Verdict: Sakana AI's Fugu-Ultra v1.1 — released July 24, 2026 — is a multi-agent orchestration model that claims to match or beat Anthropic's Claude Fable 5 on several coding and reasoning benchmarks, even though Fable 5 is not in Fugu's underlying model pool. The numbers are plausible and the architecture is genuinely novel, but every score is vendor-reported with no independent verification yet. For teams that need global availability and vendor diversification, Fugu is the strategic pick; for raw single-model capability where you can access it, Fable 5 still leads on its own terms.
Last verified: 2026-07-30 · Sakana shipped Fugu-Ultra v1.1 on July 24, 2026 · Pricing unchanged at $5/M input, $30/M output · All benchmark figures are vendor-reported · Fable 5 is NOT in Fugu's model pool
What is Sakana Fugu-Ultra v1.1?
Fugu-Ultra v1.1 is the latest refresh of Sakana AI's quality-first multi-agent orchestration model, delivered through a single OpenAI-compatible API endpoint. Instead of answering your query with one large language model, Fugu routes it across a pool of frontier models — including Claude Opus 4.8, GPT-5.5, and Gemini 3.1 Pro — coordinates their work, runs verification passes, and synthesizes one clean answer. The July 24, 2026 update incorporates newer frontier models into that pool and reports gains of up to 7.9 benchmark points over the v1.0 release from June 2026 [Sakana AI, fugu-1-1-claude-code-interface].
The name comes from "sakana" (魚), the Japanese word for fish — the design metaphor is a school of fish moving as one collective entity. Sakana calls this approach "collective intelligence": a team of good but individually smaller models, working together, can outperform a single giant model on hard tasks. The company, founded in Tokyo by former Google Brain researchers David Ha and Llion Jones, frames this as a new scaling axis beyond simply training ever-larger monolithic models [Sakana Fugu Technical Report, arXiv:2606.21228].
The Fugu family now includes three models behind the same API:
- Fugu — optimized for low-latency, everyday workloads like coding assistance and chat.
- Fugu Ultra — the quality-first tier for hard, multi-step problems like deep research, paper reproduction, and cybersecurity analysis (the subject of this article).
- Fugu Cyber — specialized for security reasoning tasks like vulnerability research and threat investigation [Sakana AI Fugu product page].
How does Fugu beat Fable 5 without using Fable 5?
A coordinated team of multiple public frontier models can match or exceed a single restricted frontier model on benchmarks, because orchestration adds a verification and synthesis layer that no single model applies to itself. Fugu's trained 7B-parameter coordinator model decides which workers to assign to which sub-task, runs recursive self-correction passes, and synthesizes results — so errors one model makes can be caught by another before the answer reaches you.
This is the central and counterintuitive claim of the v1.1 release. Claude Fable 5 is Anthropic's most capable generally-available model, launched June 9, 2026, and it is explicitly NOT in Fugu's agent pool because Fable 5 is not publicly accessible to third-party orchestrators — it was placed under US export controls on June 12, 2026, restricting availability in multiple regions [Anthropic, Fable/Mythos access update]. So when Sakana reports Fugu-Ultra beating Fable 5 on benchmarks, they are comparing a team of public models against a rival system that none of them can individually call.
Sakana's CEO David Ha (who posts online as "hardmaru") stated publicly that the team "beat Fable 5 without even having Fable 5 in the pool" — attributing the result to collective intelligence rather than any single model's capability [Sakana AI X announcement].
The honest caveat: every number comes from Sakana's own testing. No independent third-party has reproduced the v1.1 results yet, and the baseline scores for Fable 5 and other models in Sakana's comparison table are provider-reported (not re-run by Sakana). Treat the figures as a directional signal, not verified fact [Coursiv, Fugu-Ultra v1.1 analysis].
What benchmarks did Fugu-Ultra v1.1 improve on?
Sakana reports the strongest v1.1 gains on two coding and agentic benchmarks, with coding and terminal tasks moving the most across the suite. The "up to 7.9 points" figure is a ceiling — the maximum delta on any single benchmark — not the average improvement.
| Benchmark | Fugu-Ultra v1.0 | Fugu-Ultra v1.1 (claimed) | Fable 5 (provider-reported) | What it tests |
|---|---|---|---|---|
| ProgramBench | 72.4% | 79.8% | Not published | Real-world programming |
| Terminal Bench 2.1 | 82.1% (~90.0% claimed v1.1) | ~90.0% | 84.6% | Running real terminal tasks |
| GPQA-D | 95.5% | 95.6% | 92.6% | Graduate-level science reasoning |
| SWE-Bench Pro | 73.7% | Not separately published | 80.3% | Software engineering on real repos |
| LiveCodeBench | 93.2% | Not separately published | 89.8% | Live coding problem-solving |
Sources: Sakana AI published benchmark table; Fable 5 figures from Anthropic. v1.1 specific deltas are vendor-claimed and not independently verified. [Sakana AI Fugu product page; Anthropic Fable 5; AIToolsReview, Fugu-Ultra v1.1]
Two important reading notes:
- "Up to 7.9 points" is a maximum, not an average. Sakana has not published a per-benchmark v1.0-vs-v1.1 score table, so the gain distribution is marketing-shaped — ProgramBench and Terminal Bench 2.1 are the categories Sakana chose to highlight.
