Cost methodology
Sakana Fugu pricing explained
Ultra and Max can be estimated from published fixed rates. Fugu cannot be reduced to a made-up average, and Cyber is now quote-only. The useful answer is to calculate what is knowable, expose what is variable, and reconcile estimates against the complete usage object.
1. Start with the published rate table
| Model and context | Input | Output | Cached input |
|---|---|---|---|
| Fugu Ultra (v2.0), up to 272K context | $5.00 | $30.00 | $0.50 |
| Fugu Ultra (v2.0), over 272K context | $10.00 | $45.00 | $1.00 |
| Fugu Max (v1.0), any context length | $2.00 | $6.00 | $0.25 |
| Fugu Cyber | Not published on September 28, 2026. Official pages say to contact sales after access is approved. | ||
Fugu Max has no long-context surcharge, so the published 272K threshold applies to Ultra. Standard Fugu remains dependent on its underlying-agent rates; no blanket 272K rule is published for it. Fugu Cyber is available only through pay-as-you-go billing and requires approved access, and this site no longer publishes a Cyber token table because the live pricing page withdrew it. Ultra and Max both remain available in pay-as-you-go and the published subscription tiers. The 272K threshold applies to request context, not to your monthly total.
A month containing many short requests can total millions of tokens while every individual request remains below the threshold.
2. Include orchestration tokens, and subtract each cached subset
Sakana AI documents additional token fields for Ultra and Cyber. Orchestration input, cached input, and output are real usage and are billed at the same rate as their category. The ordinary cached field behaves differently: as with OpenAI, cached_tokens is a subset of input_tokens, and orchestration_input_cached_tokens is a subset of orchestration_input_tokens. Each cached count is therefore subtracted from its own input total before that input is charged the full rate, not added on top of it.
Uncached input = (user input − cached user input) + (orchestration input − cached orchestration input)
Total cached input = cached user input + cached orchestration input
Total output = final output + orchestration output
Estimated cost = uncached input / 1M x input rate + total cached input / 1M x cached rate + total output / 1M x output rate
Validate each pair before pricing: a cached count larger than its own input total is a parse error, not a discount. Store the raw values and the rate version used for calculation. That is more useful than storing only one currency number because it lets you recalculate after a pricing change or fix an interpretation error.
3. Worked examples
| Scenario | Raw usage fields (cached is a subset of input) | Estimate | What it teaches |
|---|---|---|---|
| Small Ultra month | 1M input (250K cached), 250K output | $11.38 | A light fixed-rate workload can sit below the $20 entry subscription price |
| Same visible work with orchestration | 1M input (250K cached), 2M orchestration input (500K cached), 250K output, 500K orchestration output | $34.13 | Ignoring orchestration would understate this example by $22.75 |
| Ultra, output-heavy | 2M input, 2M output, no cache | $70.00 | Output dominates because its unit rate is six times standard input |
| Ultra, long context | 3M input (1M cached), 500K output | $43.50 | Higher context rates must be applied only to qualifying requests |
| Max, standard context | 1M input (250K cached), 250K output | $3.06 | The same mix that costs $11.38 on Ultra costs about $3.06 at flat Max rates |
| Max, long context | 3M input (1M cached), 500K output | $7.25 | Max has no long-context surcharge, so the Ultra long-context example costs about 6 times more |
These examples are arithmetic demonstrations, not usage forecasts. Multi-agent token volume depends on the task, route, tool activity, and retry behavior.
Enter your own usage in the calculator4. Why this site does not publish a fake fixed Fugu rate
The official pricing rule for Fugu is route dependent:
- When one agent is active, you pay that underlying model's standard rate.
- When multiple agents are active, fees do not stack; billing uses a single rate based on the highest tier model involved.
Without knowing the chosen agents and their applicable rate, one number would create false precision. The practical budget method is to run a controlled sample, export actual Fugu usage and charges, then calculate a distribution such as median, 90th percentile, and worst observed cost per task type.
A single blended “cost per request” hides whether a cheap routine route and an expensive deep route are being mixed. Track cost by named task class so a changing workload does not silently break the budget.
5. Subscription prices are clear, exact token equivalence is not
| Plan | Price | Published positioning | Published usage relationship |
|---|---|---|---|
| Standard | $20/month | Light daily use, occasional API calls, small experiments | Baseline allowance |
| Pro | $100/month | Regular coding, review, research, and analysis sessions | 10x Standard usage |
| Max | $200/month | Heavy, long-running workloads | 20x Standard usage |
Every published subscription tier includes Fugu, Fugu Ultra, and Fugu Max — not Fugu Cyber. These are Sakana AI API Platform subscriptions; Sakana Chat is a separate, free product with its own usage limits, and an API subscription does not change them. The public page does not provide one exact token allowance that maps cleanly across all orchestration routes, so a calculator cannot honestly declare a universal subscription break-even point.
A narrow statement is still possible: if your fixed-rate pay-as-you-go estimate is below $20, it is below the entry subscription price. If it is above $20, you still need the console's current allowance and service-priority terms to decide whether the subscription covers your workload. Pay-as-you-go tokens are also served at higher priority than monthly-plan tokens, which matters for latency-sensitive production traffic.
6. Reconcile estimates against actual usage
- Save the raw response usage object with a request correlation ID.
- Separate standard-context and long-context requests before aggregating.
- Subtract each cached count from its own input total, then add user and orchestration token fields by category.
- Apply the published model and context rates for the request date.
- Compare the estimate with console charges and investigate differences rather than hiding them.
If external pricing data cannot be fetched, keep the last verified rate with its date, label the estimate as potentially stale, and link users to the official console. Silent fallback to invented values is worse than a visible limitation.
7. Common pricing mistakes
Counting only final input and output
This omits orchestration fields and can make a complex run look much cheaper than it was.
Adding cached tokens on top of input
A cached field is a subset of its own input total, so adding it again charges those tokens twice and overstates the bill. Subtract it before applying the full input rate.
Treating monthly tokens as request context
The 272K threshold concerns context length per request, not a monthly sum.
Inventing a Fugu fixed rate
Its billing depends on the routed model or highest participating tier. Use actual samples.
Calling the cheapest subscription the winner
Price alone does not show whether its allowance, priority, or workload fit is sufficient.
Forgetting retries
A failed or disconnected client can trigger additional expensive work if retry behavior is not bounded and observable.
Assuming Max inherits Ultra's long-context rate
Fugu Max is flat at any context length, so a budget model that applies Ultra's above-272K surcharge to Max overstates the bill.
Primary sources
- Sakana AI Pricing: rates, Fugu billing rules, subscription tiers, and Fugu Max tool fees
- Sakana AI Models: usage field definitions and model availability
- Sakana AI Fugu product page: product-level pricing and FAQ
Rates are time sensitive. Recheck the official console before purchase or production deployment.