Moonshot AI's Kimi K2.6 scores 58.6% on SWE-Bench Pro, beating GPT-5.4 (57.7%), Claude Opus 4.6 (53.4%), and Gemini 3.1 Pro (54.2%). It is an open-weight model you can download today and self-host. On paper, that looks like the best deal in frontier coding AI. In practice, one number in Artificial Analysis's evaluation changes the math considerably - and it is not one most coverage mentions.
Artificial Analysis measured K2.6 producing 170 million output tokens across their Intelligence Index evaluation, compared to a median of 47 million for similarly-sized models. That is a 3.6× output multiplier. At $4.00 per million output tokens on Moonshot's API, it compounds fast on agentic workloads. The headline benchmark win is real. The cost story requires more care.
What actually changed between K2.5 and K2.6
Per Moonshot's model card, the architecture is identical to K2.5 down to the parameter count - K2.6 is a re-trained model with a revised post-training pipeline rather than a new topology. That distinction matters when you are evaluating whether to upgrade.
K2.5 is one of the largest open-weight models with 1.04 trillion parameters and 32B activated parameters per token, using 384 experts with 8 activated per token, MLA attention, SwiGLU activation, and a 256K context window. K2.6 shares that entire spec. The hallucination rate on AA-Omniscience fell from 65% on K2.5 to 39% on K2.6 - a calibration jump that matters more for production deployment than most top-line benchmark gains. The gains concentrate in agentic coding and tool use.
The practical upshot: if you are already running K2.5 in a non-agentic workflow, the upgrade is mainly a reliability improvement. If you are running agentic coding tasks, K2.6 is the version worth switching to - the tool-use reliability gap is large enough to change completion rates on real tasks.
Native INT4 quantization via QAT cuts the footprint to ~594 GB with ~2× generation speedup. Unlike post-training quantization, QAT trains with quantization constraints baked in, so quality degradation is limited.
The pricing picture and where self-hosting makes sense
The Kimi API costs $0.95/$4.00 (input/output per million tokens) for K2.6, $0.60/$3.00 for K2.5, and $3.00/$15.00 for the newer K3 flagship. On OpenRouter, K2.6 is available from 21 providers; on OpenRouter specifically, it costs $0.5684/M input tokens and $3.332/M output tokens, with a cache-read rate of $0.0925/M tokens.
For agentic coding sessions that resend large repository context repeatedly, the cache-hit input rate is $0.19 per million tokens on K2.7 Code, and $0.15 on older Kimi K2; for coding agents that resend large repository context on every turn, that cache lane is where most of the real savings live.
The INT4 quantized version needs approximately 500 GB of VRAM (4× A100 80GB or 8× RTX 4090); the full FP16 version requires around 2 TB of VRAM.
The API at $0.60 per million input tokens (K2.5 tier) is very affordable. Self-hosting makes sense above roughly 50 million tokens per day.
| Moonshot API (K2.6) | OpenRouter (K2.6) | Self-hosted INT4 | |
|---|---|---|---|
| Input ($/M tokens) | $0.95 | $0.57 | Hardware cost only |
| Output ($/M tokens) | $4.00 | $3.33 | Hardware cost only |
| Cache-read input | $0.19 | $0.09 | - |
| Min. hardware | None | None | ~500 GB VRAM |
| Break-even | Below ~50M tokens/day | Below ~50M tokens/day | Above ~50M tokens/day |
The inference paradox K2.6 sits inside
This model lands in a market-wide dynamic that no benchmark score escapes. Blended AI token prices fell roughly 67% year over year - from $18.40 to $6.07 per million tokens between Q1 2025 and Q1 2026. Yet 73% of enterprises exceeded their original AI cost projections last fiscal year; Uber's CTO burned the company's full 2026 AI coding budget in four months.
The mechanism is straightforward: routing a task to an agentic reasoning model increases provider inference costs by at least five times, and often much more as task complexity grows. K2.6's own output-token behavior is a version of this. Reasoning models can consume 100× more tokens internally than they output, creating a cost paradox where cheaper per-token pricing leads to higher total bills.
The strategic implication: the competitive moat in AI agent systems will not be access to cheap inference, but rather the quality of agent architecture, memory systems, tool integrations, and organizational knowledge embedded in agent behavior. Infrastructure cost will cease to be a differentiator; the quality of what agents do with that compute will be the remaining axis of competition.
K2.6 scores well here because the calibration improvement (65% → 39% hallucination) translates directly to fewer retries and fewer wasted tool calls. A more reliable model uses fewer output tokens to finish the same task.
Beagle in action#engineering, 10:42amOne licensing detail that matters at scale
The Modified MIT license requires visible Kimi K2.6 branding on products with 100 million or more monthly active users or $20 million or more in monthly revenue. For most companies this is irrelevant. For any consumer-facing product at hyperscaler scale, it is a legal review item before launch.
Per multiple reports in mid-2026, K2.6 also appeared in Perplexity's model picker, making it one of the first Kimi models in a major Western AI product. That is a practical data point on production readiness: the model has gone through at least one external operator's quality bar.
Kimi K2.6 open-weight model: common questions
How does Kimi K2.6 compare to GPT-5.4 on coding benchmarks?
K2.6 scores 58.6% on SWE-Bench Pro, ahead of GPT-5.4 at 57.7% and Claude Opus 4.6 at 53.4%. The gains concentrate in agentic coding and tool use. On pure reasoning tasks, the advantage narrows. It costs roughly 80% less per million tokens than GPT-5.5 at standard API rates.
What hardware does self-hosting Kimi K2.6 require?
The INT4 quantized version needs approximately 500 GB of VRAM, achievable with 4× A100 80GB or 8× RTX 4090; the full FP16 version requires around 2 TB of VRAM.
vLLM is recommended for production API serving; SGLang is better for structured generation and agent frameworks.
Is Kimi K2.6 actually open-source?
Most models called "open source" are really "open weight": only the model weights are publicly available, but the training data and pipeline remain proprietary. K2.6 fits that description. Weights are on Hugging Face under a Modified MIT License; the training recipe is not published.
Why does K2.6 cost more in agentic workflows than the token price suggests?
Artificial Analysis measured K2.6 producing 170 million output tokens in their Intelligence Index evaluation, versus a median of 47 million for similarly-sized models. If you are running long-context agentic tasks in thinking mode, the output cost can erode the input cost advantage faster than you expect. Always benchmark your specific workload before projecting spend.
How is K2.6 different from Kimi K3?
K3 substantially outperforms K2.6 on coding benchmarks (76.8% vs 58.6% on SWE-bench) while requiring roughly 8 times more GPU infrastructure.
K3 costs $3.00 per million cache-miss input tokens and $15.00 per million output tokens, with a 1M-token context window
- a different cost tier entirely. K2.6 remains the practical choice for teams that do not need K3-class performance or cannot provision the hardware for it.