evals.report
BenchmarksLabsCompareRun guidesIn the wild

Fugu Ultra

Sakana AI · Fugu. Released Jun 15, 2026.

Fugu Ultra is a model from Sakana AI in the Fugu family, released Jun 15, 2026. evals.report tracks 7 reported Fugu Ultra benchmark scores across SWE-bench Pro, Terminal-Bench 2.1, Humanity's Last Exam, GPQA Diamond, CharXiv, SciCode, LiveCodeBench — each shown with its benchmark, metric, source status, and date, and never combined into a single ranking.

7 results

Benchmark results 7

Compare this model
BenchmarkCategoryScoreMetricStatusDate
SWE-bench ProCoding73.7%% resolvedVerifiedJun 15, 2026Details
Terminal-Bench 2.1Agents82.1%task successVerifiedJun 15, 2026Details
Humanity's Last ExamReasoning50.0%accuracyVerifiedJun 15, 2026Details
GPQA DiamondReasoning95.5%accuracyVerifiedJun 15, 2026Details
CharXivMultimodal86.6%accuracyVerifiedJun 15, 2026Details
SciCodeCoding58.7%accuracyVerifiedJun 15, 2026Details
LiveCodeBenchCoding93.2%Pass@1VerifiedJun 15, 2026Details

In the wild 7

See all

Real-world feedback on Fugu Ultra from people using it on actual prompts — praise and criticism alike, each linked to its source. Qualitative, never scored.

Report tone

Report type

Topic

AIAcademy
X·@AIAcademykorea·
Mixed
The 5-hour limit has been exceeded, so I have to wait 4 hours. However, it kindly provides guidance … I like this one better because it is user-oriented, offering friendly guidance for beginners and general users.
anecdotal
AM9:21
X·@AM921543266·
Positive
It discovered 27 bugs that Fable 5 couldn't find and fixed all of them. The code quality is impeccable … it implemented about 70,000 lines of new features, resolved 4 issues, and created 7 PRs.

Task Introduced Fugu into a repo previously worked on with Claude Fable 5; ~1 hour of use.

anecdotaloutput shown
Ethan Mollick
X·@emollick·
Negative
it is incredibly slow: my typical coding tests (shaders, interactive scenes) take 30 minutes to run. And the results are... fine. It does not match Fable in real use.

Task Coding tests (shaders, interactive scenes) on Fugu Ultra-high; each ~30 min.

high-signal useroutput shown
Harshith
X·@HarshithLucky3·
Positive
My First Test with Sakana Fugu Ultra … SVG of Xbox Controller.

Task First test — generating an SVG of an Xbox controller.

anecdotaloutput shown
am.will
X·@LLMJunky·
Negative
The game was pretty bad and notably worse than GPT 5.5. … GPT 5.5 by contrast did a pretty good job and required no follow ups.

Task Asked it to build a Three.js replica of Rocket League via Codex.

anecdotalprompt shownsingle run
Field testQuotaComparisonCodingAgents
View on X
Mark Santos
X·@markksantos·
Mixed
In terms of model speed and performance, Fugu on Opencode won … inverted directional turn, wonky camera, no sfx, not identical to Crossy Road game.

Task Head-to-head vs Claude Opus 4.8: a single-file Three.js Crossy Road game.

prompt shownsingle run
elvis
X·@omarsar0·
Positive
OMG! Fugu Ultra is ridiculously good at these 3D renders.

Task Generating 3D renders.

high-signal useroutput shown