Skip to content
News Product launch

Sakana Fugu Ultra reaches OpenRouter at $5/$30 per million tokens with SWE-Bench Pro benchmark lead over GPT-5.5 and Gemini

· by Pondero Newsdesk

The short version

Sakana AI's Fugu Ultra multi-agent orchestration model is available on OpenRouter at $5 input and $30 output per million tokens, offering a single API endpoint that routes tasks across a swappable LLM pool without single-vendor lock-in.

Sakana Fugu Ultra reaches OpenRouter at $5/$30 per million tokens with SWE-Bench Pro benchmark lead over GPT-5.5 and Gemini

Sakana AI's Fugu Ultra, the multi-agent orchestration model that routes each task across a swappable pool of frontier LLMs, landed on OpenRouter on June 24, 2026, priced at $5 per million input tokens and $30 per million output tokens, per the OpenRouter listing. The availability on the popular third-party routing layer gives developers a drop-in endpoint without direct integration with Sakana's own API.

What

Fugu Ultra is not a single model in the conventional sense. It is a language model trained to classify incoming queries, route them to one or more frontier sub-models from a dynamic pool, and synthesize a single response, all through one OpenAI-compatible API endpoint. Users send a request; the orchestrator decides internally how many models to involve and which to call, per Sakana's release post.

The technical report, published on arXiv on June 19 and revised June 23, arXiv:2606.21228, covers benchmark performance across SWE-Bench Pro, Terminal Bench, LiveCodeBench, GPQA-Diamond, Humanity's Last Exam, and CharXiv Reasoning. Sakana reports Fugu Ultra leads all publicly accessible models across 10 of 11 benchmarks tested, per the technical report. The specific SWE-Bench Pro score of 73.7, compared with 69.2 for Opus 4.8 and 58.6 for GPT-5.5, appears in the technical report's benchmark table but the table was published as an image rather than structured data, so only Sakana's own figures are available.

Analysis by Requesty, which reverse-engineered Fugu's behavior by comparing token counts between the Sakana API and its own router, found the model runs in two modes: a single-model pass for simple queries with roughly 1,260 orchestration input tokens of overhead, and a multi-model synthesis pass for complex queries where orchestration token consumption runs 5x to 12x higher than visible output, per Requesty's breakdown. That overhead is billed at standard input/output rates, not separately.

The $30 per million output rate is materially higher than comparable frontier models billed individually. The cost calculus depends on whether the orchestration yields a quality lift that justifies the markup for a given task type.

Why it matters

The OpenRouter listing makes Fugu Ultra practically accessible to any developer already routing through that platform, no new account or direct API contract required. That distribution path matters for adoption speed, particularly among teams prototyping multi-agent workflows who want to test the orchestration-as-model hypothesis without a procurement step.

The export-control framing from Sakana's launch, citing the June 2026 restrictions on Anthropic's Fable and Mythos models as evidence of single-vendor dependency risk, adds a specific operational reason for non-US organizations to evaluate this architecture. Because the underlying agent pool is swappable, a provider restriction triggers a model substitution rather than a service disruption, per Sakana's positioning.

The pricing structure does create a real evaluation question. At $30 per million output tokens, complex multi-step queries that trigger the full orchestration path cost significantly more than a direct call to GPT-5.5 at its standard rates. Teams considering Fugu Ultra will need to benchmark their specific workload before committing.

What to watch next

Sakana said it plans to add its own in-house models to the agent pool over coming months, which would reduce dependence on third-party providers and potentially change the cost profile. Whether any peer-reviewed, third-party benchmark replication of the SWE-Bench Pro claims appears is the clearest near-term signal on whether the reported 73.7 score holds up outside Sakana's own testing environment.

Sources