Anthropic released a coding model called Fable for three days before the US government issued an export control order restricting access, citing concerns about model jailbreaking and weaponization. The author questions whether this intervention is genuinely about safety or a narrative tool to increase government regulation and reduce competition in the AI industry.
Fusion is a new dual-model architecture for Devin Desktop and CLI that pairs a frontier model for planning and review with a cost-effective model for execution, achieving up to 39% better efficiency on coding benchmarks. The system runs two parallel agents with separate contexts, allowing the lead model to maintain control while the sidekick handles implementation, avoiding the pitfalls of traditional model routing. Devin reports that using more expensive, token-efficient models can reduce overall costs by delegating effectively and maintaining prompt caches.
AI benchmarks like BioMysteryBench and Terminal-Bench are unreliable measures of model quality, with scores often failing to predict real-world performance or user preference. Inconsistencies between reported scores and public leaderboards, combined with frequent benchmark version changes, make these metrics misleading rather than useful for evaluating AI models.
Anthropic reduced Fable 5.1's cache read price to $0.25 per million tokens, making it cheaper than Opus 5 for reads but more expensive for writes and output. The analysis determines that Fable 5.1 becomes cost-effective compared to Opus 5 after approximately 108 API calls with sufficient cached context, with a break-even point of roughly 250 minutes for cache refresh intervals.
A developer successfully designed a PCB for an RP2350-based development board with an E-ink display using Claude (Fable 5) with minimal manual intervention, following strict rules of no manual edits. The 31.8 × 37.32 mm, 4-layer board cost €26 per unit, though the design process revealed component package mismatches and routing challenges that were resolved through iteration with Claude.
Claude Fable 5.1 solved the Cyphral Distich, a 370-year-old cipher by Sir Thomas Urquhart that had remained unsolved for centuries. The solution uses the 32 Proquiritations in Urquhart's text as a key, where each number in the cipher line indexes a word whose first letter reveals the hidden message: a royalist prayer for King Charles II. Fable 5.1 also deciphered Urquhart's larger Cyphral Octastich using the same principle with numbered pages.
A user discusses their preference for Claude as their default AI model, noting that while they initially used Fable, they found it consumed credits too quickly for mobile app planning tasks. They have settled on Opus 4.8 as their go-to model for web and mobile development with cloud backends, citing it as sufficiently capable for their needs.
RTK, a popular terminal output compression tool with 79k GitHub stars, claims to reduce AI coding token costs by up to 60–90%, but benchmarking by JetBrains's SkillsBench and independent testing on Terminal-Bench 2.1 found minimal or negative savings. Testing across 1,740 attempts showed costs fell 5% for Fable but rose 5% for DeepSeek; RTK's reported token savings miscount removed output rather than actual billing impact and can mask increased costs from additional model turns.
Anthropic's AI Data Retention Tracker maps five commercial Claude model-and-tier combinations, finding that four retain user prompts by default. Claude Fable 5.1 in GitHub Copilot requires data retention to operate safety classifiers, with zero-data-retention access limited to approved enterprises through end of 2026. Anthropic's retention policies vary by surface: Enterprise conversations retain data indefinitely by default, Covered Models under Business Associate Agreements require 30-day retention with no ZDR option, and most other Copilot models operate under zero retention.
Cognition unveiled SWE-2, an advanced coding model achieving 50% on FrontierCode benchmarks at 64% lower cost than Fable 5.1, using scaled reinforcement learning in the multi-trillion-parameter regime. The model demonstrates improved efficiency by requiring fewer exploration steps and better engineering judgment compared to its predecessor SWE-1.7.