- SWE-Bench Pro tells a more nuanced story. On this widely-watched software engineering benchmark, Fable 5's own 80.3% score (per Anthropic) still tops Fugu-Ultra's 73.7% v1.0 figure. Sakana has not published a v1.1 SWE-Bench Pro number at time of writing, so the gap may have narrowed — but it has not been demonstrated to have closed on this specific test [Coursiv; Anthropic Fable 5 benchmarks].
Fugu-Ultra v1.1 vs Fable 5: how do they actually compare?
The real comparison is not "which model is smarter" — it is "which approach fits your situation, given who can access what." Here is the honest side-by-side.
| Dimension | Fugu-Ultra v1.1 | Claude Fable 5 |
|---|---|---|
| Architecture | Multi-agent orchestrator (7B coordinator + frontier model pool) | Single Mythos-class frontier model |
| Release date | July 24, 2026 (v1.1 refresh) | June 9, 2026 |
| Input price | $5 / 1M tokens | $10 / 1M tokens |
| Output price | $30 / 1M tokens | $50 / 1M tokens |
| Context window | 1M tokens (price doubles above 272K) | 1M+ tokens |
| Global availability | Generally available globally (not yet in EU/EEA) | Restricted by US export controls since June 12, 2026 |
| Vendor lock-in | None — swappable model pool, can exclude any provider | Single vendor (Anthropic) |
| Latency | Slower (multi-model orchestration overhead) | Faster (single model) |
| Token cost per task | Higher (orchestration consumes extra tokens) | Lower per query, but higher sticker price |
| Data retention | Not disclosed as covered-model retention | 30-day mandatory retention (Covered Model) |
| Best for | Hard multi-step work where quality > speed; vendor diversification | Highest single-model capability where you can access it |
| Evidence level | Vendor-reported benchmarks only | Provider-reported + broader industry scrutiny |
Sources: Sakana AI pricing page; Anthropic Fable 5 pricing; Anthropic Fable/Mythos access notice. [Sakana console pricing; Anthropic Fable 5; Anthropic access update]
The key insight: Fugu-Ultra is half the per-token sticker price of Fable 5 ($5/$30 vs $10/$50), but the effective cost per task is closer than the rates suggest because orchestration consumes additional tokens for verification and synthesis. You are trading raw speed and token efficiency for resilience, vendor diversification, and (Sakana claims) higher quality on multi-step work [Felloai, Fugu-Ultra v1.1].
Who should actually use Fugu-Ultra v1.1?
Fugu-Ultra is built for hard, high-stakes, multi-step work where getting the right answer matters more than getting a fast one. It is NOT the tool for "write me a quick email" or simple chat — the orchestration overhead makes it slower and more token-expensive for trivial tasks.
Use Fugu-Ultra v1.1 if:
- You need deep research, scientific paper analysis, or cybersecurity reasoning where verification passes materially improve quality.
- Your organization needs vendor diversification or a hedge against export-control risk (Fable 5 is restricted in multiple regions).
- You are doing Kaggle-style competitions, literature reviews, or patent investigations where multiple specialist perspectives help.
- You serve regions where Fable 5 is not available.
Use Fable 5 instead if:
- You have access to it and your workload tolerates the 30-day data retention requirement.
- You need the highest single-model capability on software engineering (Fable 5 leads on SWE-Bench Pro at 80.3%).
- Latency matters more than orchestration resilience.
- You are doing everyday coding, agents, or chat where Opus 4.8 or the Opus tier is the better cost-to-capability call [Anthropic Fable 5].
Use a single cheaper model (Opus 4.8, Sonnet, GPT-5.5) if:
- Your tasks are well-defined and do not benefit from multi-model verification.
- You need minimum latency and predictable per-token cost.
What does real-world reception say about Fugu?
The original Fugu launch in June 2026 received a lukewarm reception — independent coverage noted high token usage, slower response times, and output quality that did not consistently match the benchmark claims. Orchestrating multiple models is, almost by definition, more expensive in tokens and latency than calling one model directly. That structural trade-off showed up in real usage on v1.0, and it has not fundamentally changed in v1.1 — the refresh improves the model pool, not the orchestration cost [AIToolsReview].
Roughly 75% of community reactions to v1.1 were positive, with the remaining 25% citing slowness and cost. Both criticisms are structurally inherent to multi-agent orchestration rather than bugs to be fixed — running a team costs more than running one worker. The speed objection matters less when you are running Fugu as a background agent (you hand it a job and walk away), and more when you are staring at the screen waiting for a single answer.
How to test Fugu-Ultra v1.1 before committing
Because the v1.1 numbers are vendor-reported, the most responsible move is to run your own evaluation before switching production traffic. Here is a practical re-test checklist:
- Pick 20-50 representative tasks from your actual workload — not synthetic benchmarks.
- Separate a coding/terminal subset to probe the areas Sakana claims improved most.
