Anthropic Brings Its Most Powerful Model to the Public

SAN FRANCISCO—The pricing page went live Tuesday with a number that told the story: $10 per million input tokens and $50 per million output tokens for Claude Fable 5, less than half the price of the restricted model that preceded it. The model itself is the most capable Anthropic has ever trained, and for the first time, anyone can use it.

Anthropic released Fable 5 to the general public on June 9, along with a second version, Claude Mythos 5, that stays behind closed doors. The two share the same underlying model; the difference is the safety layer wrapped around it. Fable 5 comes with safeguards that route high-risk queries, mainly in cybersecurity and biology, to Anthropic’s previous flagship, Claude Opus 4.8. Mythos 5 has those safeguards lifted and is available only to a small group of partners under Project Glasswing, a program run with the U.S. government.

The split is a bet that Anthropic can have it both ways: sell the strongest model it has built to the widest possible audience while keeping its most dangerous capabilities out of reach. Company officials said the safeguards trigger in fewer than 5% of sessions on average, a rate they describe as a reasonable price for releasing the model quickly.

The capabilities on offer are substantial. Anthropic said Fable 5 is state-of-the-art on nearly all benchmarks it measures, showing exceptional performance in software engineering, knowledge work, vision and scientific research. In internal tests, the model wrote code, planned complex projects and handled long-horizon tasks that earlier models could not complete without human help.

The release is notable for what it does not include. Anthropic said it tuned the safeguards conservatively, meaning some harmless requests will be bounced to the older model. For researchers and power users who hit the guardrails, the experience is a visible fallback to Opus 4.8 rather than a refusal, a design choice the company says makes the system more transparent.

The pricing is aggressive. Mythos Preview, the earlier restricted version, cost roughly five times Opus 4.8. Fable 5 at about twice Opus 4.8’s price puts frontier capability within reach of startups and developers who previously could not afford it. Analysts said the move is aimed as much at enterprise adoption as at consumers, since the model’s reliability in coding tasks makes it a candidate to replace armies of junior engineers.

The timing is not accidental. Anthropic is widely reported to be preparing for an initial public offering, and a widely used, highly regarded model is the strongest possible advertisement for an AI company’s value. A public release also puts distance between Anthropic and the image of a lab that hoards its best work for governments.

Competition frames the decision too. OpenAI, Anthropic’s chief rival, has moved toward releasing frontier models with fewer restrictions, arguing that openness accelerates safety research. Anthropic’s approach is different: keep the sharpest edges private, sell the rounded version broadly. The company’s safety team, which has been its most distinctive institutional feature since its founding in 2021 by former OpenAI researchers, retains a veto over what ships.

Investors have rewarded the strategy. Anthropic’s valuation has climbed through repeated funding rounds, most recently to a level that puts it among the most valuable private companies in the world. Amazon, its largest backer, has invested billions and runs the company’s models on its own chips, a relationship that deepens with every public release.

The government connection adds a layer. Project Glasswing, which deploys Mythos 5 for defensive cybersecurity work, gives Anthropic a channel into federal agencies that competitors lack. Company officials describe the arrangement as a way to learn how the most capable models behave in high-stakes settings, knowledge that feeds back into the safeguards on the public version.

The risks are real. A frontier model in the hands of millions of users is a target for jailbreaks, and Anthropic acknowledged that its safety classifiers will not catch everything. The 319-page system card published alongside the release documents the known failure modes, an unusual degree of disclosure for the industry.

For now, the market verdict is in the numbers. Within a day of release, developers were running Fable 5 on workloads from code review to legal analysis, and early reports of quality were strong. The company says the model will roll out across its app, API and cloud partners in the coming weeks.

The rollout itself will be watched closely. Anthropic says Fable 5 will reach its app, its API and its cloud partners over the coming weeks, and enterprise customers have been queued for early access. The company has expanded its support teams ahead of what executives expect to be the busiest release week in its history, and it has briefed regulators in Washington and Brussels on the safeguards. Early demand, by all accounts, has been strong.

The broader question is whether the two-track approach holds. If Fable 5 proves too capable in the wrong hands, regulators may demand tighter controls; if the safeguards prove too restrictive, users may defect to OpenAI. Anthropic is betting it can thread that needle, one model release at a time. Tuesday’s launch was the first real test.

Related Posts

  • September 6, 2026
  • 7 views
Anthropic Moves Its IPO Filing to Late September

The bankers and lawyers running Anthropic’s initial public offering had told investors to expect the company’s registration documents as soon as this week. The calendar has moved. Anthropic now plans…

  • September 6, 2026
  • 6 views
OpenAI Quietly Revises GPT-6 Astra Scores After Launch

When OpenAI released GPT-6 Astra on Sept. 3, the launch post carried the usual furniture of a modern model debut: coding results, speed comparisons and a figure for how often…