Language modelActive
Fugu Ultra v2
Sakana AI
- Released
- September 11, 2026
- Data date
- October 7, 2026
- Input
- Text, image
- Output
- Text
Fugu Ultra v2 focuses Sakana Fugu on complex reasoning, autonomous research, and software development. Rather than using one checkpoint, the system orchestrates a swappable pool of open and specialist models.
Sakana has offered the version through its OpenAI-compatible API since September 11, 2026. Its API model ID is fugu-ultra-v2.0.
Specifications and access
| Specification | Value and source |
|---|---|
| API model ID | fugu-ultra-v2.0Source |
| System type | Multi-agent orchestration for reasoning, autonomous research, and software development; not one monolithic modelSource |
| Input modalities | Text, imageSource |
| Output modalities | TextSource |
| Model pool | Sakana describes a swappable pool of open and specialist models; according to the provider, Fugu Ultra v2 does not use Fable 5, Fable 5.1, or GPT-6-AstraSource |
| Availability | Available through the Sakana Fugu API since September 11, 2026Source |
Page 1 of 2
Measurements without matching peer values
These measurements have no matching peer values under the same test conditions. Their original values and sources remain available here.
Sakana Fugu Ultra v2.0: HLE (text) (Show measurement, test conditions, and source)
- Source value
- 56.3
- Score
- 56.3
- Metric
- Vendor-reported orchestration-system benchmark score; chart does not specify a unit
- Unit
- No unit provided
- Category
- Sakana AI Fugu Ultra v2.0 vendor chart
- Direction
- Higher is better
- Source type
- vendor-reported
- Evaluator
- Sakana AI, Fugu Ultra v2.0 benchmark chart
- Status
- active
- Retrieved at
- 2026-10-07
- Methodology
- Vendor-reported result. System=Fugu Ultra v2.0; this measures the vendor's coordinated orchestration system; the chart does not publish a unit, exact harness configuration, or isolated base-model result.
- Context
- This is a vendor-reported result for the end-to-end Fugu orchestration system, not a score for an isolated underlying checkpoint. Sakana's chart does not state the score unit, exact benchmark protocol, model-pool composition for this run, or per-run uncertainty. Treat the value as the vendor's displayed system score only.
Sakana Fugu Ultra v2.0: GDP.pdf (Show measurement, test conditions, and source)
- Source value
- 34.3
- Score
- 34.3
- Metric
- Vendor-reported orchestration-system benchmark score; chart does not specify a unit
- Unit
- No unit provided
- Category
- Sakana AI Fugu Ultra v2.0 vendor chart
- Direction
- Higher is better
- Source type
- vendor-reported
- Evaluator
- Sakana AI, Fugu Ultra v2.0 benchmark chart
- Status
- active
- Retrieved at
- 2026-10-07
- Methodology
- Vendor-reported result. System=Fugu Ultra v2.0; this measures the vendor's coordinated orchestration system; the chart does not publish a unit, exact harness configuration, or isolated base-model result.
- Context
- This is a vendor-reported result for the end-to-end Fugu orchestration system, not a score for an isolated underlying checkpoint. Sakana's chart does not state the score unit, exact benchmark protocol, model-pool composition for this run, or per-run uncertainty. Treat the value as the vendor's displayed system score only.
Sakana Fugu Ultra v2.0: Chartography (Show measurement, test conditions, and source)
- Source value
- 48.3
- Score
- 48.3
- Metric
- Vendor-reported orchestration-system benchmark score; chart does not specify a unit
- Unit
- No unit provided
- Category
- Sakana AI Fugu Ultra v2.0 vendor chart
- Direction
- Higher is better
- Source type
- vendor-reported
- Evaluator
- Sakana AI, Fugu Ultra v2.0 benchmark chart
- Status
- active
- Retrieved at
- 2026-10-07
- Methodology
- Vendor-reported result. System=Fugu Ultra v2.0; this measures the vendor's coordinated orchestration system; the chart does not publish a unit, exact harness configuration, or isolated base-model result.
- Context
- This is a vendor-reported result for the end-to-end Fugu orchestration system, not a score for an isolated underlying checkpoint. Sakana's chart does not state the score unit, exact benchmark protocol, model-pool composition for this run, or per-run uncertainty. Treat the value as the vendor's displayed system score only.
