Sakana Fugu API Pricing
Sakana AI describes Sakana Fugu as a multi-agent system delivered as one model API. Fugu dynamically orchestrates a pool of expert models for coding, reasoning, research, security analysis, and other complex multi-step work. Since September 11, 2026 it comes in two fixed-price tiers: Fugu Ultra for the highest capability and Fugu Max for the lowest cost.
Fugu Ultra
Sakana prices Fugu Ultra as a fixed pay-as-you-go card: $5/M input, $30/M output, $0.50/M cached input, and $10/$45/$1.00 once context exceeds 272K.
Fugu Max LAUNCHED SEP 2026
Fugu Max is Sakana's cost-optimised multi-agent model, released 2026-09-11 alongside Fugu Ultra v2: one OpenAI-compatible model API that routes each task across Sakana's largest pool of open-weight…
Sakana Namazu
Sakana's first in-house single model, as opposed to Fugu, which orchestrates other labs' frontier models.
| Model | Input /M | Output /M | Cached | Context | Max output | Vision | Tools | Tier |
|---|---|---|---|---|---|---|---|---|
| Fugu Ultra FLAGSHIP | $5.00 | $30.00 | $0.50−90% | 1M | — | ✓ | ✓ | Active |
| Fugu Max | $2.00 | $6.00 | $0.25−88% | 1M | — | ✓ | ✓ | Active |
| Sakana Namazu | $0.95 | $4.00 | $0.15−84% | 256K | — | ✓ | ✓ | Active |
Pricing across the lineup.
Sakana Fugu pricing notes.
AI//COST tracks the Fugu tiers Sakana sells at a fixed token price - Fugu Ultra and, since September 2026, Fugu Max - plus the single model Sakana Namazu. Regular Fugu remains variable-rate because its price depends on the active underlying model tier.
$2/$6 with $0.25 cached at any context length; the fugu-ultra id moves to v2.0 on its unchanged card$0.95/$4.00 with $0.15 cached; built on Kimi K2.6 and priced at that base model's own rateOpenAI-compatible API for Fugu and Fugu Ultra.
PRICINGCanonical source for fixed Fugu Ultra token prices and variable Fugu billing notes.
TECHNICAL REPORTResearch background for learned orchestration and multi-agent coordination.
Sakana Fugu is positioned as "Multi-Agent System as a Model": one OpenAI-compatible model API that dynamically coordinates a pool of expert models for complex, multi-step tasks. The public page emphasizes coding, reasoning, Kaggle competitions, paper reproduction, cybersecurity analysis, and literature or patent investigations.
AI//COST tracks three token-priced rows here. Fugu Ultra carries a fixed pay-as-you-go card with a higher step above 272K of context; since September 11, 2026 the fugu-ultra alias serves v2.0 on that same card, while the dated fugu-ultra-20260615 id remains the vendor's name for v1.0. Its orchestration tokens are billed on top of visible input and output. Fugu Max, released the same day, is the cost-first tier: a flat card at any context length over Sakana's largest pool of open-weight and specialised models, with web search and fetches at $0.007 per call. Sakana Namazu, added in August 2026, is the lab's first in-house single model rather than an orchestrator: Japanese-specialised, built on Kimi K2.6, pay-as-you-go only, with built-in tools billed separately at $7.00 per 1,000 search calls and $0.12 per code-execution hour. Regular Fugu remains variable-rate: Sakana says a single active agent is billed at that underlying model rate, and multiple active agents are billed as one rate based on the top-tier model involved.
OpenAI
GPT-6 Sol lists at $2/$10 per million, the same input rate as Fugu Max and well under Fugu Ultra, without Sakana's orchestration on top. Sakana says GPT-6 Astra is not in Fugu Ultra v2's pool.
AGENT MODELAnthropic
Sakana benchmarks Fugu Max against Claude Sonnet 5 on price, and says Fugu Ultra v2 reaches its scores without Claude Fable 5 or Fable 5.1 in its pool - the closed models it is measured against rather than built on.
LONG CONTEXTGemini provides lower-cost long-context alternatives when a single-model workflow is enough.
OPEN-WEIGHT GIANTAlibaba (Qwen)
The Qwen (Tongyi Qianwen) model family from Alibaba Cloud's Tongyi Lab. First released in 2023, Qwen is the most-downloaded open-weight model family in the world — most tiers ship under Apache 2.0 on Hugging Face and ModelScope, while the proprietary Max tier is served through Alibaba Cloud's Model Studio.
GLM ARCHITECTZhipu (Z.ai / GLM)
The Beijing lab behind the GLM models, also operating internationally as Z.ai. Spun out of Tsinghua University in 2019, Zhipu ships strong coding/agentic models, multiple free tiers, and open weights — and was the first of China's "Six Tigers" to pursue an IPO. It also sits on the US Entity List.
MEDICAL-AI SPECIALISTBaichuan
The Beijing lab behind the Baichuan general models and the medical M-series. Founded in 2023 by Wang Xiaochuan (former CEO of search engine Sogou), Baichuan has pivoted hard toward healthcare AI — its open-source Baichuan-M3 tops several medical benchmarks — while keeping a general-purpose 4-series lineup.
Frequently asked.
Practical notes on fixed Fugu Ultra prices, variable Fugu billing, and availability limits.
Q · 01 What Fugu prices are tracked here? +
Q · 02 Why is regular Fugu not shown as a normal fixed price? +
Q · 03 Does Fugu Ultra have a long-context surcharge? +
$10/M input, $45/M output and $1.00/M cached input for Fugu Ultra. Fugu Max has no such step: its $2/$6 card applies regardless of context length.