Tech • IA • Crypto
Fable 5 and Sonnet 5: what is true and what is not
9.2/10TL;DR
The return of Fable 5 and the launch of Son 5 are reshaping Anthropic’s offering, combining tighter security, evolving costs, and competing interpretations of performance.
KEY POINTS
Return of Fable 5 under stricter constraints
Fable 5 is available again as of July 1 across several cloud environments, but with a deeply revised security architecture. A new classifier, co-developed with the U.S. government, blocks the mi-join bypass technique that led to its withdrawal. Effectiveness exceeds 99%, but at the cost of more aggressive filtering.
Filtering that impacts legitimate use cases
This tightening has side effects. Standard requests, especially in code or debugging, may be blocked. Instead of outright refusal, requests are redirected to Opus 4.8, maintaining continuity while limiting risk.
Limited access and a transitional business model
Pro, Max, and Team subscribers can use Fable 5 within 50% of the weekly quota until July 7. After that, access shifts to a paid credit system, signaling a move toward tighter control and gradual monetization.
Mythos 5 still under restricted access
The Mythos 5 model remains reserved for selected partners in the Glass Wing program, approved on June 26 by the U.S. Department of Commerce. Public rollout has not yet begun.
A vulnerability not specific to Fable 5
Internal analyses show that models like Opus 4, GPT 5.5, and Kimi 4AT 2.7 could identify the same vulnerabilities as Fable 5. Anthropic is also launching, with Amazon, Microsoft, and Google, a joint standardization effort against jailbreaks.
Son 5 becomes the new standard
Launched on June 30, Son 5 replaces previous versions as the default model for free and pro tiers. Pricing is set at $3 input and $15 output per million tokens, with a promotion at $2 / $10 until August 31, 2026. By comparison, Opus 4.8 costs $5 / $25.
A tokenizer that changes the economics
Son 5 introduces a new tokenizer, which can generate up to 1.35× more tokens for the same text. This mechanically increases cost per task without changing unit pricing, explaining part of the criticism around its cost.
Improved but nuanced performance
On Bench Pro, Son 5 reaches 63.2%, compared to 58.1% for Son 4.6 and 69.2% for Opus 4.8. On Terminal Bench 2.1, it climbs to 80.4%. It approaches Opus in knowledge and agentic tasks, but still lags in pure coding.
Controversies over costs and biased readings
Some calculations claim costs up to 57× higher than certain competitors. Data from Artificial Analysis places Son 5 at 2.29 per task, confirming gaps, but these mainly stem from token consumption. Comparisons often rely on standard pricing without factoring in promotions.
A controversial chart corrected
A viral chart placed Son 5 consistently below Opus 4 on BrowseComp. Anthropic has since corrected it, acknowledging an initially simplified methodology that underestimated its performance. Widely shared screenshots were based on that erroneous version.
CONCLUSION
Between strengthened security, pricing repositioning, and debates over real-world performance, the arrival of Fable 5 and Son 5 marks a strategic transition phase for advanced AI models.