- Run v1.0 and v1.1 side-by-side on the same tasks (the API supports both model IDs).
- Log per-task cost: input tokens, output tokens, cached tokens, and any orchestration overhead.
- Measure P50 and P95 latency end-to-end, not just time-to-first-token.
- Score answer quality with human review or an agent-as-a-judge setup — see our guide to agent-as-a-judge evaluation for the methodology.
- Track retry rate — how often does v1.1 need to re-run a step to get it right?
- Compare against your current model (Fable 5, Opus 5, or whatever you run today) on the same axes.
This is the same evaluation discipline we recommend for any model change in our long-horizon agent evaluation guide — treat the vendor's benchmark table as a hypothesis, not a conclusion.
What this means for you
For developers and builders: The real lesson from Fugu-Ultra is architectural, not benchmark-specific. Tools change every week — a new "best" model drops constantly. If you build your entire workflow around one model, you are stuck when the next one comes or when access restrictions hit. The smarter move is to build (or adopt) a system that any model can plug into, so a new tool dropping is a configuration change, not a rebuild. Fugu itself is one realization of that idea, but the pattern matters more than any single product. If you are setting up an agent workflow for the first time, our 6-step productivity workflow guide walks through the same plug-and-play principle at a personal scale.
For small businesses: You probably do not need Fugu-Ultra for most day-to-day work — a single good model like Opus 4.8 or Sonnet will be faster and cheaper for emails, content, and customer support. Reach for orchestration-grade tools when you have a genuinely hard, multi-step problem (research-heavy, verification-sensitive) where a wrong answer is expensive. For choosing between models by price and capability, see our Claude Opus 5 vs Fable 5 price comparison and the Claude Opus 5 benchmarks and pricing developer guide.
For teams with compliance or access constraints: Fugu's swappable model pool is genuinely valuable if you operate in regions where Fable 5 is restricted, or if you need to exclude specific providers for data-privacy or compliance reasons without retraining. You can configure the pool to favor particular providers, exclude specific models, or respect data constraints — that is a real architectural advantage over single-vendor dependency [Sakana Fugu Technical Report].
FAQ
Q: Is Fugu-Ultra v1.1 actually better than Claude Fable 5? A: On several of Sakana's own published benchmarks, Fugu-Ultra v1.1 reports higher scores than Fable 5 — including Terminal Bench 2.1 (~90% vs 84.6%) and GPQA-D (95.6% vs 92.6%). However, every number is vendor-reported with no independent verification, and on SWE-Bench Pro, Fable 5's own 80.3% still tops Fugu-Ultra's published 73.7%. Fable 5 is also not in Fugu's model pool, so the comparison is a coordinated team versus a single rival system. Treat the claims as directional until third-party benchmarks land.
Q: How much does Fugu-Ultra v1.1 cost? A: On the token plan, Fugu-Ultra costs $5 per million input tokens and $30 per million output tokens, with cached input at $0.50. Prices double above a 272K-token context window ($10/$45/$1.00). Subscriptions run $20/month (Standard), $100/month (Pro), and $200/month (Max). Pricing is unchanged from v1.0 [Sakana console pricing].
Q: Can Fugu-Ultra use Claude Fable 5 as one of its models? A: No. Fable 5 is not publicly accessible to third-party orchestrators due to US export controls placed June 12, 2026, so Fugu cannot route to it. Fugu's agent pool includes Claude Opus 4.8, GPT-5.5, Gemini 3.1 Pro, and other publicly available models. When Sakana claims Fugu matches Fable 5, they mean a coordinated team of other public models rivals Anthropic's restricted frontier model [Anthropic access update].
Q: Is Fugu-Ultra v1.1 available in the EU? A: No. Sakana states that Fugu is not yet available in the EU/EEA while it works toward GDPR and EU-specific regulatory compliance. This restriction applies to v1.1 the same way it applied at the June 2026 launch [Coursiv].
Q: What is the difference between Fugu, Fugu Ultra, and Fugu Cyber? A: Fugu is optimized for low-latency everyday workloads. Fugu Ultra is the quality-first tier for hard, multi-step problems where answer quality matters more than speed. Fugu Cyber is specialized for cybersecurity reasoning — vulnerability research, threat investigation, and security analysis. All three are behind the same OpenAI-compatible API [Sakana AI Fugu product page].
Q: Why is Fugu slower than a single model? A: Because orchestration involves multiple models working in sequence or parallel, running verification passes, and synthesizing results. A single model answers immediately; a coordinated team has back-and-forth overhead. This is a structural trade-off — you are exchanging speed for verification depth and vendor diversification. The speed cost matters less when Fugu runs as a background agent you hand work to and return to later.
Q: Should I trust Sakana's benchmark numbers? A: Treat them as a strong directional signal, not verified fact. All v1.1 figures are vendor-published, the baseline scores for rival models are provider-reported (Sakana does not re-run them), and no independent third-party has reproduced the v1.1 results yet. The most responsible approach is to run your own 20-50 task evaluation comparing v1.0, v1.1, and your current model on answer quality, cost-per-task, and latency before switching production traffic.
Discussion
0 comments