cost-efficient, "thinking" vs. fast modes, specialized coding models vs. general assistants), the gap in real-world capability between tiers is becoming a more visible source of user frustration and demand for pricing granularity. Coding in particular has become a proving ground for these gaps, since large, long-lived codebases stress-test a model's context retention, architectural reasoning, and ability to operate with genuine autonomy rather than surface-level pattern matching. As agentic coding tools become central to how developers evaluate AI subscriptions, expect continued pressure on providers to either narrow the capability gap between tiers or restructure pricing so that heavy users of top-tier models aren't penalized by allocation schemes built around an assumption of even usage across the model portfolio.
Read original article →