The Strange 22-Day Journey of Claude Fable 5

In the span of two days, Anthropic shipped its cheapest serious agent model and got its most powerful model un-banned by the US government. Both stories matter if you build for the web, and both have small print worth reading before you change anything in your stack.

Fable 5: released, banned, and back in 22 days

The short version of a genuinely strange month: Anthropic released Claude Fable 5 on June 9, its most capable publicly available model ever, built for long-running autonomous work with a million-token context window [1]. Three days later, a US government export directive forced it offline. The trigger was a jailbreak found by Amazon researchers that got the model to identify software vulnerabilities and, in one case, produce code demonstrating how one could be exploited [2]. The directive barred access for foreign nationals, Anthropic had no way to verify anyone's nationality in real time, so it pulled the model for everyone [3].

On June 30 the Commerce Department lifted the controls, and Fable 5 came back July 1 across Claude.ai, the API, Claude Code, and Cowork [3]. Anthropic's defense held up under review: its testing showed less capable models, including its own Opus 4.8 and competitors' models, could find the same vulnerabilities in the report, meaning Fable offered no unique offensive capability [2]. The company is pairing the return with a public jailbreak bounty program and deeper pre-release testing with the government [3].

The practical details if you're a Claude subscriber: through July 7, paid plans can put up to half their weekly usage limit toward Fable 5, after which it runs on usage credits [4]. On the API it's priced at $10 per million input tokens and $50 per million output, double Opus 4.8, and it carries a mandatory 30-day data retention designation with no zero-retention option [5]. Premium capability, premium terms.

The lesson has nothing to do with Anthropic specifically: a frontier model your workflow depends on can now vanish overnight by government directive. If your business runs on one model with no fallback, that's a single point of failure you're choosing.

Sonnet 5: near-flagship agents at mid-tier prices, with an asterisk

The quieter release is the one most working sites will actually feel. Claude Sonnet 5 launched June 30 as Anthropic's most agentic mid-tier model: it plans, drives browsers and terminals, and runs autonomously at a level that recently required flagship-priced models [6]. On agentic coding benchmarks it scores 63.2 percent against Opus 4.8's 69.2 and its predecessor's 58.1, and it edges past Opus on at least one knowledge-work benchmark [7]. It's the default model on Free and Pro plans and already live in GitHub Copilot [8].

Now the small print. The launch price is $2 per million input tokens and $10 per million output, but only through August 31; on September 1 it steps up 50 percent to $3 and $15 [6] [7]. And Sonnet 5 uses a new tokenizer that can turn the same text into up to 1.35 times more billable tokens than before. Anthropic set the intro price so the migration is roughly cost-neutral today, which is honest, but it means the September step-up lands on an already-inflated token count [9]. Third-party cost analysis puts the same unchanged workload 20 to 35 percent above its old baseline once both effects stack, while the rate card still reads "unchanged" [10].

None of that makes Sonnet 5 a bad deal. It's a real price-performance step, and cheaper than the flagship models it now competes with. It makes it a deal you should measure rather than assume: the number that matters for agent workloads is cost per completed task, not the per-token sticker. This is the same reason every tool on our platform shows its credit cost before you run it. Whoever your AI vendor is, the meter matters more than the menu.