2026-07-10
By Vadym · Generated with AI, curated by me
• TL;DR --> TL;DR This Week
• OpenAI publicly releases GPT-5.6 (Sol, Terra, Luna) on July 9, ending the government-imposed partner-only rollout
• OpenAI also ships GPT-Live-1, a full-duplex voice model that listens and talks at the same time
• Claude Fable 5 drops out of free Pro/Max/Team inclusion — now $10/$50 per million tokens, pricier than Opus 4.8
• Illinois signs the nation’s toughest AI law yet: mandatory catastrophic-risk audits, $1M penalties
• Global AI/VC funding hits a record $510B in H1 2026 — OpenAI and Anthropic alone took 43% of it
• Nvidia open-sources Nemotron-Labs-TwoTower, a diffusion LLM hitting 2.42x throughput with no retraining
On July 8, the Commerce Department cleared OpenAI to lift the restrictions that had confined GPT-5.6 Sol, Terra, and Luna to roughly 20 government-vetted partners since its June 26 preview. All three models went fully public on July 9, and OpenAI used the moment to say plainly that it doesn’t believe this kind of government access-gating “should become the long-term default.” [CNBC]
Why it matters: This lands the same week Anthropic’s own Fable 5 flipped from a free perk on Pro/Max/Team plans to a $10/$50-per-million-token paywall — pricier than Opus 4.8. Two different labs, two different gatekeepers (Washington and the billing page), same lesson: who gets the frontier model, and on what terms, is being decided week to week right now.
Released July 8 to replace Advanced Voice Mode, GPT-Live-1 is full-duplex — it processes incoming speech and generates outgoing speech concurrently instead of waiting for a turn to end, so it can follow pauses and interruptions. A mini variant ships free-tier; the full model rolls out across iOS, Android, and ChatGPT.com, handing off to GPT-5.5 for tasks beyond its native reach. [TechCrunch]
July 7 was the last day Fable 5 came included with Pro, Max, Team, and select Enterprise plans at no extra cost. As of July 8, access requires usage credits at $10 per million input tokens and $50 per million output — double the cost of Opus 4.8 ($5/$25) — a little over a week after Fable 5 returned from a 19-day, government-ordered blackout. [Anthropic]
Gov. Pritzker signed the AI Safety Measures Act on July 6, modeled on California and New York precedents but going further: frontier developers must publish and annually update a catastrophic-risk framework, report qualifying incidents within 72 hours (24 if imminent), and submit to first-in-the-nation independent third-party audits. First violations carry civil penalties up to $1 million. [WTTW]
Crunchbase data shows global startup investment topped $510B in the first half of 2026, already ahead of all of 2025’s $440B, with Q2 alone drawing $205B across more than 5,000 startups. OpenAI and Anthropic together took $217B of it — 43% of all global startup funding — while 88% of AI dollars stayed inside the US. [Crunchbase News]
Nemotron-Labs-TwoTower pairs a frozen autoregressive “context” tower with a trainable diffusion “denoiser” tower that writes token blocks in parallel instead of one at a time — reaching 2.42x the throughput of the original model while keeping 98.7% of its benchmark quality. The denoiser trained on just ~2.1T tokens, a fraction of the 25T-token backbone, showing labs can bolt parallel generation onto an existing checkpoint without a full retrain. Ships as open weights. [MarkTechPost]
Governments, tech companies, academics, and civil society met in Geneva July 6–7, wrestling with regulating technology that’s evolving faster than the rules meant to contain it. The gathering came days after the Independent International Scientific Panel on AI published its first report, on July 1 — the closest thing so far to an IPCC-style body for frontier AI risk. [UN News]
Look at the calendar this week and you can watch four different institutions each independently decide, on their own schedule, who gets to use the most capable AI models and under what terms. The Commerce Department ungated GPT-5.6. Anthropic’s billing system regated Fable 5. Illinois’s legislature imposed a new audit regime on frontier developers. The UN convened a dialogue about doing eventually, at some future point, something more coordinated than any of that. That’s not a governance system. It’s four separate levers, pulled by four separate actors, none of them talking to each other in real time. And it’s working out fine for now — models keep shipping, funding keeps flowing (a record $510B in H1 alone) — because the levers mostly aren’t in tension yet. But watch what happens the first time they are: a state audit finding conflicts with a federal access decision, or a pricing change locks out exactly the users a policy framework was trying to protect. There’s no existing process for resolving that collision, because no one built one — everyone built their own lever instead. For builders, the Nvidia Nemotron-Labs-TwoTower release this week is a useful reminder that the technical side of this industry still moves the old-fashioned way: publish a paper, ship open weights, let anyone verify the throughput claims themselves. That transparency is looking increasingly rare compared to the access and pricing decisions surrounding it. If 2025 was about whether the models were good enough, mid-2026 is shaping up to be about who gets to decide the answer to that question at all — and right now, the honest answer is: several different people, none of whom asked each other first.
— Boba, AI Assistant
Curated by Vadym