Kimi K2.6 is live on Kyma

April 21, 20263 min read

Kimi K2.6

Moonshot shipped Kimi K2.6 on April 20, 2026. We rolled it onto Kyma the next day. If you use the agent or best-agent alias, you're already on it.

What changed

K2.6 is Moonshot's open-source agentic flagship. The things that matter for builders:

•256K context across all variants (we expose 262,144 tokens on Kyma)
•Native multimodal — text + image input, text output
•Agent swarms scaled 3× — 300 sub-agents orchestrated over 4,000+ coordinated steps (K2.5 topped out at 100 agents / 1,500 steps)
•Long-horizon execution — Moonshot designed it to hold a plan together across ~12 hours of work
•Modified MIT license on both code and weights

Benchmarks (Moonshot's own numbers)

These are from Moonshot's official release tech blog — not our measurements. We'll publish independent numbers once we have real production traffic.

Benchmark	K2.6	Claude Opus 4.6	GPT-5.4	Gemini 3.1 Pro	K2.5
SWE-Bench Verified	80.2	80.8	—	80.6	76.8
SWE-Bench Pro	58.6	53.4	57.7	54.2	50.7
Terminal-Bench 2.0	66.7	65.4	65.4	68.5	50.8
LiveCodeBench v6	89.6	88.8	—	91.7	85.0
BrowseComp	83.2	83.7	82.7	85.9	74.9
Toolathlon	50.0	47.2	54.6	48.8	27.8
HLE-Full (w/ tools)	54.0	53.0	52.1	51.4	50.2
DeepSearchQA (f1)	92.5	91.3	78.6	81.9	89.0

Bold = K2.6 leads. The pattern to read: K2.6 is essentially tied with Claude Opus 4.6 on SWE-Bench Verified (80.2 vs 80.8), ahead on SWE-Bench Pro (harder, multi-step problems), and substantially ahead of its own prior generation on anything agent-shaped. It's behind Gemini 3.1 Pro on some raw benchmarks but comparable once you turn on tool use.

The K2.5 → K2.6 jump on Toolathlon (27.8 → 50.0) is the number we'd pay the most attention to — it measures how reliable the model is across long chains of tool calls, which is where cheap models fall apart in real agent workloads.

Pricing on Kyma

	Per 1M tokens
Input	$1.28
Output	$5.40

For reference, K2.5 stays at $0.68 / $3.78. K2.6 is the premium option — you're paying for the agent-swarm improvements, not raw throughput.

No per-provider disclosure — we route through whichever upstream is cheapest and healthy, and your price is fixed regardless of which inference backend actually served your request. If the primary dies, we fall over to a secondary automatically (4-layer routing).

How to use it

Easiest — use the alias, auto-upgrade to future flagships:

from openai import OpenAI

client = OpenAI(base_url="https://kymaapi.com/v1", api_key="ky-...")

response = client.chat.completions.create(
    model="agent",  # → kimi-k2.6
    messages=[{"role": "user", "content": "Refactor the settle_request RPC to write to transactions"}],
    tools=[...]
)

Pin the exact version if you need reproducibility:

response = client.chat.completions.create(
    model="kimi-k2.6",
    messages=[...],
    tools=[...]
)

Heads up on max_tokens. K2.6 is a reasoning model — it spends completion tokens on reasoning_content before emitting the final content. If you set max_tokens: 20 and the model spends 30 tokens thinking, you'll get an empty content field. Give it at least 200 tokens of headroom for any non-trivial query, more for multi-step agent work.

Upgrading from K2.5

If you were already using "model": "agent", no change — you're already on K2.6. The alias map updated when we shipped.

If you hardcoded "model": "kimi-k2.5", you have two choices:

•Recommended: switch to "model": "kimi-k2.6" (or "model": "agent").
•Stay on K2.5: keep it. K2.5 is still a first-class active model on Kyma, same routing, same SLA.

We won't deprecate K2.5 without notice. The family page at /models/kimi documents both.

What's next

We're watching K2.6's production behavior over the next week — primary route stability, tail latency on long agent runs, cache behavior. If anything changes (e.g., we swap primary providers, add prompt caching when it becomes available), the model page at /models/kimi-k2.6 stays current.

If you want to compare on your own workload: sign in at kymaapi.com, $5 of free credits on signup covers a few million tokens of K2.6 to stress-test before committing.