Sakana Fugu Ultra v2.0: SWEFish (Show measurement, test conditions, and source)
- Source value
- 71.8
- Score
- 71.8
- Metric
- Vendor-reported orchestration-system benchmark score; chart does not specify a unit
- Unit
- No unit provided
- Category
- Sakana AI Fugu Ultra v2.0 vendor chart
- Direction
- Higher is better
- Source type
- vendor-reported
- Evaluator
- Sakana AI, Fugu Ultra v2.0 benchmark chart
- Status
- active
- Retrieved at
- 2026-10-07
- Methodology
- Vendor-reported result. System=Fugu Ultra v2.0; this measures the vendor's coordinated orchestration system; the chart does not publish a unit, exact harness configuration, or isolated base-model result.
- Context
- This is a vendor-reported result for the end-to-end Fugu orchestration system, not a score for an isolated underlying checkpoint. Sakana's chart does not state the score unit, exact benchmark protocol, model-pool composition for this run, or per-run uncertainty. Treat the value as the vendor's displayed system score only.
Sakana Fugu Ultra v2.0: DeepSWE (Show measurement, test conditions, and source)
- Source value
- 74.3
- Score
- 74.3
- Metric
- Vendor-reported orchestration-system benchmark score; chart does not specify a unit
- Unit
- No unit provided
- Category
- Sakana AI Fugu Ultra v2.0 vendor chart
- Direction
- Higher is better
- Source type
- vendor-reported
- Evaluator
- Sakana AI, Fugu Ultra v2.0 benchmark chart
- Status
- active
- Retrieved at
- 2026-10-07
- Methodology
- Vendor-reported result. System=Fugu Ultra v2.0; this measures the vendor's coordinated orchestration system; the chart does not publish a unit, exact harness configuration, or isolated base-model result.
- Context
- This is a vendor-reported result for the end-to-end Fugu orchestration system, not a score for an isolated underlying checkpoint. Sakana's chart does not state the score unit, exact benchmark protocol, model-pool composition for this run, or per-run uncertainty. Treat the value as the vendor's displayed system score only.
Sakana Fugu Ultra v2.0: ProgramBench (Show measurement, test conditions, and source)
- Source value
- 81
- Score
- 81
- Metric
- Vendor-reported orchestration-system benchmark score; chart does not specify a unit
- Unit
- No unit provided
- Category
- Sakana AI Fugu Ultra v2.0 vendor chart
- Direction
- Higher is better
- Source type
- vendor-reported
- Evaluator
- Sakana AI, Fugu Ultra v2.0 benchmark chart
- Status
- active
- Retrieved at
- 2026-10-07
- Methodology
- Vendor-reported result. System=Fugu Ultra v2.0; this measures the vendor's coordinated orchestration system; the chart does not publish a unit, exact harness configuration, or isolated base-model result.
- Context
- This is a vendor-reported result for the end-to-end Fugu orchestration system, not a score for an isolated underlying checkpoint. Sakana's chart does not state the score unit, exact benchmark protocol, model-pool composition for this run, or per-run uncertainty. Treat the value as the vendor's displayed system score only.
Sakana Fugu Ultra v2.0: GPQAD (Show measurement, test conditions, and source)
- Source value
- 95.5
- Score
- 95.5
- Metric
- Vendor-reported orchestration-system benchmark score; chart does not specify a unit
- Unit
- No unit provided
- Category
- Sakana AI Fugu Ultra v2.0 vendor chart
- Direction
- Higher is better
- Source type
- vendor-reported
- Evaluator
- Sakana AI, Fugu Ultra v2.0 benchmark chart
- Status
- active
- Retrieved at
- 2026-10-07
- Methodology
- Vendor-reported result. System=Fugu Ultra v2.0; this measures the vendor's coordinated orchestration system; the chart does not publish a unit, exact harness configuration, or isolated base-model result.
- Context
- This is a vendor-reported result for the end-to-end Fugu orchestration system, not a score for an isolated underlying checkpoint. Sakana's chart does not state the score unit, exact benchmark protocol, model-pool composition for this run, or per-run uncertainty. Treat the value as the vendor's displayed system score only.
