Three significant stories defined an otherwise quiet week in AI: a new orchestrator model from Sakana AI that routes prompts across multiple providers, Anthropic embedding Claude directly into Slack workflows, and a sobering report that the Trump administration is intervening in how OpenAI rolls out its most capable models. Here's what you need to know.
Sakana AI's Fugu: An Orchestrator That Routes Across Models
Sakana AI launched Fugu — not a single new language model, but an orchestrator that decides which underlying model to send your prompt to, and can route to multiple models simultaneously. If one provider goes down, Fugu automatically reroutes, so you never lose progress mid-project.
Two tiers are available. Fugu balances performance with low latency — faster, cheaper, and well-suited for everyday tasks. Fugu Ultra is tuned for maximum answer quality on complex, multi-step problems where you need the model to reason carefully over a longer period.
02:10
Benchmark comparison table showing Fugu and Fugu Ultra performance against Gemini Flash 2.5 and Claude Opus on LiveCodeBench, Google-Proof Q&A, SWEBench Pro, and other metrics
Watch at 02:10 →
On benchmarks, Fugu and Fugu Ultra beat Gemini Flash 2.5 on LiveCodeBench and outperform Claude Opus on Google-Proof Q&A. Performance is roughly on par with Gemini Flash 2.5 on SCode, though it falls slightly short on SWEBench Pro. Overall, the orchestrator system is competitive with — and in some cases ahead of — the best available models.
Testing Fugu in Practice
You can access Fugu via console.sakana.ai. After adding billing details and generating an API key, the simplest path is through the Codex CLI. Running codex --fugu puts you in the OpenAI Codex CLI pointed at Fugu, and you can switch to Fugu Ultra with the /model command.
04:15
Terminal window showing Codex CLI setup with Fugu model selected, running a Mega Bonk clone build prompt
Watch at 04:15 →
Two builds were tested simultaneously using Fugu Ultra on the highest setting: a clone of the game Mega Bonk, and a rebuilt version of the FutureTools.io website. After roughly an hour, both were complete — but the combined cost came to $30 in API credits (22 million input tokens, 210,000 output tokens). That's a meaningful budget to burn through on experimental builds.
The Mega Bonk clone came out as a polished Vampire Survivors-style game rather than the 3D experience the prompt intended — complete with an upgrade shop, multiple heroes, and working level-up mechanics, but lacking the Three.js 3D rendering that a comparable Gemini Flash build had produced. The FutureTools clone replicated nearly all site functionality: working category filters, a shortlist feature, newsletter opt-in, and live AI news — though the visual design was more cluttered than the original. For $30, the results demonstrate real capability, but Fugu Ultra on maximum settings is not a casual-use configuration.
Claude Tag: Anthropic Brings Claude Into Slack as a Team Member
Anthropic announced Claude Tag, a feature that lets you mention Claude in a Slack channel the same way you'd tag a colleague. You give it a project, it breaks the work into steps, uses connected company tools, and continues working in the background while you move on to other tasks.
Key capabilities include persistent memory — Claude learns your company's context the longer it lives in a channel, eliminating the need to re-paste context into every conversation. It can also take initiative: when enabled, Claude monitors channels it has access to and proactively offers help when it identifies a relevant opportunity. Admins retain tight control over what data and tools Claude can access.
Anthropic reports that 65% of their internal code is now written using this feature. Andrej Karpathy described it as a third paradigm shift in how we interact with AI — from web chat interfaces, to standalone apps, to AI that lives natively inside the tools where work already happens. Claude Tag is currently available on Teams and Enterprise plans only.
Government Oversight of GPT-5.6 Rollout
The week's most consequential story may be the least technically exciting. The Trump administration has asked OpenAI to stagger the release of new models over national security concerns. In an internal memo, Sam Altman told staff that the government would approve access customer by customer during the preview period for GPT-5.6, with hopes for a broader release a few weeks later if that phase goes smoothly.
GPT-5.6 had been rumored to arrive imminently, with some expecting it this week or next. Those timelines are now uncertain. Early reports suggest the model is comparable in capability to Gemini Flash 2.5 — which may explain the extra caution, given that Gemini Flash 2.5 was itself restricted and has yet to return to general availability.
OpenAI stated publicly that customer-by-customer approval is "not a preferred long-term model" and that they will work with the government toward a more sustainable approach for future releases. Still, many observers see this as a turning point — the end of the era when a powerful new model could appear on a Tuesday with no warning. Government involvement in release sequencing, if it becomes standard, would represent a fundamental change to how frontier AI reaches users.
Rapid-Fire: Other Stories This Week
- Seed Dance 2.5 preview: Teased at a conference in Beijing, the next version of the leading video generation model doubles single-segment output to 30 seconds, supports up to 50 multimodal reference assets (text, image, audio, video), and promises finer control over motion, camera, and post-generation editing. No release date announced.
- Crea AI goes open weights: Crea released the Crea 2 image model as open weights, allowing anyone to download, fine-tune, or train custom styles and likenesses into it — similar to what has long been possible with Stable Diffusion and Flux models.
- The Atlantic's AI training data search tool: The Atlantic published a searchable database of music and video content used to train AI audio models. Users can look up specific artists or creators to see how many of their works appear in training datasets. Blink-182 accounts for 150 songs; various YouTubers' videos appear as well.
- OpenAI and Broadcom building custom inference chips: The two companies announced a jointly developed inference chip, purpose-built for ChatGPT. This is separate from their existing partnership with Cerebras and their reliance on Nvidia hardware (which is used primarily for training). Custom inference silicon should make responses faster, cheaper, and potentially more capable.
- Meta Ray-Ban expands to 26 styles: Meta announced a significant expansion of its AI glasses lineup beyond the original Ray-Ban collaboration, including new frames designed with Kylie Jenner and styles named Adventurer and Fury, across a range of colors and lenses.
