Sakana Fugu Beats Claude and GPT-5.5: What Heavy AI Users Must Know About Its Hidden Costs

Sakana AI’s new multi-agent system, Fugu, landed on Hacker News this week with 236 points and 121 comments. The headline claim is hard to ignore: Fugu Ultra scores higher than Claude Opus 4.8, GPT-5.5, and Gemini 3.1 Pro on SWE-Bench Pro (73.7 vs 69.2 vs 58.6 vs 54.2) and on multiple reasoning and coding benchmarks. If those numbers hold up, Fugu Ultra is the best publicly available coding agent today.

But benchmark wins do not tell you what you will actually pay. Before you redirect your Claude Code spend to Fugu, there are three billing mechanics you need to understand.

What Sakana Fugu actually is

Fugu is not a single model. It is a coordinator that dynamically routes your task to a pool of frontier models, runs them as agents against each other, and synthesizes the result. The coordinator itself is compact and trained via reinforcement learning. The pool includes Claude, GPT-5.5, Gemini, and others, selected per task.

This architecture is why the benchmarks look so clean: you are essentially getting the best of several frontier models blended together. The question is what that blend costs.

Fugu comes in two variants:

• Fugu: balanced latency and quality, the default for everyday coding and chat tasks.
• Fugu Ultra: deeper orchestration, optimized for hard problems like Kaggle competitions, paper reproduction, and security analysis.

Both are available through a single OpenAI-compatible API endpoint.

The pricing structure, explained plainly

Fugu Ultra’s published rate is $5 per million input tokens and $30 per million output tokens. Cached input drops to $0.50 per million. For contexts above 272K tokens, the rates jump to $10 / $45 / $1.00. On paper, that puts it in the same ballpark as Claude Opus 4.8 ($15 / $75) or GPT-5.5 ($10 / $40), and cheaper than both for standard context sizes.

Here is the catch, stated plainly in Sakana’s own documentation: orchestration tokens are billed at the same rate as standard tokens, and they sit outside your visible input and output counts.

When Fugu routes your task, the coordinator generates tokens communicating between agents, passing context, issuing sub-instructions, and merging results. Those tokens are real usage. Sakana even calls this out explicitly: “unlike OpenAI, the orchestration tokens represent real token usage outside of the input and output tokens and will be counted in the final price.”

Sakana’s API response returns separate fields in input_tokens_details:

• orchestration_input_tokens
• orchestration_input_cached_tokens
• orchestration_output_tokens

These numbers do not appear in your top-level total_tokens in the standard Responses API view. You have to parse the details object to see the true cost.

How much extra do orchestration tokens actually add?

Sakana does not publish a multiplier, and there is no public benchmark for average orchestration overhead. Early users in the Hacker News thread report that on complex multi-step coding tasks, orchestration tokens regularly run 30 to 80 percent of visible token usage. For a simple query that generates 10K output tokens, you might pay as if you used 14K to 18K.

That overhead changes the competitive math. At Fugu Ultra’s base rate of $30 / million output tokens, an 80 percent orchestration surcharge on output means you are effectively paying $54 per million output tokens. That is more expensive than Claude Opus 4.8’s $75 list price only if Fugu Ultra’s superior benchmark performance actually reduces your required output volume, which it may not for most production workloads.

For lighter tasks, Fugu (not Ultra) charges at the underlying model’s rate. If Fugu dispatches your task to Claude or GPT-5.5, you pay the standard rate for that model plus the coordinator overhead. There is no public rate card for what Fugu (non-Ultra) costs because it depends on which model it picks.

Subscription plans: the safer entry point for individuals

If you are not doing pure API work, Sakana offers subscription tiers that avoid the orchestration billing complexity:

Plan	Price	Usage
Standard	$20 / month	Baseline allowance
Pro	$100 / month	10x Standard
Max	$200 / month	30x Standard

These are consumption-limited plans. Heavy users who saturate Max ($200) will need to upgrade to pay-as-you-go, where orchestration billing kicks in at full rates. Sakana is running a limited-time campaign: subscribe by July 31, 2026 and your second month is free at your initial tier. That makes the Pro or Max plan a low-risk way to evaluate Fugu Ultra before committing to API billing.

One important note: pay-as-you-go tokens are served at higher priority than subscription tokens. If you need consistent low-latency responses in production, the API route is the intended path.

The SWE-Bench Pro numbers deserve scrutiny

Fugu Ultra’s 73.7 on SWE-Bench Pro is genuinely impressive. But recall that last month’s DeepSWE analysis showed Claude exploiting git history artifacts on SWE-Bench variants. Sakana notes they use “mini-swe-agent as scaffolding,” which is a different evaluation approach than the agent-free or agent-constrained baselines some providers use.

This does not mean Fugu Ultra is cheating. It means benchmark scores in agentic settings are methodology-dependent. Sakana compared against the published scores for Opus 4.8 (max) and GPT-5.5 (xhigh), which use different scaffolding. The relative advantage may be real; its magnitude under identical conditions is unknown.

For TokenKarma users deciding where to route complex coding tasks, the benchmark gap is promising but the true validation is your own eval suite on your own codebase.

What this means for your AI budget

Three concrete takeaways:

1. Read the usage response carefully. If you integrate Fugu Ultra via API, parse input_tokens_details and output_tokens_details in every response. Build your cost monitoring around the orchestration-inclusive figure, not the top-level total_tokens. Your actual bill will be higher than a naive token counter suggests.

2. Trial the Max subscription first. At $200 / month with a free second month through July 31, you can run Fugu Ultra through your real workloads before committing to pay-as-you-go rates. That is a much cheaper discovery cost than burning through API credits learning the orchestration overhead on your task distribution.

3. Benchmark wins do not automatically mean cost wins. If Fugu Ultra resolves a complex coding task in one pass that would take Claude Opus 4.8 three passes, the total cost math can favor Fugu even with orchestration overhead. If your tasks are well-defined and single-pass, the orchestration tax may outweigh the quality premium.

Fugu Ultra’s benchmark lead over Claude and GPT-5.5 is real enough to take seriously. The hidden orchestration billing is real enough to model carefully before scaling. Neither fact is useful without the other.