Anthropic: Claude Fable 5.1 (batch)

Review this Python function for potential race conditions. We're seeing intermittent failures in production under high load, but unit tests pass consistently.

The function modifies a shared cache dictionary without synchronization. Under concurrent access, two threads could read the same missing key simultaneously, both compute the expensive result, and both write back—wasting resources. More critically, if the cache uses a custom `__setitem__` that isn't atomic, you risk corrupting internal state. The unit tests pass because they run single-threaded. Consider using `threading.Lock` around the read-check-write sequence, or switch to a thread-safe structure like `functools.lru_cache` if the use case fits.

Claude Fable 5.1's 1M token context window allows pasting entire codebases for holistic analysis. This example shows its strength in connecting symptoms (intermittent production failures) to root causes (concurrency bugs invisible in tests). The batch pricing model ($5 input / $25 output per Mtok) makes it cost-effective for large-scale code review workflows, though the async nature means it's unsuitable for interactive debugging.

Summarize the key regulatory changes in these 47 PDF pages of EU AI Act amendments, focusing on obligations for general-purpose AI providers.

The amendments introduce tiered obligations based on systemic risk classification. Providers of general-purpose AI models must now: (1) maintain technical documentation on training data provenance and bias mitigation, (2) implement incident reporting within 72 hours of detecting serious malfunctions, (3) conduct annual third-party audits if monthly active users exceed 10 million. The most significant change is Article 52a, which mandates watermarking for synthetic content—text, image, and audio—with cryptographic traceability. Penalties scale to 6% of global revenue for non-compliance.

The model's file ingestion and 1M token capacity handle lengthy regulatory documents without chunking. This example demonstrates its ability to extract domain-specific requirements from dense legal text and present them in actionable terms. However, batch processing means you can't iteratively refine the summary in real-time—you'd need to submit a new job with clarifying instructions.

Generate a patient-friendly explanation of this radiology report. The patient is a 34-year-old with no medical background who's anxious about the findings.

Your chest X-ray shows a small area of increased density in the lower left lung. This often happens after a respiratory infection—your immune system is still cleaning up. The radiologist noted it's stable compared to last month's scan, which is reassuring. They've recommended a follow-up in 8 weeks to confirm it continues resolving. There's no sign of anything urgent. The 'ground-glass opacity' term sounds alarming, but it just means the area looks hazy rather than solid—think of it like fog versus a cloud. Your doctor will discuss whether you need antibiotics or just rest.

Claude Fable 5.1's training includes medical literature, enabling accurate translation of clinical jargon into lay terms while preserving nuance. This example highlights its empathetic tone calibration—acknowledging anxiety without being patronizing. The image input modality would allow direct analysis of the X-ray alongside the report. The main limitation: batch processing introduces latency unsuitable for urgent clinical communication.

When batch processing cuts your document ops cost by 80%

A 12-person legal ops team processes 2,000 vendor contracts monthly, tagging each by risk tier and extracting key dates. Running this synchronously during business hours with standard Claude costs $180/month at current volumes. Switching to Claude Fable 5.1 batch drops that to $36/month—same 1M token context window, same multimodal file handling, but you queue the job at 5pm and pull tagged results at 8am. The catch: batch jobs can take up to 24 hours to complete, so this only works if your SLA is next-day or slower. If you need answers in under 4 hours, pay the synchronous premium. Otherwise, route all non-urgent classification and extraction work here and bank the difference.

How a 6-person finance team automated board decks for $12/month

A startup's finance lead spends 8 hours each month turning Stripe exports, bank statements, and investor updates into a formatted board deck. The workflow ingests 40-60 pages of mixed PDFs and CSVs, then outputs a 12-slide narrative with variance commentary. At 150k tokens in and 25k tokens out per run, standard Claude costs $1.38 per deck. Fable 5.1 batch brings that to $0.28—under $4/year if you run it quarterly. The model's 1M token window handles the full document set in one call, and the 24-hour batch window fits a month-end close cycle perfectly. This is the poster child for batch: infrequent, high-token jobs where overnight turnaround beats real-time cost by 5x.

When to batch 500 support tickets into one weekly insight report

A 20-person SaaS support team closes 500 tickets weekly and wants a single executive summary of themes, bugs, and feature requests. Each ticket averages 800 tokens; the summary outputs 3,000 tokens. Running this Friday afternoon as a batch job costs $2.75/week versus $13.75 synchronous—a $572 annual difference. The trade-off: if your exec needs the report Monday morning, you're fine. If they want it Friday at 4pm for an end-of-day standup, batch won't land in time. The model's image and file support also means you can throw in screenshot attachments without preprocessing. Route this to batch if your reporting cadence tolerates a 12-24 hour lag; stay synchronous if leadership wants same-day insights.