Kimi K3 deep dive: how far the 2.8T open flagship sits from Claude and GPT

What it is

Kimi K3 is Moonshot AI’s flagship, shipped 2026-07-16: 2.8 trillion parameters in Stable LatentMoE (16 of 896 experts active per token), built on Kimi Delta Attention (KDA) and Attention Residuals (AttnRes), with native vision and video input and a 1M-token context window. In its official tech blog, Moonshot states K3 still trails Claude Fable 5 and GPT 5.6 Sol overall, while showing frontier-level results across its own evaluation suite.

As of 2026-07-22, K3 is available on Kimi.com, Kimi Work, Kimi Code, and the Kimi API; full weights are promised by 2026-07-27 and are not downloadable yet—Artificial Analysis still classifies it as a proprietary API model.

Spec sheet and architecture

By the numbers: a 1M-token context window; input modalities text, image, and video (API docs), output text. Architecturally, K3 delivers roughly 2.5× overall scaling efficiency versus Kimi K2 (vendor claim); quantization-aware training from SFT uses MXFP4 weights with MXFP8 activations.

Reasoning is always on, with reasoning_effort=max by default (low/high coming later). Multi-turn and tool-call flows must return the complete assistant message—including reasoning—or quality can collapse, the first item in Moonshot’s official Limitations. API model ID: kimi-k3; OpenRouter: moonshotai/kimi-k3. K3 is a recharge-unlocked flagship SKU (minimum $1 top-up).

Pricing and where to run it

Official Moonshot API pricing (flat across the full 1M window, no length tiers): cache-hit input $0.30/MTok, cache-miss input $3/MTok, output $15/MTok. This snapshot verifies $3/$15 on the OpenRouter primary listing (is_primary_listing=1); moonshot-direct and SiliconFlow also show cache $0.30; Groq lists $3/$15.

Moonshot claims a cache hit rate above 90% on coding workloads—long-prefix agent loops can materially lower effective input cost. Output is 5× cache-miss input; long generations and reasoning tokens inflate the bill.

Inside the family: K2.6 → K3

K3 is not a “bigger K2.6” but a generational architecture jump. Parameters move from ~1T to 2.8T; context from 262K to 1M; modalities from vision to native video; attention shifts to KDA + AttnRes.

On the OpenRouter primary listing, K2.6 is $0.684/$3.42 versus K3 at $3/$15—roughly 4.4× on input and output, about 4–5× blended. Capability-wise, Artificial Analysis Intelligence Index v4.1 rises from ~44 (K2.6) to ~57 (K3) per the independent third party as of 2026-07-22.

Verdict: if your Kimi workflow is long autonomous coding, cost-sensitive, and 256K is enough, K2.6 remains the sensible default. Pay K3’s 4–5× premium only when you need a 1M repo window, native video, or a ceiling closer to closed frontier models.

The real gap vs. frontier leaders

Composite intelligence: Artificial Analysis Intelligence Index v4.1 (single methodology, independently run) scores Claude Fable 5 ≈60, GPT 5.6 Sol ≈59, Kimi K3 ≈57, and Claude Opus 4.8 ≈56—a three-point spread that puts all four in the same frontier tier, not a blowout. Moonshot’s official blog likewise states K3 still trails Fable 5 and Sol overall.

Coding splits (Moonshot official benchmark table plus third-party reads; different agent harnesses—do not treat as apples-to-apples): Sol leads DeepSWE (73.0) and Terminal-Bench 2.1 (88.8); K3 leads SWE Marathon (42.0 vs Fable 35.0), Program Bench (77.8), and Frontend Code Arena (#1); Fable leads FrontierSWE (86.6) and GDPval-AA v2 Elo (1760 vs K3 1668).

Agent/knowledge work: Fable leads JobBench and AA-Briefcase; K3 stands out on BrowseComp (91.2), AutomationBench, and SpreadsheetBench 2. Experience-wise, official Limitations flag thinking-history sensitivity, excessive proactiveness, and UX still behind Fable/Sol; AA measures K3 as relatively slow (~36 t/s) and verbose on the Index—cheap list price is not cheap per task. These are external benchmarks, not AI Hippo’s context-per-dollar proxy score.

Cost in practice: two scenarios

Modeled with this snapshot’s OpenRouter primary listing (cache-miss input), excluding extra reasoning tokens.

Scenario A—agent coding turn (100K input + 20K output): K3 about $0.60 ($0.30+$0.30), K2.6 about $0.14, GPT-5.6 Sol about $1.10, Fable 5 about $2.00. K3 runs roughly 4.3× K2.6 but far below Fable.

Scenario B—long-document pass (500K input + 5K output): K3 at miss pricing about $1.58; if all 500K input cache-hits at $0.30, the same pass is about $0.23 input + $0.075 output ≈ $0.30 total. Sol about $2.65, Fable about $5.25, K2.6 about $0.36.

Takeaway: K3’s price edge shows up most in long-prefix, high-cache-hit agent loops; once reasoning inflates output, real bills drift toward AA’s per-task cost. These are spec and cost comparisons, not capability benchmarks.

When to choose it—and when not to

Choose K3 for long-horizon agent coding (terminal/browser loops), 1M whole-repo or long-document work, native video understanding, and testing an open-weight path near the frontier via API before weights land—provided you accept mandatory thinking and harness constraints.

Skip K3 or look elsewhere for one-shot professional deliverables (Fable 5), a mature Claude Code ecosystem with lowest UX friction (Fable/Sol), pure throughput with tight budgets and 256K enough (stay on K2.6), or downloadable weights for self-hosting (wait until after 2026-07-27 and validate license plus cluster cost).

Decide with AI Hippo’s model, Token, and compare pages—not the composite rank alone (that score is a context-per-dollar proxy, not a capability benchmark).

Deployment checklist

Before you integrate: confirm the harness returns complete thinking history (officially verified stacks like Kimi Code); avoid switching to K3 mid-session from another model; constrain excessive proactiveness in the system prompt or AGENTS.md.

Operationally: design long prefixes to benefit from prefix cache; monitor tokens per task—including reasoning—not list price alone. If weights ship after 2026-07-27, self-hosting 2.8T MoE will demand a far larger cluster than 1T-class K2.6—size the hardware before you call it “fully open.”

Sources

• Kimi K3 on AI Hippo
• Kimi K3 official tech blog
• Kimi K3 API quickstart
• Artificial Analysis — Kimi K3
• Moonshot AI
• OpenRouter models API