Fugu AI Pricing: Every Rate, Plan, and Hidden Cost

Fugu’s pricing is unusual in three ways: the base model has no price list at all, the flagship bills tokens you never see, and the top rate roughly doubles once your context crosses a threshold. This guide is the full reference — every number verified against Sakana’s official pricing page on August 19, 2026. When Sakana changes a rate, we update this page; if a number matters to you, verify it on the official page before spending.

Base Fugu: no rate table, by design

Base Fugu has no fixed per-token price, because what you’re billed depends on what the orchestrator uses. The official rule has two halves:

• One active agent: you pay only the standard rate of the specific underlying model that handled your request.
• Multiple agents: fees are never stacked. You pay a single rate based on the top-tier model involved.

That second rule is friendlier than it might sound: a five-model collaboration doesn’t cost five model fees. But it also means base-Fugu costs are variable by nature — a simple request routed to a modest model is cheap; a hard request that pulls in a top-tier model bills at that model’s rate. Your effective average depends on your workload mix, and the only way to learn it is to run a representative batch and read the usage reports.

If your budget needs a hard number per token before you start, that’s what Fugu Ultra’s fixed table is for — predictability is part of what Ultra sells.

Fugu Ultra: the fixed rate table

Fugu Ultra has fixed pricing per million tokens, with two tiers depending on context size (as of August 2026):

Token type	Standard	Context > 272K
Input	$5	$10
Output	$30	$45
Cached input	$0.50	$1.00

The same table applies to both current Ultra versions, so you don’t need to track which build you’re on to budget. Three things to read out of it:

The output rate is six times the input rate. Verbose responses dominate most bills. If you’re paying for prose you don’t read, asking for concise output is a real cost lever, not a style preference.

The cache discount is 90%. Cached input bills at $0.50 instead of $5. Long system prompts, shared codebase context, and stable conversation prefixes get an order of magnitude cheaper when the prompt cache can serve them. If your application re-sends the same preamble on every request — most do — cache-friendly prompt structure is the single biggest saving available to you.

The long-context tier is a step, not a slope. When a request’s context exceeds 272K tokens, input doubles and output rises 1.5×. There’s no partial blending — plan workloads to stay under the line unless the task genuinely needs the room. (The full window is 1M tokens, so the capacity is there when you need it; you just pay tier-two rates to use it.)

The tokens you don’t see: orchestration usage

This is the fact that surprises almost everyone. Fugu Ultra is a multi-agent system, and the work its internal agents do — the routing, the inter-agent discussion, the verification — consumes tokens. Sakana’s pricing page is explicit: these orchestration tokens “represent real token usage outside of the input and output tokens and will be counted in the final price of the request,” billed at the same rates as standard input and output tokens.

In other words, the tokens you can count from your prompt and the visible response are a floor, not the bill. Every response’s usage object itemizes the orchestration usage, so the costs are auditable after the fact — but not precisely predictable before it, because harder tasks trigger more internal work.

There’s a second, related trap: capping the response length does not cap the cost. Output-token limits apply only to the final visible response; the orchestrator still uses its own internal maximum. We unpack the whole mechanism — the exact usage fields, the audit arithmetic, and the controls that do work — in the dedicated guide to Fugu orchestration tokens, and the cost calculator has an overhead slider so you can model it.

Pay-as-you-go vs. subscription plans

Sakana offers two ways to pay, and they differ on more than price.

Pay-as-you-go is the per-token billing described above, funded by prepaid credits. Sakana positions it for “heavy production workloads that need maximum reliability,” and backs that with a documented scheduling rule: consumption-based tokens are served at higher priority than monthly plan tokens. If latency and throughput matter, this is your lane.

Subscription plans are monthly tiers aimed at individuals and hands-on use. Every tier includes both Fugu and Fugu Ultra (as of August 2026):

Plan	Price	Allowance
Standard	$20/month	Baseline allowance
Pro	$100/month	10× Standard usage
Max	$200/month	20× Standard usage

Note what Sakana does not publish: a token-denominated definition of the “baseline allowance.” The tiers are defined relative to each other (10× and 20× Standard), so nobody can tell you exactly how much usage $20 buys — you find out by using it. For light personal use, a subscription is usually the cheaper and more predictable entry point; a flat fee also absorbs the orchestration-overhead variability that makes per-token Ultra costs hard to forecast. Our plan-by-plan comparison works through who should pick what.

Credits: how you actually pay

Pay-as-you-go usage is funded by prepaid Credits, and the terms of service attach real strings (as of August 2026):

• Credits expire six months after purchase, and unused credits are forfeited automatically at expiration (except where law requires otherwise).
• They are non-refundable, not redeemable for cash, and bound to your account — no transfers, sales, or gifts.
• Payments run through third-party payment processors; Sakana says it does not collect or store your card details itself.

Practical consequence: buy credits in portions you’ll spend within six months, and put the expiry date on a calendar if you pre-load a serious balance. “Use it or lose it” is written into the terms.

Fugu Cyber: pricing via sales only

Fugu Cyber, the security-specialized model, publishes no rates. Getting access requires a request form detailing your use case plus manual review; billing is pay-as-you-go only — a key must be in pay-as-you-go billing mode for the model to even appear in its model list — and pricing details go through Sakana’s sales team. Budget-wise, treat Cyber as an enterprise procurement, not a metered add-on.

