Kimi K3: 2.8 Trillion Parameters, Half the Cost Per Task of Opus 4.8

What shipped

Kimi K3 is a mixture-of-experts model with 2.8 trillion total parameters, of which 16 of 896 experts activate per token, giving it an activation ratio under 2%. It uses Moonshot's Kimi Delta Attention and an Attention Residuals mechanism, ships MXFP4 weights with MXFP8 activations, and holds a 1-million-token context window with native multimodal (text and image) input. Moonshot calls it “the world's first open 3T-class model” while stating plainly in its own launch post that it “still trails the most powerful proprietary models, Claude Fable 5 and GPT-5.6 Sol” — an unusually direct admission for a launch announcement, and an accurate one.

Where it actually lands

On the Artificial Analysis Intelligence Index, K3 scores 57, ranking #4 of 189 tracked models: behind Fable 5 and GPT-5.6 Sol, ahead of Claude Opus 4.8 and GPT-5.5. On agentic and coding-specific evaluations the picture is more mixed and, in a few places, better than the headline index suggests:

The AutomationBench-AA result is the one worth sitting with. Yesterday's deep dive on this exact benchmark showed the entire frontier field topping out under 50%, with Claude Fable 5's headline 48.6% score quietly depending on a silent fallback to Claude Opus 4.8 on roughly 18% of tasks. K3 clears that ceiling outright, at 53%, as a single model with no fallback disclosed — on a benchmark built specifically to catch agents that look good on paper and fall apart on real, rule-bound business work. Read the full AutomationBench-AA breakdown for the methodology this result is measured against.

The price: Sonnet-tier, half the cost per task of Opus

K3's sticker price matches Claude Sonnet 5's: $3 per million input tokens, $15 per million output tokens, with automatic prompt caching at a 90% discount ($0.30 per million on cache hits). Moonshot reports 90%+ cache hit rates on coding workloads, and independent tracking puts K3's average cost per completed taskat $0.94, against $1.80 for Claude Opus 4.8 and $1.04 for GPT-5.6 Sol — roughly half of Opus 4.8's per-task cost, achieved with 21% fewer output tokens than its own predecessor, Kimi K2.6, while gaining 13 points on the Intelligence Index.

Weights, timing, and reach

Full open weights ship by July 27, 2026; the model is already live via the Kimi API, Kimi.com, Kimi Work (desktop), and Kimi Code (terminal). Had Moonshot shipped K3 even a month earlier, it would plausibly have held the #1 spot outright instead of trailing Fable 5 and GPT-5.6 Sol — the frontier moved in the same window K3 was training. What it holds today, unambiguously, is the top spot among open-weights models on general intelligence, and the top spot among every model, open or closed, on the one benchmark built to measure real business work instead of exam questions.

Available now on MegaBrain

Kimi K3 is live on the MegaBrain gateway today under moonshotai/kimi-k3, at the same zero-markup, pass-through pricing as every other model on the gateway. It is now the default Kimi model surfaced in the model picker. Switching to it from any other model already on your MegaBrain integration is a one-line change:

import openai

client = openai.OpenAI(
    base_url="https://getmegabrain.com/api/gateway/v1",
    api_key="mb-your-key-here",
)

response = client.chat.completions.create(
    model="moonshotai/kimi-k3",
    messages=[{"role": "user", "content": "Refactor this module to async and add tests"}],
)

Browse live pricing and the exact model ID on the models page. For a long-running coding or research agent on mhermes, K3's combination of a 1M-token context window and roughly half Opus 4.8's cost per completed task is worth routing to for anything that doesn't specifically need Fable 5 or Sol's extra headroom.