Kimi K3

The Best Open-Weights Model Nobody Priced This Well

Start with the combination, because it is unprecedented in our dataset: Kimi K3 is the highest-overall open-weights model we track (56.0), it scores 85.7 on Terminal-Bench - better than DeepSeek’s flagship and second only to the closed frontier pair - and it costs $4.33 per million tokens with a 1M context window. Moonshot AI has positioned K3 exactly where developer demand is: agentic coding quality near the closed frontier, at a price that survives high-volume production use.

Benchmark Deep Dive, In Plain English

• Terminal-Bench 85.7 - The live-terminal evaluation: clone the repo, reproduce the bug, fix it, make the tests pass, leave the environment working. K3’s 85.7 puts it in the same conversation as Claude Fable 5 (88.0) while costing 3.3 times less per token. If your coding agent spends its day in a shell, this is the budget pick that does not feel like one.
• GPQA 64.2 and MATH-500 59.8 - Graduate-level science reasoning and competition math. K3 sits between DeepSeek-V4-Pro (weaker math) and GPT-5.6 Sol (stronger) - a genuine mid-tier in reasoning, not a coding specialist pretending to be general.
• BrowseComp 46.7 - Web-research tasks, marginally ahead of Sol’s 45.8. Combined with the 1M context window, K3 handles long research-and-synthesis chains respectably.

When to Choose K3

Choose K3 when you need near-frontier agentic coding at a price that scales: startup products, agent fleets, CI pipelines that review every pull request. It is the strongest open-weights option for teams that want self-hosting rights without giving up Terminal-Bench performance - the exact gap between K3 and DeepSeek-V4-Pro-0813 is terminal proficiency, and it is large (85.7 vs 80.3). If your work is math-heavy instead, check GPT-5.6 Sol or compare the two directly on the comparison engine.

Pricing Reality Check

$4.33 flat per million tokens puts the standard 10,000-request workload (15K tokens average) at about $650 - roughly 30% of Claude Fable 5’s bill for the same volume, with 80.5 vs 84.5 SWE-bench as the quality trade. For teams running continuous background agents (code review, triage, documentation sync), that difference decides whether the project gets budget approval at all.

Real Limitations

Peak capability is not the frontier: on the hardest reasoning and the most novel debugging, Fable 5 and GPT-5.6 Sol still pull away, and K3’s arena Elo (1682) sits below both. Throughput is 77 tokens per second - slower than DeepSeek’s 176, which matters for long generation loops. And open-weights operations apply: self-hosting means GPU capacity planning, and the API tier carries the usual provider rate limits at scale.

Where to Go Next

• Direct comparison: Kimi K3 vs DeepSeek-V4-Pro
• The complete value analysis: Best LLM for Coding
• Standings across all pillars: AI Model Leaderboard