Pick Fugu Ultra v2.0 for sakana's own evaluation reports best-or-joint-best results on 5 of 8 benchmarks tested, including a 74.3 on the deepswe coding-repair benchmark or claims to outperform flagship models like gpt-6 astra and claude fable 5.1 despite its underlying orchestration pool reportedly excluding both of them. Pick Kimi K3 for largest open-weight model at release — 2.8t sparse moe, self-hostable or 1m-token context with native vision (text, image and video). Choose Kimi K3 if you need self-hosting or data privacy; Fugu Ultra v2.0 if you want a managed API.
Fugu Ultra v2.0 (Sakana AI) and Kimi K3 (Moonshot AI) are two of the models people most often weigh against each other in 2026. Fugu Ultra v2.0 is sakana AI's flagship multi-agent orchestration model, claiming best-or-joint-best results on 5 of 8 self-reported benchmarks against monolithic frontier models — $5/$30 per million tokens. Kimi K3 is moonshot's 2.8-trillion-parameter open-weight model — the largest open AI at release, with a 1M context and strong self-reported coding scores. They diverge most on price, context window and open vs. closed weights — each quantified below from the models' real specs.
Key differences
Price: Kimi K3 is about 1.7× cheaper on input ($3/$15 per 1M tokens vs $5/$30 per 1M tokens) — modest, but it adds up at steady volume.
Context window: 1M tokens vs 1M — within a few percent of each other, so treat this as a tie and test on your own long inputs, since usable recall varies by model.
Recency: Fugu Ultra v2.0 is the newer model by about 45 days (released September 10, 2026), usually meaning fresher training data and capabilities.
Specifications
Spec
Fugu Ultra v2.0
Kimi K3
Provider
Sakana AI (Global)
Moonshot AI (China)
Released
September 10, 2026
July 27, 2026
Context window
1M tokens (~1,500 pages)
1M (~1,573 pages)
Price (in/out)
$5/$30 per 1M tokens
$3/$15 per 1M tokens
Open weight?
No — API only
Yes — self-hostable
Modalities
text
text, image, video, code
SWE-Bench Verified
Not published
Not published
MRCR v2 @ 1M
Not published
Not published
Who wins what
Sakana's own evaluation reports best-or-joint-best results on 5 of 8 benchmarks tested, including a 74.3 on the DeepSWE coding-repair benchmark: Fugu Ultra v2.0 — Kimi K3 is comparatively weak here — coding scores use Moonshot FrontierSWE, not standard SWE-Bench Verified
Claims to outperform flagship models like GPT-6 Astra and Claude Fable 5.1 despite its underlying orchestration pool reportedly excluding both of them: Fugu Ultra v2.0 — Sakana AI's flagship multi-agent orchestration model, claiming best-or-joint-best results on 5 of 8 self-reported benchmarks against monolithic frontier models — $5/$30 per million tokens — and it is the newer of the two.
1M-token context window: Fugu Ultra v2.0 — Fugu Ultra v2.0 lists 1M-token context window among its strengths; Kimi K3 does not.
Largest open-weight model at release — 2.8T sparse MoE, self-hostable: Kimi K3 — Open weights make this possible at all — Fugu Ultra v2.0 is API-only, so it cannot leave the vendor's servers.
1M-token context with native vision (text, image and video): Kimi K3 — Fugu Ultra v2.0 is comparatively weak here — pricing rises to roughly $10/$45 per million tokens for prompts above 272K tokens
Vendor reports 81.2 FrontierSWE and 88.3 Terminal-Bench 2.1 on its own harness: Kimi K3 — Fugu Ultra v2.0 is comparatively weak here — not a single trained model — Sakana describes Fugu as "a Multi-Agent System, Delivered as One Model" that dynamically routes tasks across a pool of other models, so head-to-head benchmark framing against monolithic frontier models should be read skeptically
Lowest cost at scale: Kimi K3 — At $3/$15 per 1M tokens, it is the cheaper of the two — the gap dominates the bill on high-volume workloads.
Which should you pick?
A cost-sensitive startup shipping high volume: Kimi K3 — At $3/$15 per 1M tokens it undercuts Fugu Ultra v2.0, and on millions of tokens that margin decides the monthly bill.
Someone analysing very long documents or codebases: Kimi K3 — Larger 1M window fits more in one prompt.
A team with data-privacy or self-hosting needs: Kimi K3 — Open weights let you run it on your own hardware; Fugu Ultra v2.0 is API-only.