Sakana Namazu: the cheap comparison point

Sakana Namazu isn’t part of the Fugu family — it’s a separate Japanese-specialized model — but it shares the console and makes a useful budget reference. As of August 2026: input $0.95, output $4.00, cached input $0.15 per million tokens, with thinking tokens billed at the output rate. Namazu is pay-as-you-go only; no subscription includes it.

Namazu’s built-in tools bill separately, and the accounting is worth understanding if you use them. Web search costs $7.00 per 1,000 calls — and each search query and each page-body fetch counts as one call, so a single “search” that reads five pages is six calls. Code execution costs $0.12 per hour of retained session. And because each internal tool-loop call re-sends the full prompt and conversation so far, tool-heavy requests generate substantial input-token usage — typically softened by the cache, since the shared prefix bills at the $0.15 cached rate.

What a real request costs: worked examples

A typical coding question to Fugu Ultra: a 2,000-token prompt, an 800-token answer. The visible tokens cost one cent for input ($5 × 0.002) plus 2.4 cents for output ($30 × 0.0008) — about 3.4 cents. Now add orchestration. If the orchestrator’s internal usage adds 150% overhead on top of the visible tokens (a mid-range assumption; real overhead varies by task), the request lands around 8.5 cents. A hundred such requests a month: roughly $8.50.

The same request with a cached preamble: if 1,500 of those 2,000 input tokens are a stable system prompt served from cache, visible input drops from $0.010 to $0.0033. Small per request — but cache savings apply to orchestration input too, and they compound across every request you send all month.

A long-context task: feed Ultra 300K tokens of codebase and the >272K tier applies to the request — the input alone is $3.00 at tier-two rates, before output and orchestration. Long-context work on Ultra is priced for occasions, not routine. If you’re doing it daily, restructure toward retrieval or caching.

The Namazu sanity check: the same 2,000-in/800-out request shape on Namazu costs about half a cent ($0.0019 input + $0.0032 output), with no orchestration overhead. When a task doesn’t need frontier-team reasoning, the price gap is enormous — which is exactly why knowing your workload mix matters more than knowing any single rate.

Run your own numbers in the calculator; it handles the cache discount, the overhead slider, and the long-context toggle.

Where these numbers come from

Sakana spreads its pricing facts across three official pages, which is worth knowing when you verify. The rate tables, plan tiers, and the orchestration-billing rule live on the pricing page. The credits fine print — the six-month expiry, non-refundability, and third-party payment processing — lives in the terms of service, not on the pricing page. And Fugu Cyber’s billing constraints (pay-as-you-go only, access gating) are documented on the models page, while the pricing page only says to contact sales. If you check one page and a claim seems missing, it’s probably on one of the other two — every figure in this guide was re-verified across all three on August 19, 2026.

How to audit your Fugu bill

Because Fugu’s costs have moving parts — routing on base Fugu, orchestration overhead on Ultra, cache hit rates everywhere — a monthly total on its own tells you very little. Auditing well means breaking the bill into questions you can act on.

What share of input was cached? Every response reports cached tokens separately. If your cache share is low and your prompts share long prefixes, restructure so the stable parts come first — the difference between $5 and $0.50 per million is the biggest lever on this list.

What is your orchestration overhead ratio per task type? Divide internal tokens by visible tokens, tracked separately for each kind of job you run. A rising ratio on a stable workload means something changed — a prompt regression, or a task drifting into harder territory. Overhead ratios are also how you calibrate the calculator to your reality instead of our default assumption.

Did anything cross 272K? Tier-two requests are easy to spot in cost-per-request outliers. If they weren’t intentional, cap your retrieval or context assembly below the line.

Are the right tasks on the right model? Sort spend by task type and ask which Ultra jobs base Fugu could have handled — and which base-Fugu jobs were actually routed to top-tier underlying models anyway, making Ultra’s predictable pricing a better fit.

Subscriptions, credits, or both: a decision path

If you’re one person exploring: start with the $20 Standard plan; it includes both models and converts an unpredictable per-token bill into a flat fee while you learn your usage. If you’re building something that ships: pay-as-you-go from the start — production traffic gets scheduling priority, and you’ll want the usage telemetry against real credit spend anyway. If you’re a team doing both, run both: plans for humans, credits for pipelines. The two modes are not mutually exclusive, and the priority rule means your production traffic never queues behind your teammates’ chat sessions.

Whichever mode you choose, re-verify the numbers on this page against Sakana’s official pricing page before committing a budget — prices are current as of August 2026 and can change without our noticing on the same day.

Seven rules of thumb

1. Budget from total billed tokens, not visible tokens — orchestration usage is real money.
2. Engineer for cache hits: stable prompt prefixes cut input costs by 90%.
3. Watch the 272K line — crossing it doubles input rates mid-workload.
4. Use base Fugu by default and reserve Ultra for problems that earn it.
5. Light personal use? Start with a $20 subscription, not credits.
6. Buy credits in portions you’ll spend within six months.
7. Date-check every price — this page says “as of August 2026” for a reason.

New to Fugu entirely? Start with the pillar guide: What is Sakana Fugu?