AI News: The New Model That's As Good As Fable

The host presents a weekly recap of underreported AI news, framing a relatively quiet week as nonetheless packed with major developments. Sakana AI launched Fugu, an orchestrator model that routes prompts to the best underlying model and automatically switches providers during outages; a hands-on test of Fugu Ultra produced a game and a Future Tools clone for roughly $30, with the game playing like a Vampire Survivors clone and the site clone preserving functionality almost perfectly. Nexo offers a consolidated AI dashboard with a no-code agent builder, while Claude's Slack integration lets it act like a teammate—a shift Andrej Karpathy calls a third major redesign in AI interaction. The Trump administration is reportedly asking OpenAI to stagger GPT-5.6 releases over security concerns, signaling an end to sudden, unvetted model launches. Finally, Seed Dance 2.5 was teased with major video editing improvements, though no access has been granted yet.

• ▶ 0:00 Host opens with "Here's the AI news you probably missed this week," framing the segment as a weekly recap of underreported AI developments.
• ▶ 0:05 He notes a deliberate format change: "let's change up the scenery a little bit," signaling a different presentation style for this episode.
• ▶ 0:11 Acknowledges a quiet week overall, but emphasizes at ▶ 0:13 that the few stories were "really, really big" and need attention.
• ▶ 0:34 Fugu is not a standalone model like a typical GPT or Claude release, but rather an orchestrator/manager model.
• ▶ 0:38 Its core function is routing intelligence: deciding which underlying model to send your prompt to, sometimes routing to multiple models at once.
• ▶ 0:46 Because it abstracts provider routing, Fugu automatically switches to another model if a provider like OpenAI or Anthropic has an outage, preventing loss of progress.
• ▶ 1:03 Sakana AI launched two model variants: Fugu and Fugu Ultra.
• ▶ 1:07 Fugu balances strong performance with low latency, making it a fast, cheaper default for everyday work, but with less extraordinary output.
• ▶ 1:18 Fugu Ultra is tuned for maximum answer quality on hard multi-step problems, prioritizing deep reasoning over speed and cost.
• ▶ 1:27 Fugu is an orchestrator model designed for complex, multi-step tasks, with its strength shown in benchmarks against Fable 5 and Mythos.
• ▶ 1:35 Fugu (and Fugu Ultra) outperform Fable 5 on Live Code Bench, beat Mythos on Google Proof Q&A, and match Fable 5 on Sci Code.
• ▶ 1:46 Fugu is weaker on Sweet Bench Pro, but overall performs as well as or better than Fable across most benchmarks.
• ▶ 2:15 Go to console.sakana.ai, set up billing (pay-as-you-go, e.g. $20), and generate an API key to access Fugu.
• ▶ 2:35 The easiest setup is via the Codex CLI: copy the command from Sakana's "Get Started" guide, paste it in your terminal, then run codex-fugu to launch Codex with Fugu as the active model.
• ▶ 3:02 Inside the Codex CLI, type /model to switch models—for example, to Fugu Ultra.
• ▶ 3:04 Matt tests Fugu Ultra with extra-high effort by prompting it to clone the game Megabonk.
• ▶ 3:20 In parallel, he launches a second task in another terminal: clone futuretools.io and improve its UI/UX, using the same "full on YOLO" settings.
• ▶ 3:48 After about an hour, both apps finish—producing Bonk Storm (Megabonk clone) and a Future Tools clone—and the run consumed roughly 22M input tokens and 210K output tokens.
• ▶ 4:08 The session's charges were driven primarily by output tokens generated during testing.
• ▶ 4:15 Wolf's two builds (Mega Bonk and Future Tools clones) cost roughly $30 in total, burning through his initial $20 credit and an extra $10 top-up.
• ▶ 4:18 His practical takeaway: don't start with Fugu Ultra's "extra high" setting — it racks up costs quickly, so use a more measured approach for first experiments.
• ▶ 4:21 Fugu's "Bonk Storm" is a functioning Mega Bonk-inspired game, including an upgrade shop and a single hero (Mallet Sprout) from the start.
• ▶ 4:48 The narrator admits it plays nothing like Mega Bonk, instead calling it a "spot-on Vampire Survivor clone" with working level-ups and upgrades.
• ▶ 5:19 Comparing tools, Fable offers better 3D graphics and camera control, while Fugu excels at level-ups, meta upgrades, and menu polish — but neither delivers a true Mega Bonk experience.
• ▶ 5:53 Gameplay is described as “quite addictive”; dying returns the player to the main menu.