Anyone whose priority is sakana's own evaluation reports best-or-joint-best results on 5 of 8 benchmarks tested, including a 74.3 on the deepswe coding-repair benchmark: Fugu Ultra v2.0 — It is specifically built for that.
Anyone whose priority is largest open-weight model at release — 2.8t sparse moe, self-hostable: Kimi K3 — That is its strongest area.
Fugu Ultra v2.0: where it fits
Sakana AI's flagship multi-agent orchestration model, claiming best-or-joint-best results on 5 of 8 self-reported benchmarks against monolithic frontier models — $5/$30 per million tokens. Released September 10, 2026 by Sakana AI, it is built for sakana's own evaluation reports best-or-joint-best results on 5 of 8 benchmarks tested, including a 74.3 on the DeepSWE coding-repair benchmark, claims to outperform flagship models like GPT-6 Astra and Claude Fable 5.1 despite its underlying orchestration pool reportedly excluding both of them, and 1M-token context window.
Its trade-offs are real: not a single trained model — Sakana describes Fugu as "a Multi-Agent System, Delivered as One Model" that dynamically routes tasks across a pool of other models, so head-to-head benchmark framing against monolithic frontier models should be read skeptically, pricing rises to roughly $10/$45 per million tokens for prompts above 272K tokens, and benchmark comparisons against GPT-6 Astra and Claude Fable 5.1 are Sakana's own self-reported figures, not independently verified head-to-head scores. At $5 in / $30 out per million tokens, it sits in the premium price band.
Kimi K3: where it fits
Moonshot's 2.8-trillion-parameter open-weight model — the largest open AI at release, with a 1M context and strong self-reported coding scores. Released July 27, 2026 by Moonshot AI, it is built for largest open-weight model at release — 2.8T sparse MoE, self-hostable, 1M-token context with native vision (text, image and video), vendor reports 81.2 FrontierSWE and 88.3 Terminal-Bench 2.1 on its own harness, and fresh-input pricing of $3/M (cached $0.30/M), flat across the full 1M context.
Its trade-offs: coding scores use Moonshot FrontierSWE, not standard SWE-Bench Verified, 2.8T params need serious hardware to self-host — weights are free, running is not, no independent benchmark reproduction yet at release, and image input but no audio or video. At $3 in / $15 out per million tokens, it sits in the mid price band.
The bottom line for this matchup
The defining split here is open vs. closed. Kimi K3 gives you weights you control — self-host it, fine-tune it, keep data in-house, pay only for hardware. Fugu Ultra v2.0 gives you a managed, always-updated API with no infrastructure to run. Teams with GPUs, privacy requirements, or huge volume often favour the open model; teams that want zero ops and the latest capabilities favour the closed one. Capability is close enough that this operational question, not the benchmark, usually decides it.
Frequently asked questions
Is Fugu Ultra v2.0 or Kimi K3 better for coding?
Public SWE-Bench figures are not available for either model, so the honest test is your own repository — run an identical real bug through both. By design, Fugu Ultra v2.0 leans toward sakana's own evaluation reports best-or-joint-best results on 5 of 8 benchmarks tested, including a 74.3 on the deepswe coding-repair benchmark while Kimi K3 leans toward largest open-weight model at release — 2.8t sparse moe, self-hostable, and that positioning usually predicts which feels better on your codebase.
Which is cheaper, Fugu Ultra v2.0 or Kimi K3?
Kimi K3 is open-weight, so self-hosting means no per-token fee (you pay for hardware instead), while Fugu Ultra v2.0 is API-metered at $5/$30 per 1M tokens. For most teams without GPUs, the API model is cheaper to start; at very high volume, self-hosting can win.
Which has the bigger context window?
Effectively neither — 1M tokens vs 1M is a difference of a few percent. Remember advertised ≠ usable: recall typically degrades before the ceiling.
Can I use both Fugu Ultra v2.0 and Kimi K3 together?
Yes — a multi-model platform like LumiChats gives you Fugu Ultra v2.0, Kimi K3 and 40+ others under one ₹69/day pass (about $1/day), so you can draft with one and cross-check with the other instead of buying two subscriptions.
Which is newer, Fugu Ultra v2.0 or Kimi K3?
Fugu Ultra v2.0 — released September 10, 2026, about 45 days after Kimi K3.
