Kimi K375
Family profile
Best published score in each covered benchmark family.
75/100 avg
- Knowledge
- 48
- Reasoning
- 94
- Coding
- 59
- Agentic
- 85
- Multimodal
- 81
- Long context
- 83
6 tested benchmark families
VerdictPal · editorial desk · updated 5 Sep 2026VerdictPal
Models
Model-first evidence atlas.
We score models, not vendors. Below: every tooled model the public benchmark ledger has evidence for, sorted by intelligence receipts (Vals Index first, then Artificial Analysis). Every row carries a date and a URL, not a universal verdict.
Current Vals evidence anchor · Vals Index
Kimi K3Moonshot · Kimi K3 · 75 on the Vals intelligence index216
Models with pages
21
Providers
171
Priced in USD/1M
8
In score ledger only
Models with public intelligence rows
Vals Index rows appear first. Models without a Vals row fall back to Artificial Analysis. Source badges show whether a row is single-source or corroborated — not a cross-benchmark winner.
- 175Vals
- 273Vals
- 373Vals
- 470Vals
- 569Vals
- 668Vals
- 767Vals
- 867Vals
- 966Vals
- 1066Vals
Top 10 of 192
Intelligence × blended cost
Blended cost per million tokens (input + 3× output) on the X axis, log scale. Intelligence index on the Y axis — Vals when present, otherwise Artificial Analysis. Dots are sized by context window and connected by the Pareto frontier (best intelligence at each cost).
All 159 models in this chart
Input-only scatter
The earlier cut, using input cost alone. Useful when your workload is input-heavy (RAG, document Q&A) rather than output-heavy.
Hover any dot for name, score, and input price. Keyboard: tab through the top twelve, or use the ranking below.
AnthropicOpenAIMistralAlibabaDeepSeekGoogleMoonshotMetaMiniMaxxAIZhipuXiaomiOtherMeituanCohereNVIDIA
Every flagship model
8 models with benchmark rows but no model page
Older releases, special configurations (Codex CLI, ForgeCode), and ad-hoc aggregations. We surface their sourced scores but we don't promote them to the model atlas until they earn a full model page.
Family profile — strongest current evidence rows
The highest intelligence receipt in each family slice. Use this as a sanity check that a single number isn't hiding a weak spot.
Claude Mythos Preview73
Family profile
Best published score in each covered benchmark family.
91/100 avg
- Reasoning
- 95
- Coding
- 94
- Agentic
- 84
3 tested benchmark families
GPT-5.6 Sol73
Family profile
Best published score in each covered benchmark family.
79/100 avg
- Knowledge
- 59
- Reasoning
- 94
- Coding
- 73
- Agentic
- 88
- Multimodal
- 83
- Long context
- 78
6 tested benchmark families
Which families does the ledger actually cover?
A model with no row in a family is a model that hasn't been publicly tested there. We surface that instead of pretending it's weak.
Knowledge57models covered
Reasoning195models covered
Math51models covered
Coding189models covered
Agentic188models covered
Multimodal15models covered
Long context184models covered
Tool use20models covered
Performance0models covered
Safety9models covered
Human preference13models covered
Where these models show up in the atlas
Every tool on this list is wired to at least one model in the atlas above. The same wrapper that hides the harness can also hide the price.