GPT-5.5 Beat Mythos. Google Just Bet $40B on Anthropic.
Two weeks earlier, Oscar and I called GPT-5.5 noise. It shipped Thursday and beat Anthropic's gated Claude Mythos on Terminal-Bench. The same day, Google put forty billion dollars into Anthropic and DeepSeek dropped a 1.6-trillion-parameter open-source model that runs on Chinese chips. When we're wrong, we say so — so we called every one of them again.
Here's how I read it.
Signal or Noise
GPT-5.5 (Spud) ships and narrowly beats Claude Mythos on Terminal-Bench 2.0
We under-called this one two weeks ago, and it's worth owning. The frontier lead is measured in weeks now, and it changes hands fast. Signal.
DeepSeek V4 ships open source on Huawei chips — 1M context, one-sixth the inference cost
This is the strategically important one. Open weights, running on non-US silicon, at a fraction of the cost. You can't un-ship that. Signal.
Google invests up to $40B in Anthropic
The cloud war just moved inside the lab. When the hyperscalers own stakes in the model makers, "neutral infrastructure" gets complicated. Signal.
Cognition (Devin) in talks at a $25B valuation
Autonomous coding keeps attracting serious money. Signal, lightly.
Grok 4.3 Beta lands at $300/month — still no cross-session memory
Premium price, missing a basic capability. Noise until it remembers what you told it yesterday.
Stack Check
What I'd reach for:
- Vercel Labs Agent Browser — for agents that need to actually drive a browser.
- WorkOS FGA + AuthKit — fine-grained authorization, which you'll need the moment agents touch real data.
- AI SDK (ai-sdk.dev) — the pragmatic way to wire models into a TypeScript app.
- Claude Code /remote-control — worth a look for driving Claude Code from elsewhere.
The through-line: the frontier lead is now a matter of weeks, and the open models on cheap silicon are the ones that compound. Watch or listen below.