Fugu Ultra v2.0 vs Kimi K3
Sakana AI · Global | Moonshot AI · China · Updated June 2026
Quick verdict
Pick Fugu Ultra v2.0 for sakana's own evaluation reports best-or-joint-best results on 5 of 8 benchmarks tested, including a 74.3 on the deepswe coding-repair benchmark or claims to outperform flagship models like gpt-6 astra and claude fable 5.1 despite its underlying orchestration pool reportedly excluding both of them. Pick Kimi K3 for largest open-weight model at release — 2.8t sparse moe, self-hostable or 1m-token context with native vision (text, image and video). Choose Kimi K3 if you need self-hosting or data privacy; Fugu Ultra v2.0 if you want a managed API.
Fugu Ultra v2.0 (Sakana AI) and Kimi K3 (Moonshot AI) are two of the models people most often weigh against each other in 2026. Fugu Ultra v2.0 is sakana AI's flagship multi-agent orchestration model, claiming best-or-joint-best results on 5 of 8 self-reported benchmarks against monolithic frontier models — $5/$30 per million tokens. Kimi K3 is moonshot's 2.8-trillion-parameter open-weight model — the largest open AI at release, with a 1M context and strong self-reported coding scores. They diverge most on price, context window and open vs. closed weights — each quantified below from the models' real specs.
Key differences at a glance
▸Price: Kimi K3 is about 1.7× cheaper on input ($3/$15 per 1M tokens vs $5/$30 per 1M tokens) — modest, but it adds up at steady volume.
▸Context window: 1M tokens vs 1M — within a few percent of each other, so treat this as a tie and test on your own long inputs, since usable recall varies by model.
▸Recency: Fugu Ultra v2.0 is the newer model by about 45 days (released September 10, 2026), usually meaning fresher training data and capabilities.
Side-by-side specs
Spec
Fugu Ultra v2.0
Kimi K3
Provider
Sakana AI (Global)
Moonshot AI (China)
Released
September 10, 2026
July 27, 2026
Context window
1M tokens (~1,500 pages)
1M (~1,573 pages)
Price (in/out)
$5/$30 per 1M tokens
$3/$15 per 1M tokens
Open weight?
No — API only
Yes — self-hostable
Modalities
text
text, image, video, code
SWE-Bench Verified
Not published
Not published
MRCR v2 @ 1M
Not published
Not published
Who wins what
Sakana's own evaluation reports best-or-joint-best results on 5 of 8 benchmarks tested, including a 74.3 on the DeepSWE coding-repair benchmark
Fugu Ultra v2.0
Kimi K3 is comparatively weak here — coding scores use Moonshot FrontierSWE, not standard SWE-Bench Verified
Claims to outperform flagship models like GPT-6 Astra and Claude Fable 5.1 despite its underlying orchestration pool reportedly excluding both of them
Fugu Ultra v2.0
Sakana AI's flagship multi-agent orchestration model, claiming best-or-joint-best results on 5 of 8 self-reported benchmarks against monolithic frontier models — $5/$30 per million tokens — and it is the newer of the two.
1M-token context window
Fugu Ultra v2.0
Fugu Ultra v2.0 lists 1M-token context window among its strengths; Kimi K3 does not.
Largest open-weight model at release — 2.8T sparse MoE, self-hostable
Kimi K3
Open weights make this possible at all — Fugu Ultra v2.0 is API-only, so it cannot leave the vendor's servers.
1M-token context with native vision (text, image and video)
Kimi K3
Fugu Ultra v2.0 is comparatively weak here — pricing rises to roughly $10/$45 per million tokens for prompts above 272K tokens
Vendor reports 81.2 FrontierSWE and 88.3 Terminal-Bench 2.1 on its own harness
Kimi K3
Fugu Ultra v2.0 is comparatively weak here — not a single trained model — Sakana describes Fugu as "a Multi-Agent System, Delivered as One Model" that dynamically routes tasks across a pool of other models, so head-to-head benchmark framing against monolithic frontier models should be read skeptically
Lowest cost at scale
Kimi K3
At $3/$15 per 1M tokens, it is the cheaper of the two — the gap dominates the bill on high-volume workloads.
Which should you pick?
A cost-sensitive startup shipping high volume
→ Kimi K3
At $3/$15 per 1M tokens it undercuts Fugu Ultra v2.0, and on millions of tokens that margin decides the monthly bill.
Someone analysing very long documents or codebases
→ Kimi K3
Larger 1M window fits more in one prompt.
A team with data-privacy or self-hosting needs
→ Kimi K3
Open weights let you run it on your own hardware; Fugu Ultra v2.0 is API-only.
