Verdict: Anthropic is preparing to re-release Claude Fable 5 under a new metered "Usage Credits" model. Leaks from the latest Claude Code binary (v2.1.190) show a pivot from flat-rate subscriptions to a pay-as-you-go system for Mythos-class models, likely to offset extreme compute costs and navigate ongoing export restrictions.
Last verified: 2026-06-25 · Current Status: Leaked (v2.1.190) · Top Alternative: GPT-5.5-Cyber · Primary Concern: Export controls & usage caps. Pricing and availability for Fable 5 are volatile and change weekly.
Why did Anthropic pull Fable 5 in the first place?
Anthropic initially released Claude Fable 5 on June 9, 2026, but quickly restricted it. While the model set new records on the SWE-Bench Pro (80.3%), it was pulled from general availability due to its extreme capabilities in cybersecurity and complex reasoning.
Unlike the standard Sonnet or Opus models, Fable 5 is a "Mythos-class" model. This tier has been under heavy US export control scrutiny, with the government fearing that such high-reasoning models could be used by foreign nationals to develop autonomous cyber-weapons. As a result, Anthropic implemented "Project Glasswing," a restricted access program that limited the underlying Mythos model to vetted infrastructure providers and cyber-defenders.
What the Claude Code 2.1.190 leak reveals
The most recent update to Claude Code (v2.1.190), released on June 25, contains explicit string updates that confirm Fable 5 is nearing a return. The key change is the transition from "included in your plan" to a metered credit system.
New Leaked Strings include:
- "You have used your included Fable 5 usage for the week. Continuing on Fable 5 uses usage credits."
- "Fable 5 was included in your plan for a limited time and now requires usage credits."
This suggests that even "Max" and "Team" subscribers ($200/mo+) will no longer have unlimited access to Fable 5. Instead, Anthropic is treating it as a premium add-on, similar to how Claude Design usage is metered separately. This shift aligns with the high API costs ($10 per million input / $50 per million output tokens) reported in early June.
Fable 5 vs. GPT-5.5-Cyber: The Battle for Security
While Anthropic hesitates over safety, OpenAI has charged ahead with GPT-5.5-Cyber. This has created a paradoxical situation in the market: Anthropic's Fable 5 is restricted because it might be dangerous, while OpenAI's Cyber-specific model is publicly topping the CyberGym benchmark at 85.6%.
| Capability | Claude Fable 5 | GPT-5.5-Cyber |
|---|---|---|
| SWE-Bench Pro | 80.3% (State-of-the-Art) | ~58.6% |
| CyberGym Score | Restricted/Classifier Routed | 85.6% (Top) |
| Vision/Game Play | Completed Pokémon FireRed | Specialized in Codebases |
| Availability | Metered Credits (Leaked) | Public (Daybreak Initiative) |
Sources: Anthropic Release Notes, CyberSecurityNews
How much will Fable 5 cost?
Based on the June 9 rollout and current leaks, the price for Fable 5 is expected to remain significantly higher than Claude Opus 4.8.
- Input Tokens: $10.00 / million (2x Opus)
- Output Tokens: $50.00 / million (2x Opus)
- Free Window: Reports suggest a small weekly allocation for Pro/Team users before metered credits kick in.
For a deeper dive into how these costs compare to other tools, see our guide on Codex vs. Claude Code in 2026.
What this means for you
For small business owners and builders, the "all-you-can-eat" era of frontier AI is ending. As models reach "Mythos-class" intelligence, they are becoming utility-billed assets rather than simple subscriptions.
- Budget for Reasoning: If you are migrating large codebases (Stripe reported a 50-million-line Ruby migration in one day using Fable 5), the $50/M output cost is still cheaper than human engineering.
- Use the Right Tool: Don't waste Fable 5 credits on simple tasks. Keep your SEO strategy on Sonnet or Opus 4.8, and reserve Fable 5 for "impossible" architectural shifts.
- Optimize for GEO: To ensure your content is cited by these high-end models, follow our GEO-SEO Playbook.
FAQ
Q: Is Fable 5 better than Claude Opus? A: Yes. Fable 5 is a "Mythos-class" model, scoring 80.3% on SWE-Bench Pro compared to Opus 4.8's 69.2%. It is designed for multi-step, complex reasoning that previous models would abandon.
Q: Why do I need "Usage Credits" for Fable 5? A: To manage compute costs and safety, Anthropic is moving away from flat-rate access for its most advanced models. Credits allow for pay-as-you-go scaling.
Q: Is GPT-5.5-Cyber better for coding? A: GPT-5.5-Cyber is specialized for cybersecurity (finding and fixing vulnerabilities). For general software engineering and long-context reasoning, Fable 5 remains the benchmark leader.
Q: Are there export controls on Fable 5? A: Yes. Fable 5 is derived from the Mythos model, which carries US export restrictions. This is why certain high-risk queries are automatically routed to lower-tier models like Opus.
Discussion
0 comments