Fugu Ultra v2.0: where it fits

Sakana AI's flagship multi-agent orchestration model, claiming best-or-joint-best results on 5 of 8 self-reported benchmarks against monolithic frontier models — $5/$30 per million tokens. Released September 10, 2026 by Sakana AI, it is built for sakana's own evaluation reports best-or-joint-best results on 5 of 8 benchmarks tested, including a 74.3 on the DeepSWE coding-repair benchmark, claims to outperform flagship models like GPT-6 Astra and Claude Fable 5.1 despite its underlying orchestration pool reportedly excluding both of them, and 1M-token context window.

Its trade-offs are real: not a single trained model — Sakana describes Fugu as "a Multi-Agent System, Delivered as One Model" that dynamically routes tasks across a pool of other models, so head-to-head benchmark framing against monolithic frontier models should be read skeptically, pricing rises to roughly $10/$45 per million tokens for prompts above 272K tokens, and benchmark comparisons against GPT-6 Astra and Claude Fable 5.1 are Sakana's own self-reported figures, not independently verified head-to-head scores. At $5 in / $30 out per million tokens, it sits in the premium price band.

Qwen3.6 27B: where it fits

A dense 27B multimodal model with its family's best coding score — it beats a 397B mixture-of-experts, but costs more per token. Released April 22, 2026 by Alibaba, it is built for the best open coding score in its family — 77.2% on SWE-Bench Verified, beating Alibaba's own 397B mixture-of-experts at a fifteenth of the size, dense, so quality per gigabyte of VRAM is high: it fits one consumer GPU when quantised, far stronger agentic work than its sparse sibling (59.3 against 51.5 on Terminal-Bench 2.0), and dense models fine-tune far more predictably than mixture-of-experts models do.

Its trade-offs: every parameter fires on every token, so it is slower and costlier per token than the sparse 35B, hosted output pricing is the harshest in its family, and provider input prices moved by roughly half in a single quarter, and its SWE-Bench score comes from Alibaba's internal scaffold rather than the standard public harness. As an open-weight model, its running cost is your own hardware rather than a per-token fee.

The bottom line for this matchup

The defining split here is open vs. closed. Qwen3.6 27B gives you weights you control — self-host it, fine-tune it, keep data in-house, pay only for hardware. Fugu Ultra v2.0 gives you a managed, always-updated API with no infrastructure to run. Teams with GPUs, privacy requirements, or huge volume often favour the open model; teams that want zero ops and the latest capabilities favour the closed one. Capability is close enough that this operational question, not the benchmark, usually decides it.

Related comparisons