Sakana Fugu Ultra v2.0: Toolathon (Show measurement, test conditions, and source)
- Source value
- 80.6
- Score
- 80.6
- Metric
- Vendor-reported orchestration-system benchmark score; chart does not specify a unit
- Unit
- No unit provided
- Category
- Sakana AI Fugu Ultra v2.0 vendor chart
- Direction
- Higher is better
- Source type
- vendor-reported
- Evaluator
- Sakana AI, Fugu Ultra v2.0 benchmark chart
- Status
- active
- Retrieved at
- 2026-10-07
- Methodology
- Vendor-reported result. System=Fugu Ultra v2.0; this measures the vendor's coordinated orchestration system; the chart does not publish a unit, exact harness configuration, or isolated base-model result.
- Context
- This is a vendor-reported result for the end-to-end Fugu orchestration system, not a score for an isolated underlying checkpoint. Sakana's chart does not state the score unit, exact benchmark protocol, model-pool composition for this run, or per-run uncertainty. Treat the value as the vendor's displayed system score only.
OpenRouter GPQA Diamond · 2026-09-15 · 3 epochs (Show measurement, test conditions, and source)
- Source value
- 0.93266
- Score
- 0.93266
- Metric
- task accuracy (correct_answers / total_questions)
- Unit
- ratio
- Benchmark version
- OpenRouter benchmark run 2026-09-15; workflow rbgq3-1789433899-pilot-gpqa.sakana/fugu-ultra-v2.auto-routing
- Category
- OpenRouter model-router evaluation
- Direction
- Higher is better
- Sample
- 594
- Source type
- independent-evaluator
- Evaluator
- OpenRouter Model Router Benchmarks
- Status
- active
- Retrieved at
- 2026-10-07
- Methodology
- Independent evaluation. Exact public route model_permaslug=sakana/fugu-ultra-v2; provider=auto-routing; benchmark_type=gpqa_diamond; timestamp=2026-09-15T01:07:48.052Z (UTC); reasoning_effort=high; epochs=3, recorded_epochs=3; max_tokens and temperature are not reported. OpenRouter recorded total_questions=594 and correct_answers=554.
- Context
- OpenRouter evaluated the routed Fugu service, not a single underlying checkpoint; Sakana does not disclose the internal model pool. total_questions counts question executions across the recorded epochs, so repeated epochs are not distinct items. This is a separate OpenRouter run, not a reproduction of Sakana's private release-chart harness.
OpenRouter τ³-Bench Banking · 2026-09-16 · 2 epochs (Show measurement, test conditions, and source)
- Source value
- 0.345361
- Score
- 0.345361
- Metric
- task accuracy (correct_answers / total_questions)
- Unit
- ratio
- Benchmark version
- OpenRouter benchmark run 2026-09-16; workflow rb-7741ebdf-ext-fugu-ultra-v2-tau3-banking
- Category
- OpenRouter model-router evaluation
- Direction
- Higher is better
- Sample
- 194
- Source type
- independent-evaluator
- Evaluator
- OpenRouter Model Router Benchmarks
- Status
- active
- Retrieved at
- 2026-10-07
- Methodology
- Independent evaluation. Exact public route model_permaslug=sakana/fugu-ultra-v2; provider=auto-routing; benchmark_type=tau3_bench_banking; timestamp=2026-09-16T02:46:05.133Z (UTC); reasoning_effort=medium; epochs=2, recorded_epochs=2; max_tokens and temperature are not reported. OpenRouter recorded total_questions=194 and correct_answers=67.
- Context
- OpenRouter evaluated the routed Fugu service, not a single underlying checkpoint; Sakana does not disclose the internal model pool. total_questions counts question executions across the recorded epochs, so repeated epochs are not distinct items. This is a separate OpenRouter run, not a reproduction of Sakana's private release-chart harness.
OpenRouter MMMU-Pro Vision · 2026-09-18 · 3 epochs (Show measurement, test conditions, and source)
- Source value
- 0.80501
- Score
- 0.80501
- Metric
- task accuracy (correct_answers / total_questions)
- Unit
- ratio
- Benchmark version
- OpenRouter benchmark run 2026-09-18; workflow rb-7741ebdf-ext4-fugu-ultra-v2-mmmu-pro-vision
- Category
- OpenRouter model-router evaluation
- Direction
- Higher is better
- Sample
- 5190
- Source type
- independent-evaluator
- Evaluator
- OpenRouter Model Router Benchmarks
- Status
- active
- Retrieved at
- 2026-10-07
- Methodology
- Independent evaluation. Exact public route model_permaslug=sakana/fugu-ultra-v2; provider=auto-routing; benchmark_type=mmmu_pro_vision; timestamp=2026-09-18T11:55:14.184Z (UTC); reasoning_effort=high; epochs=3, recorded_epochs=3; max_tokens and temperature are not reported. OpenRouter recorded total_questions=5190 and correct_answers=4178.
