Anthropic's 24 Hours: Fable 5 Returns, Sonnet 5 Launches, and a Science Lab Opens

Anthropic's 24 Hours: Fable 5 Returns, Sonnet 5 Launches, and a Science Lab Opens

Fable 5 went offline on a Friday afternoon in June, and it stayed offline for 18 days. On July 1, it came back, alongside a new model, Claude Sonnet 5, and a new product built for scientific research. The three announcements landed within a day of each other, and together they offer a rare, concrete look at how a frontier AI lab navigates government oversight, competitive pricing pressure, and expansion into new markets, all at the same time.

Why Fable 5 disappeared

Anthropic released Fable 5 and its more capable, less restricted sibling Mythos 5 on June 9. Both models share the same underlying architecture, but Fable 5 shipped with substantially stronger safeguards, intended to make it safe for general public use, while Mythos 5 went only to a small set of trusted partners working on defensive cybersecurity under Anthropic's Project Glasswing.

Three days later, the Commerce Department issued an export control directive requiring Anthropic to restrict both models from foreign nationals, a rule the company had no reliable way to enforce in real time. Anthropic pulled access for everyone rather than risk violating the order.

The trigger was a report from Amazon researchers who had found a way to prompt Fable 5 into identifying software vulnerabilities and, in one case, producing code demonstrating how to exploit one of them. When Anthropic tested other models against the same prompts, it found that Opus 4.8, GPT-5.5, and Kimi K2.7 could identify the same vulnerabilities, and that every model it tested, including older and smaller Claude models, could reproduce the single exploit demonstration. Anthropic's own account describes the incident as a borderline case involving routine defensive cybersecurity information rather than a unique Mythos-level capability, but the government's directive stood regardless, and Anthropic spent the next two and a half weeks working with Commerce Department officials, the Center for AI Standards and Innovation, and national security agencies to resolve it.

The company's response was a new safety classifier that specifically targets the reported bypass, which Anthropic says now blocks the technique in over 99% of cases, along with a wider safety margin that catches more borderline requests at the cost of occasionally flagging legitimate coding and debugging work. CAISI researchers tested both the old and new safeguards and told Anthropic they consider the updated version extraordinarily strong, according to the company's account.

A shared standard for jailbreak severity

One outcome of the episode is a proposal Anthropic is developing with Amazon, Microsoft, Google, and other Glasswing partners: a common framework for scoring how serious a given AI jailbreak actually is. The framework scores a jailbreak on four dimensions, how much capability it hands an attacker beyond what's already available, how many different offensive tasks it works for, how much effort it takes to turn into an actual attack, and how easily the technique can be found or shared. Anthropic argues that without a shared standard, both AI developers and government regulators are left guessing at how urgently to respond to any given report, a gap that becomes more consequential as more labs release models with serious cybersecurity capabilities.

Anthropic is also committing to expanded government collaboration going forward, including early access for designated government evaluators to test new frontier models and their safeguards before public release, faster sharing of information when new jailbreaks surface, and a standing team supporting joint research with government partners. Fable 5 is now rolling back out globally through the Claude Platform, Claude.ai, Claude Code, and Claude Cowork, with cloud access on AWS, Google Cloud, and Microsoft Foundry to follow. Mythos 5 access remains limited to a set of US organizations that operate critical infrastructure, with Anthropic saying it continues to negotiate broader access with the government.

Sonnet 5 closes the gap with Opus

The same day Anthropic resolved the Fable 5 suspension, it released Claude Sonnet 5, which the company describes as its most agentic Sonnet-class model so far. Anthropic's benchmark data shows Sonnet 5 performing close to Opus 4.8 on reasoning, tool use, coding, and knowledge work, at roughly 40% of the price. On evaluations like BrowseComp for agentic search and OSWorld-Verified for computer use, Sonnet 5 covers a wider range of cost-performance tradeoffs than its predecessor, Sonnet 4.6, and at higher effort settings its results approach Opus 4.8 on some tasks.

Anthropic's safety testing found Sonnet 5 shows lower rates of hallucination, sycophancy, and undesirable behavior overall compared to Sonnet 4.6, though it scored somewhat higher on the company's automated behavioral audit than the more capable Opus 4.8 and Claude Mythos Preview. On cybersecurity specifically, the company says Sonnet 5 was not deliberately trained on offensive cyber tasks and performs substantially worse than Opus 4.8 or Mythos 5 at developing software exploits. In one test built with Mozilla around patched Firefox vulnerabilities, Sonnet 5 never produced a fully working exploit, though it showed a slightly higher rate of partial success than Sonnet 4.6, a change Anthropic attributes to general capability gains rather than any cyber-specific training.

Sonnet 5 is now the default model for Free and Pro users and is available across Max, Team, and Enterprise plans, priced at $2 per million input tokens and $10 per million output tokens through August, after which it moves to $3 and $15. It is available in Claude Code and through the API as claude-sonnet-5.

Claude Science tries to consolidate the research stack

The third announcement, Claude Science, addresses a different problem entirely: the fragmented tooling that slows down scientific research. Researchers routinely move between PubMed, Jupyter, R, and cluster terminals, each with its own schema and workflow. Claude Science brings these into one environment where a coordinating agent, backed by more than 60 skills and connectors for genomics, single-cell biology, proteomics, and cheminformatics, can query specialized databases like UniProt, PDB, and ChEMBL directly from a plain-language request.

Every figure or result the system produces comes with the code and environment that generated it, along with a plain-language explanation, so the work can be checked or reproduced later. A separate reviewer agent checks citations and calculations as analyses run, flagging figures that don't match their underlying code. The platform also manages compute directly, submitting jobs to a lab's own HPC cluster or to on-demand GPUs through Modal, and keeping data on the systems where it already lives.

Early users described concrete gains. A neuroscientist at the Allen Institute used the platform to build a pipeline of about 20 custom skills for writing long-form literature reviews, cutting a process that used to take up to two years down to a matter of months. An epidemiologist at UCSF's Brain Tumor Center said the platform enabled germline genetic analysis in roughly a tenth of the time it previously required, results his lab independently validated before relying on them. Manifold Bio, a biotech company developing tissue-targeting medicines, used Claude Science to rank candidate drug targets against criteria drawn from the company's own proprietary data, something the company said distinguished it from a general-purpose coding assistant.

What the timing suggests

Taken separately, each of these three announcements is a distinct piece of news: a regulatory dispute resolved, a new model at a new price point, a product aimed at a new market. Taken together, they show a company moving on several fronts simultaneously under active government scrutiny, something that would have been unusual for an AI lab even a year earlier. The Fable 5 episode in particular illustrates how contested the boundary has become between capabilities that are genuinely novel and capabilities that were already achievable with existing tools, since Anthropic's own testing found that most of the models implicated in the Commerce Department's directive, including competitors' models, could reproduce the same behavior that triggered the suspension.

The performance figures from Claude Science and the benchmark comparisons for Sonnet 5 are self-reported by Anthropic and the labs it names, without independent replication at this stage. What is independently verifiable is the sequence of events itself: a model taken offline by government order, a joint industry framework proposed in its wake, and two new products released the moment the dispute cleared. Whether the jailbreak severity framework holds up as new incidents surface, and whether labs like Manifold Bio and UCSF report the same efficiency gains months from now as they did during the beta period, will say more about where this settles than the announcements themselves.