AI Briefing

Two days, two frontier models. OpenAI dropped GPT-5.5 just six weeks after 5.4, and now DeepSeek has answered with V4, a 1.6-trillion-parameter open-source giant Huawei is backing with its newest Ascend chips. Meanwhile Anthropic's 'too dangerous to release' Mythos model got breached on day one, Devin-maker Cognition is shopping for a $25B valuation, and Geoffrey Hinton told the UN that AI is a fast car with no steering wheel. The race and the panic are accelerating in lockstep.

01Open SourceCNBC ↗

DeepSeek Drops V4, a 1.6-Trillion-Parameter Open-Source Answer to GPT-5.5

Exactly one year after DeepSeek R1 rattled Silicon Valley, the Hangzhou lab released preview versions of its V4 model in both Pro (1.6T parameters) and Flash (284B) flavors. V4 introduces a new 'Hybrid Attention Architecture,' a 1M-token context window, and claims top-tier coding and agent benchmarks. Critically, it remains open-source, and Huawei announced same-day 'full support' with its Ascend 950 Supernode clusters, turning V4 into the clearest signal yet that China now has both the models and the domestic silicon to compete at the frontier without US chips.

02ProductTechCrunch ↗

OpenAI Ships GPT-5.5 Six Weeks After 5.4, Pitches It as the Foundation for a 'Super App'

On Thursday OpenAI released GPT-5.5 to paid ChatGPT and Codex users, hitting 88.7% on SWE-bench and claiming 60% fewer hallucinations than 5.4. Greg Brockman framed it as a big step toward the company's long-rumored 'super app,' saying users can now hand it 'messy, multi-part tasks' and trust it to plan, use tools, and self-check. The cadence is unprecedented: a frontier model release barely six weeks after the previous one, at doubled pricing ($5/$30 per million tokens) that bets enterprises will pay up for agentic reliability.

03IndustryTechCrunch ↗

Anthropic's 'Too Dangerous to Release' Mythos Model Gets Breached on Day One

Days after Anthropic locked Mythos behind Project Glasswing because it independently found thousands of zero-days, Bloomberg reported that a Discord-based group already has it. Access came through a third-party vendor, reportedly helped by a contractor inside the group who used leaked knowledge about Anthropic's storage practices to locate the model. Anthropic confirmed it's investigating and says there's no evidence its own systems were compromised, but the incident is a brutal proof point that even the strictest model-release policies leak through their vendors.

04FundingSiliconANGLE ↗

Cognition AI in Talks to Raise at $25B Valuation, More Than Doubling in Seven Months

Bloomberg reports that Cognition, the startup behind autonomous coding agent Devin and the new owner of Windsurf, is in early talks to raise hundreds of millions at a $25B valuation. That's up from $10.2B in September 2025 and just $4B in March 2025. The pace illustrates how quickly investor appetite for coding agents has widened: Cursor is reportedly raising at $50B, GPT-5.5 just shipped with agentic coding as its headline feature, and Cognition wants to cement itself as the independent AI-engineer pureplay before OpenAI and Anthropic swallow the category.

05PolicyUN News ↗

Geoffrey Hinton Tells UN Conference AI Is 'A Fast Car With No Steering Wheel'

Speaking at the UN-co-organized Digital World Conference in Geneva, Nobel laureate Geoffrey Hinton warned that rapid AI advances have outpaced any meaningful governance, comparing the moment to a car with no brakes and no steering wheel. He estimated that only about 1% of AI work is going into safety research, flagged call-center automation as an imminent employment shock, and said it remains unclear whether humanity can coexist with superintelligent systems. The warning landed during a week when OpenAI and DeepSeek both shipped frontier models and Anthropic's most powerful system leaked to an unauthorized Discord group.