- Context
- OpenRouter evaluated the routed Fugu service, not a single underlying checkpoint; Sakana does not disclose the internal model pool. total_questions counts question executions across the recorded epochs, so repeated epochs are not distinct items. This is a separate OpenRouter run, not a reproduction of Sakana's private release-chart harness.
OpenRouter τ³-Bench Banking · 2026-09-18 · 1 epoch (Show measurement, test conditions, and source)
- Source value
- 0.350515
- Score
- 0.350515
- Metric
- task accuracy (correct_answers / total_questions)
- Unit
- ratio
- Benchmark version
- OpenRouter benchmark run 2026-09-18; workflow rb-7741ebdf-ext4-fugu-ultra-v2-tau3-banking
- Category
- OpenRouter model-router evaluation
- Direction
- Higher is better
- Sample
- 97
- Source type
- independent-evaluator
- Evaluator
- OpenRouter Model Router Benchmarks
- Status
- active
- Retrieved at
- 2026-10-07
- Methodology
- Independent evaluation. Exact public route model_permaslug=sakana/fugu-ultra-v2; provider=auto-routing; benchmark_type=tau3_bench_banking; timestamp=2026-09-18T11:55:14.190Z (UTC); reasoning_effort=medium; epochs=1, recorded_epochs=1; max_tokens and temperature are not reported. OpenRouter recorded total_questions=97 and correct_answers=34.
- Context
- OpenRouter evaluated the routed Fugu service, not a single underlying checkpoint; Sakana does not disclose the internal model pool. total_questions counts question executions across the recorded epochs, so repeated epochs are not distinct items. This is a separate OpenRouter run, not a reproduction of Sakana's private release-chart harness.
Page 1 of 2
Published prices
- Input price
- $5 per 1M tokens; above 272K context: $10Source
- Output price
- $30 per 1M tokens; above 272K context: $45Source
- Cache-read price
- $0.50 per 1M tokens; above 272K context: $1Source
- Orchestration usage
- Orchestration tokens are billed in addition to visible input and output tokens at the corresponding ratesSource
- Tool price
- web_search and web_fetch: $0.007 per callSource
Sources and data date
| Type | Evidence and data date |
|---|---|
| Input | Documented modalities. Retrieved October 7, 2026. |
| Output | Documented modalities. Retrieved October 7, 2026. |
| Research status | Research date October 7, 2026. 9 source URLs checked. This documents the inspected sources, not an exhaustive inventory of every publication. |
| Not found in the inspected sources | reproduction of Sakana release-chart protocols Original wording OpenRouter independently evaluated the public route sakana/fugu-ultra-v2 on GPQA Diamond, MMMU-Pro Vision, Tau³-Bench Banking and SWE-Atlas QA. These are OpenRouter's own routed-system runs with separately reported accuracy and epochs, not reproductions of Sakana's release-chart harness. The route is separately identified as v2 in OpenRouter's recorded permaslug. Sakana's release chart still does not publish its complete harness. |
| Not published | benchmark score units and full harness configuration Original wording The release chart gives eight numeric system scores but no score unit or full benchmark/harness configuration; do not interpret these values as percentages or isolated base-checkpoint results. |
| Additional source | Sakana Fugu Ultra v2 release (retrieved October 7, 2026) · Sakana Fugu Ultra v2 release · Editorial description reviewed October 7, 2026 |
| Additional source | Sakana Fugu token and tool pricing (retrieved October 7, 2026) |
| Additional source | Sakana AI, Fugu Ultra v2.0 benchmark chart (retrieved October 7, 2026) |
| Additional source | OpenRouter Model Router Benchmarks (retrieved October 7, 2026) |
| Additional source | Model metadata (retrieved October 7, 2026) |
| Additional source | Model metadata (retrieved October 7, 2026) |
| Additional source | Model metadata (retrieved October 7, 2026) |
| Additional source | Model metadata (retrieved October 7, 2026) |