Anyone whose priority is sakana's own evaluation reports best-or-joint-best results on 5 of 8 benchmarks tested, including a 74.3 on the deepswe coding-repair benchmark
→ Fugu Ultra v2.0
It is specifically built for that.
Anyone whose priority is largest open-weight model at release — 2.8t sparse moe, self-hostable
→ Kimi K3
That is its strongest area.
Fugu Ultra v2.0: where it fits
Sakana AI's flagship multi-agent orchestration model, claiming best-or-joint-best results on 5 of 8 self-reported benchmarks against monolithic frontier models — $5/$30 per million tokens. Released September 10, 2026 by Sakana AI, it is built for sakana's own evaluation reports best-or-joint-best results on 5 of 8 benchmarks tested, including a 74.3 on the DeepSWE coding-repair benchmark, claims to outperform flagship models like GPT-6 Astra and Claude Fable 5.1 despite its underlying orchestration pool reportedly excluding both of them, and 1M-token context window.
Its trade-offs are real: not a single trained model — Sakana describes Fugu as "a Multi-Agent System, Delivered as One Model" that dynamically routes tasks across a pool of other models, so head-to-head benchmark framing against monolithic frontier models should be read skeptically, pricing rises to roughly $10/$45 per million tokens for prompts above 272K tokens, and benchmark comparisons against GPT-6 Astra and Claude Fable 5.1 are Sakana's own self-reported figures, not independently verified head-to-head scores. At $5 in / $30 out per million tokens, it sits in the premium price band.
Kimi K3: where it fits
Moonshot's 2.8-trillion-parameter open-weight model — the largest open AI at release, with a 1M context and strong self-reported coding scores. Released July 27, 2026 by Moonshot AI, it is built for largest open-weight model at release — 2.8T sparse MoE, self-hostable, 1M-token context with native vision (text, image and video), vendor reports 81.2 FrontierSWE and 88.3 Terminal-Bench 2.1 on its own harness, and fresh-input pricing of $3/M (cached $0.30/M), flat across the full 1M context.
Its trade-offs: coding scores use Moonshot FrontierSWE, not standard SWE-Bench Verified, 2.8T params need serious hardware to self-host — weights are free, running is not, no independent benchmark reproduction yet at release, and image input but no audio or video. At $3 in / $15 out per million tokens, it sits in the mid price band.
The bottom line for this matchup
The defining split here is open vs. closed. Kimi K3 gives you weights you control — self-host it, fine-tune it, keep data in-house, pay only for hardware. Fugu Ultra v2.0 gives you a managed, always-updated API with no infrastructure to run. Teams with GPUs, privacy requirements, or huge volume often favour the open model; teams that want zero ops and the latest capabilities favour the closed one. Capability is close enough that this operational question, not the benchmark, usually decides it.
Want both Fugu Ultra v2.0 and Kimi K3 without two subscriptions? LumiChats gives you these plus 40+ models under one ₹69/day pass (about $1/day) — draft with one, cross-check with the other.
Public SWE-Bench figures are not available for either model, so the honest test is your own repository — run an identical real bug through both. By design, Fugu Ultra v2.0 leans toward sakana's own evaluation reports best-or-joint-best results on 5 of 8 benchmarks tested, including a 74.3 on the deepswe coding-repair benchmark while Kimi K3 leans toward largest open-weight model at release — 2.8t sparse moe, self-hostable, and that positioning usually predicts which feels better on your codebase.
Which is cheaper, Fugu Ultra v2.0 or Kimi K3?
Kimi K3 is open-weight, so self-hosting means no per-token fee (you pay for hardware instead), while Fugu Ultra v2.0 is API-metered at $5/$30 per 1M tokens. For most teams without GPUs, the API model is cheaper to start; at very high volume, self-hosting can win.
Which has the bigger context window?
Effectively neither — 1M tokens vs 1M is a difference of a few percent. Remember advertised ≠ usable: recall typically degrades before the ceiling.
Can I use both Fugu Ultra v2.0 and Kimi K3 together?
Yes — a multi-model platform like LumiChats gives you Fugu Ultra v2.0, Kimi K3 and 40+ others under one ₹69/day pass (about $1/day), so you can draft with one and cross-check with the other instead of buying two subscriptions.
Which is newer, Fugu Ultra v2.0 or Kimi K3?
Fugu Ultra v2.0 — released September 10, 2026, about 45 days after Kimi K3.
Specifications and benchmarks reflect publicly reported figures as of June 2026 and may change as providers release updates. Always verify on your own workload.