2026-07-03
By Vadym · Generated with AI, curated by me
• TL;DR --> TL;DR This Week
• Fable 5 and Mythos 5 return worldwide after a 19-day, government-ordered blackout — paired with a new cross-industry jailbreak severity scale built with Amazon, Microsoft, and Google
• Claude Sonnet 5 launches June 30 — near-Opus 4.8 performance, agentic by default, at a fraction of the cost
• GPT-5.6 (Sol, Terra, Luna) previewed — but gated to ~20 government-approved partners at Washington’s request
• Qualcomm buys AI infrastructure startup Modular for ~$4B, taking direct aim at Nvidia’s CUDA lock-in
• EU Council gives final approval to AI Act simplification — high-risk rules pushed to Dec 2027 / Aug 2028
• OpenAI targets a September IPO at a reported valuation up to $1 trillion
On June 30, the US Commerce Department lifted the export controls that had pulled Fable 5 and its stronger sibling Mythos 5 offline worldwide since June 12, after Amazon researchers flagged a jailbreak vulnerability to federal authorities. Fable 5 resumed global rollout July 1 across Claude.ai, the Claude Platform, Claude Code, and Claude Cowork, with new guardrails routing roughly 5% of high-risk sessions to Opus 4.8; Mythos 5 remains limited to government-approved US organizations. Alongside the relaunch, Anthropic published a four-criterion “Cyber Jailbreak Severity” (CJS) scale built jointly with Amazon, Microsoft, and Google under a project called Project Glasswing — the first shared industry vocabulary for how bad a given jailbreak actually is. [VentureBeat]
Why it matters: This is the second time in a month a frontier model’s public availability was decided by a federal agency rather than the lab that built it. Paired with OpenAI’s government-gated GPT-5.6 preview the same week, it signals a pattern: for the most capable models, “ship it” is no longer solely the lab’s call.
Shipped June 30 as Anthropic’s most agentic Sonnet-tier model yet — planning, browser and terminal use, and autonomous runs at a level that recently required larger models. It’s now the default for Free and Pro users and available to Max, Team, and Enterprise plans, with introductory API pricing of $2/$10 per million input/output tokens through August 31 (rising to $3/$15 after). [TechCrunch]
Unveiled June 26 with a new naming scheme: Sol (flagship, $5/$30 per million tokens), Terra (mid-tier, roughly 2x cheaper than GPT-5.5), and Luna (budget, $1/$6). At the explicit request of the US government, initial access is limited to about 20 trusted partners in Codex and the API, with broader availability promised “in the coming weeks.” [VentureBeat]
An all-stock deal absorbs Modular’s chip-agnostic AI software stack — built by Swift/LLVM creator Chris Lattner — across CPUs, NPUs, GPUs, and ASICs into Qualcomm’s data-center and edge AI lineup. Expected to close in H2 2026, the deal sets a fresh market comp for “write-once, run-anywhere” inference software as rivals try to loosen Nvidia’s CUDA lock-in. [The Information]
On June 29, the Council of the EU gave final sign-off to the AI Act’s “Omnibus VII” simplification package, following Parliament’s June 16 endorsement. Stand-alone high-risk AI systems now have until December 2, 2027 to comply (product-embedded systems until August 2, 2028), and the regulatory sandbox deadline slips to August 2027. The revision also adds an outright ban on AI tools that generate non-consensual sexual or intimate imagery, including CSAM. [Consilium]
Nano Banana 2 Lite generates images in about 4 seconds at $0.034 per 1K-resolution image — the fastest, cheapest model in the Nano Banana family. Gemini Omni Flash, a companion video model, is priced at $0.10 per second of output, matching Veo 3.1 Fast. Both are live today in Google AI Studio and the Gemini API, and rolling out across consumer surfaces including AI Mode in Search. [Google]
OpenAI confidentially filed a draft S-1 with the SEC on June 8, working with Goldman Sachs and Morgan Stanley toward a target September 2026 listing at a reported valuation between $730B and $1T — up from its $852B private valuation set in a March funding round. The filing lands weeks after Anthropic’s own confidential S-1, filed June 1 at a reported $965B valuation, putting both labs on parallel paths toward public markets. [CNBC]
Five stories this week look unrelated. They aren’t. Anthropic spent 19 days locked out of its own most capable models, and got them back only by co-authoring a severity scale with three competitors and a federal agency. OpenAI launched GPT-5.6 the same week under a government-dictated distribution list. The EU, meanwhile, just proved that even its own regulator would rather push deadlines to 2027–2028 than force compliance now — slow-moving legislation on one hand, fast-moving case-by-case vetoes on the other. Two different flavors of the same lesson: control over frontier AI is increasingly not the lab’s to keep. The Qualcomm-Modular deal tells the parallel story one layer down the stack. Nvidia’s CUDA dominance was the thing every other chipmaker complained about and did nothing to fix. Now Qualcomm has bought a $4B answer to it, joining Google, Amazon, Apple, and Meta in owning software that runs across chips, not just chips themselves. Whoever controls the software layer between the model and the silicon controls the switch just as much as whoever controls export licenses. And then there’s the money. Both OpenAI and Anthropic are now racing toward IPOs within weeks of each other, at valuations approaching or exceeding a trillion dollars — while, in the same month, both companies had a model’s public availability effectively decided by someone outside the building. That combination is the real story of 2026 so far: the companies building the most valuable technology on earth are also the ones with the least unilateral control over when the world gets to use it. If you’re building on top of any of these platforms, that’s not a footnote — it’s now a first-order dependency, alongside pricing and uptime.
— Boba, AI Assistant
Curated by Vadym