Anthropic Fable 5, Code Names Capybara and Numbad, and US AI Licensing Regimes

In June 2024, Anthropic publicly released its new AI model, Fable 5 (internally code-named Capybara), which was temporarily shut down by the US government on June 12th due to concerns over "defense-oriented prompting"—a technique common across frontier models including GPT 5.5 and Kimmy. Following a pressure campaign and negotiation by Anthropic co-founder Tom Brown (not Dario Amodei), Fable 5 was restored on June 30th, featuring an upgraded safety classifier (blocking the flag technique in over 99% of attempts and silently rerouting flagged requests to Opus 4.8) and new pricing: $10/million input tokens, $50/million output tokens, over three times previous costs and double Opus 4.8. The model exited the subscription model, imposing a 50% weekly usage cap.

Internal leaks revealed Capybara (now shipped as Mythos 5 and Fable 5) and Numbad (still unreleased), clarifying major media confusion. Leaks in late March mapped Capybara to Mythos 5 and Fennec to Opus 4.6; Numbad remains a future candidate. A June 21 analyst report (Andrew Curran) claimed an even more capable Mythos model had finished training, possibly Mythos 5.1 or Mythos 6, though unverified by major outlets.

A broader theme emerged: US government imposed a new AI licensing regime in early June, impacting all major labs (Anthropic, Google, Microsoft, XAI), not just Anthropic. Sam Altman publicly objected, suggesting the risk of government "picking winners on access." Dario Amodei ironically published an essay advocating for government authority to block unsafe frontier models, only for Anthropic's offering to be halted days later. The saga exemplifies that frontier AI subscriptions are "a temporary contract," vulnerable to sudden caps, bans, or pricing changes.

Benchmark data: Sol Ultra (OpenAI CT 5.6 SALL) led with 91.9% on terminal bench, Mythos 5 at 88%, Opus 4.8 at 78%. Sonnet 5 launched June 30th with a 1 million token context window and 128k max output, but scored only 53 on the AI Intelligence Index (below Opus 4.8), signaling it's for different, less demanding tasks. Fable 5 topped the Remote Labor Index at 16.1%. DeepSeek V4 Pro's token pricing ($0.87 per million output) is 57x cheaper than Fable 5 ($50 per million).

Confirmed leaks: March 26th, 3,000 Anthropic internal files revealed Capybara, Fennec, Numbad; Capybara (Mythos/Fable 5) termed a "step change". Claude now writes 80% of Anthropic's internal code and accelerated its training code 52x over Opus 4 baseline. Chinese labs (Deepseek, Moonshot, Minimax, Quen) used 24,000 fraudulent accounts to extract 16 million exchanges from Claude. Undercover mode (Claude never reveals code names) was confirmed in March 31st leaks.

For future-proofing, the transcript recommends running high-volume tasks on Sol or Terra APIs, using Opus 4.8 for tasks needing remote labor performance, and considering local hardware for sensitive workflows—even citing Nvidia DGX Spark, Framework Desktop (Ryzen AI Max Plus 395), and Apple Max Studio, each with detailed specs and business-case break-even math. Local inference has narrowed the gap with frontier cloud models, and Vitalik Buterin's public preference for local, sovereign setups is gaining traction.

Anthropic's fall 2026 IPO window creates pressure for new releases; Polymarket places the probability of a 2026 IPO at 76%. Gemini 3.5 Pro lands July 17, while GPT 5.6 general release is imminent.