Sec & AI News — 26 June 2026
🟣 Sakana Ships a Model That Manages Other Models
Sakana AI launched Fugu. Not another chatbot — an orchestrator. One API that reads your prompt, picks the right frontier model, sometimes fans it out to several, then stitches the answer back. One provider goes down, it reroutes. You don't notice. Two tiers: Fugu for speed, Fugu Ultra for the hard multi-step stuff. Vendor benchmarks put Ultra at Fable 5 parity on some tests, ahead of Mythos on others. The asterisk: those export-restricted models aren't actually in Fugu's callable pool, so "parity" is a claim, not a head-to-head. Early testers also clocked 30-minute waits and burned $30 building one game and one website clone. Routing isn't free.
🔵 120+ AI Models, One Key, $0 — And Yes, There's a Catch
Routing across frontier models is one thing. Running models for free is another. So I set up a single API key that unlocks 120+ open models — DeepSeek, Kimi, Qwen, Llama — no credit card, no trial timer. The catch is exactly what you'd expect, and I'd rather you know it going in: rate limits, open-weight models only (no GPT-5 or Claude for free — those are closed, only their owners serve them), dev-and-test not production, and hosted never means private. I run frontier closed models as my daily driver, but I keep these free tiers wired up for teaching, so no student is ever locked out of learning by an API bill. I wrote up three ways to set it up — the easy path in OpenCode, the advanced path inside Claude Code, and a reusable script.
🟢 Anthropic Puts Claude In Your Slack And Calls It a Coworker
Claude Tag. You @mention Claude in a channel like it's a person. It breaks the task down, uses your company's tools, works in the background while you do something else. Everyone in the channel sees it, talks to it, redirects it. It remembers — no re-pasting context every morning. Turn on ambient mode and it volunteers, jumping in when it thinks it can help. Admins gate what it touches. Anthropic says 65% of their own code now ships this way. Karpathy called it the third redesign of how we use AI: website, then app, now a persistent entity that just hangs out on the team. Teams and Enterprise only for now. The rest of us wait.
🔴 The Government Now Decides Who Gets GPT-5.6
The Trump administration asked OpenAI to stagger GPT-5.6 over security concerns. Per a Thursday memo from Altman, the government approves access customer by customer during the preview. A broader release maybe a couple weeks later, if it goes well. The rumours had 5.6 landing any minute — Fable-class good. That's reportedly why everyone's being careful. OpenAI says it's not their preferred long-term model and they'll find a more sustainable approach. Translation: the wake-up-Wednesday-to-a-new-frontier-model era may be over. Someone in Washington now holds the release valve. We still don't have Fable back either.
🟠 That $38.5 Billion OpenAI Loss Is Mostly Fake
OpenAI's audited books leaked. The headline screamed a $38.5 billion loss. I pulled the numbers apart, and the scary figure is mostly accounting theatre — about $41.5B of it is a one-time non-cash charge from the non-profit-to-for-profit conversion. No cash left the building. The real story: $13B revenue, triple last year, and the company actually got more efficient as it scaled. The genuine risk isn't on the income statement. It's the ~$1 trillion in compute commitments and the circular financing loop — the same firms selling OpenAI chips are investing in OpenAI. That's where the word "bubble" earns its place. Here's what the leak actually means for anyone whose job or savings sit downstream of the AI boom — which is most of us.
⚡ Rapid Fire
Seed Dance 2.5 teased. ByteDance showed its next video model in Beijing. Thirty seconds of single-segment native video, double the current ceiling. Up to 50 reference assets — text, image, audio, video. Tighter motion and camera control. No release date. 2.0 already led the pack; this just widens it.
