Detailed Analysis
A Reddit user's comparison of two Claude models in diagnosing a Subaru Forester XT's rough idle reveals a striking capability gap between Anthropic's Opus 4.8 and its newer Fable model, specifically around multimodal reasoning and agentic tool use. Facing a persistent cold-start misfire, the user uploaded a .mov video of the vehicle's instrument cluster to each model in succession. Opus 4.8 acknowledged the video format as a hard limitation it could not overcome and redirected the user to human experts — a Subaru shop, enthusiast forums like NASIOC, or community subreddits — while offering to help interpret text-based outputs such as scan data or spark plug photographs if they became available. The response, while polite and technically accurate about its own constraints, delivered no diagnostic progress.
Fable's approach to the identical problem was categorically different. Rather than treating the inability to directly "watch" a video as a terminal dead end, the model recognized that a sandboxed computational environment within the chat session gave it indirect access to the file itself. It used ffmpeg to extract still image frames for visual analysis and performed an audio waveform analysis on the engine sound, converting what appeared to be an insurmountable input limitation into a tractable multi-step inference problem. The frame extraction revealed cold-start conditions — a blue coolant light, fast idle decaying from approximately 1,700 to 1,550 RPM — while the audio analysis identified a strong amplitude modulation at 13.3–13.5 Hz, a frequency that corresponds precisely to one complete four-stroke engine cycle at the ~1,600 RPM displayed on the tachometer. Fable correctly identified this as the acoustic signature of a single underperforming cylinder rather than generalized carbon buildup across all four, and outlined a specific diagnostic pathway: per-cylinder misfire counts during cold start via OBD scan tool, followed by a coil-swap test to isolate the faulty component.
The technical substance of Fable's diagnosis reflects a meaningful advance in applied reasoning under uncertainty. The model was transparent about the limitations of its methodology — phone-microphone audio with automatic gain control is far from calibrated instrumentation — yet anchored its conclusion in a physically meaningful observation: a frequency peak that matched the engine's displayed RPM with sufficient precision to be non-coincidental. This kind of quantitative grounding, where the model ties an acoustic measurement back to fundamental engine mechanics, represents a more sophisticated form of diagnostic reasoning than pattern matching or symptom-list retrieval. The specific hypothesis — one weak coil, one fouled plug, one underperforming injector, or one heavily carboned intake valve — is clinically actionable and consistent with the symptom of a misfire that diminishes under warmer or higher-throttle conditions.
The episode illustrates a broader inflection point in AI model development: the shift from models that accurately self-report limitations to models that engineer around them. Opus 4.8's response was not wrong; it correctly identified that it could not process the video format and responsibly directed the user elsewhere. But Fable's response demonstrates an emerging class of behavior in which agentic tool use — accessing compute environments, running system utilities like ffmpeg, and chaining visual and acoustic analysis — transforms what counts as a tractable problem. The availability of sandboxed compute as an internal resource fundamentally changes the model's problem-solving surface area, allowing it to approach multimodal inputs through decomposition rather than rejection. This pattern, where models fluidly recruit available tools rather than stopping at the boundary of any single modality, is increasingly central to how frontier AI systems differentiate themselves in real-world utility.
For Anthropic, this user-generated comparison functions as informal but concrete benchmarking in a high-stakes practical domain — automotive diagnostics — where the gap between "I cannot help you" and a specific, physically grounded hypothesis carries tangible value. It also underscores the role of agentic architectures in driving perceived capability improvements independent of raw language model quality. Whether Fable's diagnosis ultimately proves correct will depend on the scan tool data the user is yet to collect, but the structural quality of the reasoning chain — from video frame extraction, through RPM-correlated frequency analysis, to a narrowed differential diagnosis — demonstrates that the ceiling for what AI assistance can accomplish in skilled-trades troubleshooting has moved substantially higher.
Read original article →