Anthropic said late Tuesday that the US Department of Commerce had removed the export restrictions it imposed on June 12, and that general access to Claude Fable 5 would return on Wednesday, July 1, with Mythos 5 access already restored on June 26 for organizations that "operate and defend critical infrastructure." The shutdown began after Amazon researchers reportedly found a way to bypass Fable 5's guardrails and get it to identify software vulnerabilities and generate exploit code — prompting officials to pull both Mythos-class models offline worldwide, including for Anthropic's own staff.
In a statement, Anthropic thanked users for their patience and said it had built an updated automated classifier to close the gap the government flagged, clearing the way for a full commercial rollout across the Claude Platform, AWS, Google Cloud and Microsoft Foundry. US Commerce Secretary Howard Lutnick told CNBC he had determined that "appropriate safeguards" were now in place to permit "trusted partners" to access the models.
The episode is also reshaping how the industry polices itself: Anthropic, Amazon, Microsoft and Google have started work on a shared framework to score the severity of AI security breakdowns across four criteria — capability gain, breadth of that gain, ease of weaponization, and public discoverability — an attempt to give regulators a consistent yardstick instead of ad hoc shutdown orders. "We're grateful to our users for their patience," Anthropic wrote, while continuing to press the case that government access reviews "shouldn't become the long-term default."
Claude Sonnet 5 launched Tuesday as the default model for Free and Pro users, and is available immediately across Max, Team, Enterprise and the API. Anthropic says it substantially closes the gap with Opus 4.8 on reasoning, tool use, coding and knowledge work while staying cheaper — testers cited in the launch post described it finishing multi-step jobs, like updating Salesforce tiers and sending a launch email end to end, that previous Sonnets would have stalled on halfway through.
The launch arrives four days after OpenAI's GPT-5.6 Sol preview and lands in the middle of Anthropic's own IPO push: the company confidentially filed a draft S-1 on June 1 following a $65 billion Series H that valued it at $965 billion. Because Sonnet 5 shows a slight uptick in partial cyber-exploit success versus 4.6, Anthropic shipped it with real-time cyber safeguards on by default — though the model still scored zero percent on a full Firefox exploit-development evaluation, well behind Fable 5 and Mythos 5.
Claude Science entered public beta Tuesday for macOS and Linux on Pro, Max, Team and Enterprise plans. It isn't a new model — it runs on whatever Claude the user's plan already includes — but a workbench that folds together the tools researchers juggle daily: PubMed, Jupyter, R, a cluster terminal, and dozens of specialty databases like UniProt, PDB and ChEMBL. Sessions run on the lab's own infrastructure, so large or sensitive datasets never have to leave it, and a reviewer agent inspects outputs as they're produced, flagging incorrect citations, untraceable numbers and figures that don't match their underlying code.
The app is pre-configured for genomics, single-cell analysis, proteomics, structural biology and cheminformatics, and leans on NVIDIA's BioNeMo Agent Toolkit to reach models like Evo 2, Boltz-2 and OpenFold3. Early users include Manifold Bio, which used it to nominate drug targets by weighing surface expression and trafficking data, and an Allen Institute neuroscientist who built a 20-skill multi-agent review pipeline. Anthropic is also funding up to 50 "AI for Science" research projects with up to $30,000 in credits each; applications close July 15.
Sen. Mark Warner (D-VA) released a discussion draft of the Artificial Intelligence Access, Gatekeeper Exchange, and Nondiscriminatory Transfer Act on Monday, aimed at stopping dominant platforms from locking out competing AI agents. The bill would give users of platforms with more than 50 million monthly users the right to choose at least one FTC-certified agent provider, require agents to disclose their human operator's identity, and let the commission deregister providers that violate baseline privacy and security rules.
The push follows Morgan Stanley estimates that nearly a quarter of Americans made an AI-assisted purchase in a recent 30-day span, alongside growing reports of agents making erratic purchases or leaking data without a user's knowledge. Warner is releasing the draft for public feedback before formally introducing it, and the timing is notable: it lands the same month the Commerce Department's export-control fight with Anthropic pushed frontier-model governance to the top of the congressional agenda.
California became the first state to strike a blanket agreement with Anthropic, offering all state agencies — and extending down to cities and counties — access to Claude at a 50% discount, alongside free training for public-sector workers. Anthropic's Head of Americas, Kate Jensen, framed it as an effort to "put Claude to work for the people who keep this state running," covering tasks from permitting backlogs to constituent casework.
The deal lands the same week Anthropic launched Sonnet 5 and Claude Science, and fits a broader pattern of frontier labs courting government customers even as they navigate export-control friction with the same government. It's also a template other states are likely to study: budget-constrained public agencies get frontier AI at a discount, while Anthropic locks in the kind of durable, recurring institutional revenue that could matter to investors ahead of its planned IPO.
Meta FAIR published Brain2Qwerty v2 on Monday, a follow-up to last year's v1 that decodes full sentences in real time from magnetoencephalography (MEG) recordings, without any surgical implant. The system trained on roughly 22,000 sentences from nine volunteers who each spent about 10 hours wearing an MEG scanner while typing, and combines a convolutional encoder, a transformer and a fine-tuned language model to turn raw magnetic brain signals into coherent text — averaging 61% word accuracy, with the best participant reaching 78%.
That's a sharp jump from the roughly 8% accuracy of earlier non-invasive approaches, and Meta says the gap with surgical brain-computer interfaces should keep narrowing as more training data is collected, since accuracy scaled log-linearly with data volume in testing. Meta is open-sourcing the full training code for both v1 and v2, along with partner BCBL's underlying dataset, aiming to help the millions of people with brain lesions or conditions like ALS who have lost the ability to communicate.