Cutting-edge AI capability will become increasingly expensive and restricted
Anthropic has opened its most capable model yet to the public, threading a careful line between ambition and restraint — releasing Claude Fable 5 with built-in guardrails that redirect sensitive queries while reserving the full, unrestricted version for a select circle of institutional partners. The move arrives as the company approaches a public offering, and days after calling for a global pause on AI development, placing its stated caution in tension with its commercial momentum. What emerges is a portrait of a technology crossing a threshold: powerful enough to compress months of work into hours, yet costly and restricted enough to widen the gap between those who can afford the frontier and those who cannot.
- Anthropic released a model it had held back for three months as too dangerous, then launched it publicly within the same week it called for a global AI slowdown — the contradiction is drawing scrutiny.
- Built-in safety filters can intercept queries about cybersecurity, biology, and chemistry, rerouting them to an older, weaker model — though only 5% of sessions ever hit this wall.
- A restricted, more permissive version is quietly circulating among roughly 200 organizations across 15 countries, including U.S. government collaborators, under a program called Project Glasswing.
- Real-world results are striking: Stripe completed a two-month code migration in a single day, and developers are generating full games in hours — benchmarks confirm a measurable edge over prior models.
- Pricing has doubled, access is time-gated for subscribers, and after June 22 the model shifts to a credit-consumption model — the frontier is becoming both more capable and more expensive to reach.
Anthropic has released Claude Fable 5, a public version of its Mythos model — a system the company had developed three months earlier but held back, judging it too powerful for broad deployment. The timing is notable: the release came just days after Anthropic publicly called for a global pause on AI development, a contradiction that observers have not let pass quietly.
The model ships with safety mechanisms that can block or redirect sensitive questions — particularly those touching on cybersecurity, biology, and chemistry — routing them instead to Claude Opus 4.8, an older and deliberately limited system. Anthropic says fewer than 5 percent of sessions trigger this redirection. A less restricted version, Claude Mythos 5, is being made available to a curated group of organizations through Project Glasswing, a program that began with roughly 40 critical infrastructure partners and has since expanded to around 200 organizations in more than 15 countries, including collaborators with the U.S. government.
On performance, Fable 5 sits at the top of available benchmarks, with a measured 5 percent edge over its predecessor in coding and mathematics. The practical implications are vivid: Stripe used the model to complete a code migration that would have taken two months — finishing it in a single day. Developers have shared examples of complete games built in just a few hours.
The cost reflects that power. Input tokens run at $10 per million and output at $50 per million — roughly double the price of Opus 4.8. For casual users, the difference may feel modest. For the corporate clients and developers who drive AI adoption at scale, it is significant. Access is also being restructured: Fable 5 is currently available only to subscribers, but after June 22 it shifts to a credit-based model that charges beyond the base subscription. The message is clear — cutting-edge AI capability is becoming more expensive, more tiered, and more deliberately rationed.
Anthropic has released Claude Fable 5, a public version of its Mythos model that the company developed three months ago but deemed too powerful and dangerous to share widely. The release comes as the company pushes toward an initial public offering, creating what some observers see as a contradiction between Anthropic's recent calls for a global pause on AI development and its decision to open this advanced system to the public.
Fable 5 arrives with built-in safety mechanisms designed to manage risk. Questions touching on cybersecurity, biology, and chemistry can be blocked or redirected to Claude Opus 4.8, an older and deliberately less capable model that Anthropic presents as safer. The company reports that fewer than 5 percent of user sessions trigger this redirection—meaning 95 percent of conversations run entirely on Fable 5 without hitting these guardrails. The system is also being released in a more permissive version called Claude Mythos 5 to a limited set of organizations through a program called Project Glasswing, which includes cybersecurity partners and collaborators with the U.S. government. According to reporting, the initial Mythos version was shared with roughly 40 organizations managing critical infrastructure, and Anthropic plans to expand that circle. The Guardian reports the program now includes around 200 organizations across more than 15 countries.
Performance benchmarks place Fable 5 at the top tier of available models, particularly for programming and mathematics. Testing by Vals AI showed a 5 percent advantage over Opus 4.8. Real-world use cases underscore the capability gap: Stripe, the payments company, completed a major code migration that would have taken two months in a single day using Fable 5. Researchers and developers posting on social media have demonstrated the model generating complete games in just a few hours of work.
The pricing reflects the model's power and scarcity. Fable 5 costs $10 per million input tokens and $50 per million output tokens—roughly double the cost of Opus 4.8. Tokens are units of text used to measure and bill AI usage. For most casual users treating AI as a chatbot for advice or practical help, the performance jump will feel incremental. The real competition is happening elsewhere: among corporate customers and software developers who pay for frontier AI technology. These are the users who will feel the difference and bear the cost.
The release also signals a shift in how Anthropic will monetize its most advanced systems. For the first 15 days, Fable 5 is available only to subscription holders. After that window closes on June 22, access moves to a credit-based consumption model that charges beyond the base subscription price. This tiering suggests that cutting-edge AI capability will become increasingly expensive and restricted in the months ahead.
The contradiction between Anthropic's messaging and its actions has not gone unnoticed. The company called for a global pause on AI development days before releasing a model it had previously described as too risky for broad deployment. The New York Times reported that Anthropic is releasing a version of Mythos with what researchers believe are sufficient safety constraints for wider use, though the system relies partly on older technology to manage certain tasks. The Rundown newsletter noted that while every AI lab claims its latest release is the world's best model, it is rare for the broader AI community to agree—yet benchmarks suggest Fable 5 has delivered on its promises. The real questions now center on cost and access, with significant content restrictions and a hard deadline before pricing becomes steeper.
Notable Quotes
Anthropic is releasing a version of Mythos with safety constraints sufficient for wider use, though the system relies partly on older technology to manage certain tasks— New York Times reporting
The real questions now center on cost and access, with significant content restrictions and a hard deadline before pricing becomes steeper— The Rundown newsletter
The Hearth Conversation Another angle on the story
Why release a model the company said was too dangerous just three months ago?
The IPO timeline matters here. Anthropic needs to show growth and market leadership to investors. Releasing Fable 5 demonstrates technological dominance while the safety guardrails let them claim they're managing risk responsibly.
Does the 5 percent redirection rate actually prove the guardrails work, or does it just mean most queries don't trigger them?
That's the real question. It could mean the safety measures are well-designed and proportionate. Or it could mean they're narrow enough that most users never hit them. The fact that 95 percent of conversations stay on Fable 5 suggests the guardrails are selective, not comprehensive.
Why does it cost twice as much as the previous model?
Partly because it performs better—that's real value. But also because Anthropic is testing how much users will pay for frontier capability. The credit-based model after 15 days is the real signal: they're moving toward making advanced AI a luxury good.
What does the Glasswing program tell us about who actually gets access to the unrestricted version?
It's a two-tier system. Governments and critical infrastructure partners get Mythos with fewer restrictions. Everyone else gets Fable 5 with guardrails. It's a way to maintain relationships with powerful institutions while controlling what the public can do.
Is this actually a contradiction, or is it just business?
Both. Anthropic genuinely believes in AI safety—that's not cynical. But they also need to survive as a company and compete with OpenAI and others. The contradiction is real, but it's the contradiction of the entire industry right now.