New numbers caught my eye this week: 65% of Anthropic’s own code is now written through a single Slack feature. That one stat tells you where AI is heading, and it was buried in a quiet news week.
I pulled these stories from a weekly AI roundup by the creator behind Future Tools, Matt Wolfe. He drinks from the firehose all week so the rest of us don’t have to. I was genuinely surprised by how big the stakes were for such a light news cycle.
Here’s the breakdown of what actually matters.
🔹 Sakana’s Fugu model
The creator dug into Sakana AI’s new “Fugu” release, and it’s not a normal model. It’s an orchestrator that reads your prompt and routes it to the best model under the hood, sometimes several at once. If one provider goes down, it reroutes so you lose zero progress. It ships in two flavors: regular Fugu (fast, cheap, everyday) and Fugu Ultra (slow, deep, hard multi-step problems). On benchmarks it beats Fable 5 on Live Code Bench, edges out Mythos on Google-proof Q&A, and ties Fable on Sci Code.
🔹 Anthropic’s Claude Tag
The expert also flagged a new Anthropic feature that lets you tag Claude inside Slack like a coworker. It breaks projects into steps, works in the background, remembers context over time, and can jump in when it spots a chance to help. Admins control exactly what it touches. That 65% internal-code stat? This is the tool behind it.
🔹 Government steps into model releases
The one that rattled him: the US government reportedly asked OpenAI to stagger GPT 5.6’s rollout, approving access customer by customer. Many think this marks the end of the surprise-Wednesday-model era.
Three ways you can act on this now:
- Test Fugu yourself. The creator set it up through console.sakana.ai, added pay-as-you-go billing, generated an API key, then ran it via the Codex CLI with a simple “codex-fugu” command.
- Use Fugu Ultra only for genuinely hard jobs. For quick tasks, regular Fugu saves real money.
- If you’re on a Claude Teams or Enterprise plan, try Claude Tag in a low-stakes channel first to see how it learns your workflow.
Quick wins and pitfalls:
- Pro: Fugu’s auto-rerouting means provider outages stop killing your projects.
- Pro: Claude Tag keeps context so you stop re-pasting the same background.
- Con: Cost adds up fast. The creator ran two builds on Fugu Ultra at “extra high” and burned through $30 in about an hour (22 million input tokens). His honest take: maybe don’t start on max settings.
- Con: Results were mixed. His “Mega Bonk” clone came out as a solid Vampire Survivor-style game instead, and he felt Fable still handled the 3D graphics better.
Rapid fire he also covered: Seed Dance 2.5 teased with 30-second clips and up to 50 reference assets, Krea 2 going open weights, The Atlantic’s searchable database of AI training music, OpenAI and Broadcom building their own inference chip, and a wave of 26 new Meta Ray-Ban styles.
Want the full head-to-head and the on-screen build results? Watch the original video, it’s worth your time.