The AI Brief

On Tuesday Anthropic launched Claude Fable 5, the first publicly available version of its Mythos model, which the company says excels at software engineering, knowledge work, and vision but ships with hard safety limits. Mythos-class models are a tier that sits above the Opus class in capability; the first, Claude Mythos Preview, was released in April through Project Glasswing, followed by Fable 5 and Mythos 5 on June 9.

Fable 5 is described as a Mythos-class model made safe for general use, with capabilities exceeding any model Anthropic has made generally available. It is state-of-the-art on nearly all tested benchmarks, showing exceptional performance in software engineering, knowledge work, vision, and scientific research. The longer and more complex the task, the larger Fable 5's lead over other models. In high-risk areas including cybersecurity, biology, chemistry, and distillation, the model blocks responses and falls back to Claude Opus 4.8. Anthropic says the safeguards trigger on average in fewer than 5% of sessions. Pricing for both Fable 5 and Mythos 5 is $10 per million input tokens and $50 per million output tokens, double the price of Opus 4.8.

Access on subscription plans rolls out in stages: through June 22, Fable 5 is included in Pro, Max, Team, and seat-based Enterprise plans at no extra cost, after which Anthropic will pull it from those plans and require usage credits going forward. Anthropic is simultaneously deploying an updated Claude Mythos 5 to organizations already approved under Project Glasswing. Wary of misuse, Anthropic says it ran an external bug bounty with over 1,000 hours of testing that produced no universal jailbreaks, then engaged external red-teaming organisations that also failed, though novel attacks remain possible. Third-party assessments from Stripe, Hex, and Genspark reported significant gains on complex coding and analytics tasks, per Anthropic's announcement.

Disclosure: Anthropic, mentioned in this item, is the company that develops Claude, which generates this brief.

SpaceX plans to market its IPO at a fixed price of $135 per share, implying a valuation of $1.77 trillion, which would place it above Tesla's market cap of approximately $1.6 trillion. The company plans to sell 555.6 million shares, amounting to a $75 billion fundraise. SpaceX is planning to debut on Nasdaq on June 12. Musk will own over 82% voting control after the offering, per the filing.

SpaceX took the rarer path of a fixed offer price; the practical effect is that demand expresses itself entirely through the secondary market once SPCX opens. With a retail slice potentially as high as 30%, at least triple the 5–10% typical of standard deals, the opening session is where price discovery actually happens. SpaceX, Anthropic, and OpenAI could all go public within months of each other, a concentration of high-stakes offerings the markets have not seen since the dot-com boom. Anthropic recently disclosed a revenue run rate of $47 billion, up from roughly $10 billion in annual revenue last year, and closed a funding round at a $965 billion valuation; OpenAI was valued at $852 billion in late March.

CaveatValuation figures are based on roadshow materials and broker disclosures as of June 2026. Final pricing occurs after market close on June 11; figures may change. Vendor-interest framing from financial intermediaries should be weighed accordingly.

Anthropic has made a change to the way it handles business customer data for Fable 5, Mythos 5, and future models with similar or higher capability. It will require 30-day retention for all traffic on Mythos-class models, on both first- and third-party surfaces. Anthropic says it will not use this data to train new Claude models or for any non-safety-related purpose, and has instituted new privacy protections including logging all human access and ensuring deletion after 30 days in almost all cases. The data is intended to defend against complex and novel attacks, including new jailbreaks and attacks that operate across many requests.

Every other Claude model available through the API, including Opus 4.8, Sonnet 4.6, and Haiku 4.5, can operate under ZDR agreements. Fable 5 cannot. The change applies specifically to organisations that have set up workspaces with ZDR in Claude Console, use Claude Code with ZDR in Claude Enterprise, or access Claude through AWS Bedrock, Google Cloud Agent Platform, or Microsoft Foundry with ZDR. For all other organisations, there is no change. When Amazon announced that Claude Fable 5 was available on AWS, a detail tucked into the infrastructure section noted that once data retention is opted in, data will leave AWS's data and security boundary.

TechCrunch notes the policy could set an industry precedent in which access to increasingly powerful models comes bundled with mandatory data-retention policies framed as a safety measure. Enterprises in sectors with matter-confidentiality obligations should classify which workloads can tolerate 30-day third-party storage before enabling Fable 5; on GitHub Copilot, admins must explicitly acknowledge the requirement as an opt-in policy before the model activates.

Disclosure: Anthropic, mentioned in this item, is the company that develops Claude, which generates this brief.

OpenAI is backing away from fully autonomous AI research by 2028, now talking about a "tandem" between humans and machines. CEO Sam Altman and chief researcher Jakub Pachocki also call for an international body that could slow frontier development if needed. Last fall, OpenAI had laid out an ambitious goal: by March 2028, the company wanted to build a fully autonomous AI system capable of conducting research on its own.

The "Built to Benefit Everyone" post published June 8 states that "entirely automating everything is not the future we want," characterizing it as both unfulfilling and dangerous, and arguing that AI should help people pursue their goals, not become untethered from them, with the human role in setting direction, making tradeoffs, and applying judgment becoming more important as systems grow more capable. The filing also states that "by March 2028 we may have a significant fraction of our research being done by AI systems in tandem with our own researchers," replacing the earlier fully-automated framing.

The softened language lands at a notable moment. OpenAI, which was last valued at $852 billion, submitted a draft registration statement to the SEC for a proposed IPO. The firm projects burning $85 billion in 2028 even after doubling sales from the prior year, and by its own projections will not generate more cash than it spends for at least four more years. Analysts noted the "tandem" framing serves dual purposes: it reduces liability exposure around autonomous-system risks while offering a margin-expansion narrative for the S-1, per FourWeekMBA's read of the document.

Microsoft cut off access to dozens of its open-source projects hosted on GitHub as it investigates how hackers apparently breached the projects and injected password-stealing malware. Many of the affected projects relate to Azure and tools used by developers to code with AI development apps, including Claude Code, Gemini's command line interface, and VS Code. Security firm Cloudsmith and malware analysis site OpenSourceMalware were among the first to flag the hack; the malware allowed attackers to steal users' passwords and other sensitive credentials when they opened the compromised tools in their AI coding apps.

Microsoft confirmed it is investigating and has already contacted a small number of potentially affected customers, though the full scale remains unclear. This is Microsoft's second known incident in recent weeks; Ars Technica reported an earlier breach in May involving the Durable Task project, a tool for building and managing applications. OpenSourceMalware believes the latest case may link to the earlier one, describing it as a possible re-compromise suggesting attackers retained access after the first breach.

The incident is an instance of the software supply chain attack pattern that has been documented in the AI tooling ecosystem since early 2026. Per the Cisco State of AI Security 2026 report, 83% of organisations plan to deploy agentic AI, but only 29% feel ready to do so securely, and only 34.7% of organisations have deployed dedicated prompt injection defences. A compromised dependency in an AI coding tool is a particularly high-value vector because the credentials obtained grant access to the same cloud and model environments the developers are using to build production systems.