Fugu Max vs Claude Sonnet 5

These two land on exactly the same input price, $2 per million tokens, and diverge on everything else. Claude Sonnet 5 is one model at $2 and $10 with a 1 million token context, adaptive thinking up to xhigh effort, and vendor-reported scores of 85.2 on SWE-bench Verified and 80.4 on Terminal-Bench 2.1. Fugu Max, released September 11, 2026, is not a model at all: it is an orchestrator that routes each task to the leanest model in a large pool weighted toward open weights, including NVIDIA Nemotron. Its output rate is $6, which Sakana says runs 40 to 60 percent under Sonnet 5, GPT-5.6 Terra, and Kimi K3, and cached input is flat at $0.25 at any context length against $0.20 for a Sonnet 5 cache hit. Sakana reports best overall score on six benchmarks including Terminal Bench 2.1, GPQAD, AA-LCR, GDP.pdf, AutomationBench, and its internal SWEFish, and says it expands the cost-performance frontier on seven of ten. Two things stop this from being a clean price win. Because the open models in its pool lack native web search, web_search and web_fetch bill at $0.007 per call, and Sakana says one query may need several calls, so the per-token rate understates any search-heavy request. And the router, not the caller, decides which model answers, so a team that needs one fixed model behind every call does not get that here. Against that, Anthropic prices one predictable path and publishes a longer track record, while Sakana offers a swappable pool that is genuine insulation against one vendor changing terms. There is no benchmark both ran on a shared harness, so no row below is like-for-like.