Sakana's Fugu Max & Ultra v2: top benchmarks, a fraction of the cost
Fugu is a learned multi-agent router behind an OpenAI-compatible API, not one model — swap it in with one line. Plus: Codex desktop adds quick chat.
Copy markdown
Fugu Max: benchmark-topping agents at $2/$6 per million
Sakana's cost tier tops six agentic benchmarks — Terminal Bench 2.1, GPQAD, SWEFish, AutomationBench and more — while charging $2 per million input and $6 per million output. That output rate runs 40-60% below Sonnet 5, GPT-5.6 Terra and Kimi K3, so token-heavy agent loops get a direct bill cut.
Fugu Ultra v2 is an orchestra, not a soloist
Ultra v2 isn't a single model but a learned multi-agent router that farms each step out to a swappable pool of open-weight and specialist models (NVIDIA Nemotron among them). It posts DeepSWE 74.3 — beating models that cost 3-5x more per token — and Chartography 48.3 vs Opus 5's 27.3, with a 1M-token context and 128K-token output at $5/$30 per million (cache reads $0.50).
Drop-in: OpenAI-compatible, one line to switch
Both tiers speak the OpenAI API, so existing code moves over by changing base_url and the model name — no rewrite. You keep configurable reasoning effort, function calling, structured outputs, image and PDF input, and built-in web search, and they're already live on OpenRouter behind a single key.
Build this weekend: an agent you can leave running
Point your existing agent harness at Fugu Max and keep your tools — you get near-frontier coding and research at roughly a third of the per-token cost. That's what makes long, many-step jobs (repo-wide refactors, overnight research sweeps) cheap enough to actually let run to completion.
Elsewhere: Codex desktop adds quick chat and app sharing
OpenAI's Sept 11 Codex desktop update adds a floating quick-chat control, 'appshots' (Alt+Alt) to drop a specific app window straight into a conversation on Windows and macOS, and Codex Micro inline text insertion — small cuts to the day-to-day friction of driving the agent.