• ▶ 5:56 The main menu features an upgrade shop where players spend currency on meta upgrades, such as “pocket gravity” (with 87 silver available).
• ▶ 6:05 Earning more currency allows players to purchase new heroes, like a “tin can” character.
• ▶ 6:11 Fugu generated the entire Future Tools site, but Matt does not consider the design an improvement, calling it more cluttered and confusing than his current site.
• ▶ 6:25 Core functionality works well: tool filtering by pricing/category, "Matt's Picks" reflects his real picks, and a category analytics graph is included.
• ▶ 7:16 Clicking tools correctly redirects to external pages, and the shortlist feature (shortlisting tools like ChatGPT) populates a sidebar and copies actual URLs — a feature Matt likes and wants to add to his real site.
• ▶ 8:06 Overall, the clone replicated the site's functionality "almost perfectly," though Matt prefers the light mode over dark in the clone.
• ▶ 8:13 The Fugu AI demo cost only $30 to generate, but Matt jokes the quality is rough and thinks Fable would have done a slightly better job.
• ▶ 8:30 Nexo is pitched as a single AI dashboard replacing multiple subscriptions—giving access to ChatGPT, Claude, and Gemini to avoid paying up to $200/month across separate tools.
• ▶ 9:21 Nexo's standout feature is a no-code agent builder: users describe a task in plain language and get a working AI agent, with integrations like Google Drive and Slack for context.
• ▶ 10:04 Claude tag lets you tag Claude in Slack like a human teammate, enabling it to break down projects, use company tools, learn context over time, and even take initiative in channels.
• ▶ 11:08 Admins get granular control over Claude's access, and Anthropic reports that 65% of its own code is now written using this Slack integration.
• ▶ 11:43 Andrej Karpathy calls it a third major redesign in AI interaction—moving from web chat and standalone apps to embedding AI directly into tools like Slack.
• ▶ 12:31 The Trump administration is asking OpenAI to stagger GPT-5.6 releases over security concerns, with customer-by-customer government approval during preview and a possible general release weeks later.
• ▶ 13:54 Both Fable and GPT-5.6 release dates remain uncertain, signaling a possible end to the "wild west era" of sudden, unvetted AI model launches.
• ▶ 14:37 Seed Dance 2.5 was teased with 30-second single-segment video, up to 50 multimodal reference inputs, and major control/editing improvements—but no access or release date yet.
• ▶ 15:26 Krea 2 was released as open weights, letting users download, run, fine-tune, or train custom content into the model like Stable Diffusion and Flux.
• ▶ 15:53 The Atlantic launched a free, searchable database of music used to train AI models, targeting companies like Suno and Udio, available at theatlantic.com/category/aiwatchdog.
• ▶ 17:15 OpenAI and Broadcom are developing their own inference chip, competitive with Cerebrus, which could make ChatGPT faster, cheaper, and smarter—though Nvidia still handles training.
• ▶ 18:06 Meta announced 26 new styles for Ray-Ban smart glasses, including a Kylie Jenner-designed model and new Meta Adventurer and Meta Fury frames.
• ▶ 18:33 Although it wasn’t a news-heavy week, major stories like Sakana AI, the Claude tag feature, and OpenAI’s government review made it high-impact.
• ▶ 18:56 Matt curates AI news all week so viewers get a concise, digestible Friday recap without the overwhelm.
• ▶ 19:19 Viewers are asked to like and subscribe; AI tutorials drop midweek and news videos drop Fridays.

• ▶ 0:00 Introduction & Sakana Fugu Deep Dive (0:00 - 8:10) - AI news intro, Sakana Fugu model details, benchmarks, setup/testing, and generated clones.
• ▶ 8:10 AI Game/Website Demo & Nexo Sponsor (8:10 - 10:03) - Showcases a cheap AI-generated game/website and presents the Nexo AI sponsor segment.
• ▶ 10:03 Anthropic Claude Slack Integration (10:03 - 12:20) - Covers Claude in Slack and the broader redesign of AI integrations.
• ▶ 12:20 OpenAI GPT-5.6 & Rapid-Fire Previews (12:20 - 15:53) - Details GPT-5.6's staggered release and previews Seed Dance 2.5 and Krea 2.
• ▶ 15:53 Atlantic Music DB, OpenAI Chip & Meta Glasses (15:53 - 18:36) - Rapid-fire news on The Atlantic's music database, OpenAI/Broadcom chip, and Meta Ray-Ban styles.
• ▶ 18:36 Summary & Wrap-Up (18:36 - 19:53) - Recaps the week's big stories, release schedules, and transitions to the next segment.