Editor's note: today's lead is the biggest model release of the year so far, so Models & Benchmarks leads, Enterprise and Open Source take the other two mains, and Regulation sits in the quick hits — where the freshest regulatory story is in any case a direct development of Monday's lead (the Apple–Brussels standoff over Siri AI). The MiniMax M3 weights watch hits its due date tomorrow.
The Frontier Goes Public With Strings Attached: Claude Fable 5 Arrives at Double the Price, With Safeguards and Mandatory Retention
Anthropic on Tuesday released Claude Fable 5, the first Mythos-class model available to the general public — a tier the company positions above its Opus line, and the same underlying model as the Claude Mythos system that has so far been restricted to vetted cyberdefence organisations under Project Glasswing (Anthropic, TechCrunch). What separates Fable 5 from its restricted twin, Claude Mythos 5, is not the weights but the wrapper: classifier-based safeguards that intercept requests touching cybersecurity, biology and chemistry, or suspected distillation attempts, and hand those queries to Claude Opus 4.8 instead. Anthropic says the fallback fires in under 5% of sessions, that a 1,000-hour bug bounty produced no universal jailbreak — though the UK AISI "made progress towards one" in a brief testing window — and prices both models at $10 per million input tokens and $50 per million output: double Opus 4.8, but less than half the Mythos Preview rate.
Two details matter more to a European regulated buyer than the benchmark table. First, capability tiering is now explicit product architecture: the unrestricted version of the same model is reserved for organisations the vendor (in consultation with the US government) chooses to trust, with a broader "trusted access programme" promised. Second, Anthropic will require 30-day data retention on all Mythos-class traffic — on first- and third-party surfaces, explicitly overriding existing zero-retention agreements — as a safety measure against novel jailbreaks. Whatever the merits, that is a unilateral change to a contractual term many EU institutions treated as load-bearing for GDPR and professional-secrecy analyses, and it sets a precedent: access to the most capable models now comes bundled with the vendor's retention policy, not yours. It is exactly the kind of discretionary term that makes the case for keeping a parallel open-weight stack on infrastructure you control — so the workloads that can't accept the terms have somewhere to run.
The Token Bill Comes Due: A Standards Body for AI Cost Discipline
Per-token prices have fallen roughly 98% since late 2022 — GPT-4-equivalent performance now costs about $0.40 per million tokens — and yet enterprise AI bills are up an estimated 320%, because agentic workloads multiplied consumption faster than prices fell (TheNextWeb, TechCrunch). The FinOps Foundation says companies were already three times over their full-2026 token budgets by April–May; Uber reportedly exhausted its year's AI coding budget by April; and GitHub Copilot's repricing was abrupt enough that one enterprise nicknamed it the "Tokenpocalypse" (TechCrunch). The Linux Foundation's response is institutional: a new Tokenomics Foundation, a standards body aiming to do for token spend what FinOps did for cloud cost management.
The timing is pointed. Fable 5 lands at double Opus pricing in the same week OpenAI moved its workspace agents onto credit-based pricing from 6 July, and the most capable models increasingly decompose one request into many billable sub-tasks — so headline per-token rates understate real cost growth. For European deployments the practical lesson is the one this brief has tracked all month: with capability plateauing at the top, model choice is becoming a per-workload cost-and-jurisdiction decision, not a loyalty decision, and routing the easy 80% of traffic to models that cost a tenth as much is now a finance-department mandate rather than an engineering optimisation.
Microsoft's Open-Source Repos Were Weaponised Against AI Developers
Microsoft disabled 73 GitHub repositories across four of its organisations — Azure, Azure-Samples, Microsoft and MicrosoftDocs — after attackers injected password-stealing malware into open-source projects widely used by developers building with AI coding tools (TechCrunch). According to analyses from Cloudsmith and OpenSourceMalware, the malicious code harvested credentials when the compromised tools were opened inside AI development apps, including Claude Code, Gemini's CLI and VS Code. It is Microsoft's second such incident in weeks, after May's Durable Task compromise — and the choice of target is the story: the attackers went after the people wiring agents into production.
The lesson is not "open source is unsafe" — proprietary supply chains fail too, and these were Microsoft-stewarded repos, not abandoned hobby projects. The lesson is that the agentic toolchain has become a high-value attack surface, and that provenance checking, dependency pinning and curated internal registries are now table stakes for any organisation letting coding agents pull tooling at will. Agents that install their own dependencies inherit their operators' credentials; a curated catalogue with a human gate is cheap insurance by comparison.
Quick Hits
- Apple and Brussels are now blaming each other for the Siri AI gap. Following Monday's EU-exclusion news, the Commission's spokesperson said "the decision not to roll out Siri AI in the EU is Apple's and Apple's only" and called the DMA non-negotiable, while Apple argues the required interoperability would force it to open user data and system functions to rival assistants — no EU timeline exists (AP, MacObserver).
- SpaceX prices tomorrow. The IPO prices 11 June at a fixed $135/share and trades Friday as SPCX (~$75B raise, ~$1.75T valuation, the largest listing on record); fresh AI line: Google will reportedly pay SpaceX $920M per month for compute (CNBC, TechCrunch).
- MiniMax M3's weights hit their due date. The promised open weights and technical report (within 10 days of the 1 June launch) are due by tomorrow and still pending — the open-weight cohort's next reproducibility test (MiniMax, the-decoder).
- Gemini 3.5 Pro is still the missing frontier model. Google's I/O promise of a June GA remains unshipped — 2M-token context and Deep Think reasoning in limited Vertex preview, with pricing expected around $15/$60 — leaving Fable 5 a clear window at the top of the price ladder (TechTimes).
- 53 days to the EU AI Act's GPAI obligations. The 2 August date holds — it adds Commission enforcement to the general-purpose-model transparency and documentation duties that have covered models like Fable 5 since placement on the market — while the Digital Omnibus postponing the high-risk regime to 2027/2028 still awaits Official Journal publication (European Commission).