Summary
Fahd Mirza tests Anthropic’s Fable 5.1 against Volayer, a non-trivial trading desk simulator managing an $8.5 million portfolio that streams live market headlines and updates portfolio value based on trading decisions. The core bug: large macro events like a CPI miss or chip export ban barely move the portfolio while minor news produces outsized swings — a logic inversion that would actively mislead real investors. Fable 5.1, accessed through the Hermes agent framework with no hints about where the bug lives, identifies and resolves a core inversion bug plus four additional issues including a connection leak and a validation gap, validating the fix with targeted automated tests.
Mirza covers Fable 5.1’s pricing structure in detail: $10 per million input tokens and $50 per million output tokens, with cache reads now 75% cheaper — translating to up to 45% cost savings for heavy agentic workloads that cycle through cache-heavy agent loops. Benchmark figures cited include Fable 5.1 clearing 50% on scientific research tasks versus Fable 5’s sub-30% ceiling, and 52.6% on the same task versus Opus 5’s 29%.
The video closes with a multimodal multilingual test: Fable 5.1 is given an image of a bar scene and asked to generate contextually appropriate pickup lines in dozens of languages, each in the correct native script. Mirza finds the model falls short on this specific challenge, making the video notable for documenting both a clear strength (complex agentic code debugging) and a concrete limitation (multilingual multimodal generation) rather than presenting one-sided praise.
📺 Source: Fahd Mirza · Published September 01, 2026
🏷️ Format: Hands On Build
