Overview
Kimi K2.6 is Moonshot AI's April 20, 2026 release of its flagship open-weight assistant, landing about three months after K2.5 and pushing the model further into long-horizon, tool-heavy agentic work. Where K2.5 centered on native multimodal reasoning and the first generation of Agent Swarm, K2.6 is an operational upgrade aimed at developers who need a model that stays coherent across thousands of tool calls and many-hour autonomous sessions.
The release ships simultaneously across Kimi's main surfaces: kimi.com for web, the Kimi mobile app, the Kimi API at platform.kimi.ai, and Kimi Code, the command-line coding assistant. The standard context window remains 262,144 tokens, matching the prior 256K-class generation, while the headline gains are measured: Terminal-Bench 2.0 rises from 50.8% to 66.7% and SWE-Bench Pro from 50.7% to 58.6%; the 96.60% tool invocation figure is from CodeBuddy's internal evaluation quoted in Moonshot's launch blog, not the public benchmark table. K2.6 also introduces a research preview of "Claw Groups" — heterogeneous human-AI multi-agent collaboration — and deeper support for 24/7 proactive agents like OpenClaw and Hermes.
What's New
Long-Horizon Coding Execution
K2.6's biggest behavioral change is endurance. Moonshot demonstrates autonomous sessions running 12+ hours with 4,000+ sequential tool calls on real engineering tasks — Zig compiler optimization passes, full refactors of financial trading engines, and similar long-running workloads. K2.5 already supported large multi-step agentic workflows, but K2.6 is the first K-series release where Moonshot publicly showcases 12+ hour, 4,000+ tool-call autonomous coding runs as a headline capability.
Agent Swarm Scaled 3× to 300 Sub-Agents
Agent Swarm in K2.5 coordinated up to 100 parallel sub-agents. K2.6 scales that ceiling to 300 and explicitly ties it to 4,000-step workflows, meaning more decomposed subtasks can run simultaneously without hitting the orchestration cap. This is aimed at production pipelines — parallel code review, multi-document research, and parallelized test generation — where the K2.5 ceiling was already the bottleneck.
Coding-Driven Design and Simple Full-Stack Workflows
K2.6 adds end-to-end coding-driven design: natural-language prompts produce front-end interfaces including animations, interaction behaviors, and connected back-end stubs, not just standalone UI components as in K2.5. Moonshot's launch materials cite partners including CodeBuddy ("significant improvements"), Factory.ai ("+15% on benchmarks"), and Vercel (">50% improvement on Next.js benchmarks") as third-party indicators of the coding uplift.
Claw Groups (Research Preview)
A new research preview called Claw Groups introduces heterogeneous multi-agent collaboration where specialized agents — including human agents — work alongside Kimi inside a single task graph. This is positioned as exploratory: Moonshot has not yet exposed Claw Groups as a generally available API and is framing it as a glimpse of where the platform is heading.
Proactive 24/7 Agents
K2.6 is tuned for reliability under continuous operation, and Moonshot explicitly calls out integration with proactive agent frameworks such as OpenClaw and Hermes. These run as 24/7 background workers that initiate their own tool calls, and the K2.6 release notes emphasize fewer drift failures across multi-day runs.
Performance Benchmarks
All numbers from Moonshot's official K2.6 release materials, with K2.5 as the direct baseline where available.
| Benchmark | K2.6 | K2.5 | Δ |
|---|---|---|---|
| Terminal-Bench 2.0 | 66.7% | 50.8% | +15.9 pp |
| SWE-Bench Pro | 58.6% | 50.7% | +7.9 pp |
| BrowseComp | 83.2% | 74.9% | +8.3 pp |
| DeepSearchQA (F1) | 92.5% | 89.0% | +3.5 pp |
| Tool invocation success | 96.60% | — | New headline |
Moonshot also reports internal-evaluation gains of roughly +12% in code generation accuracy and +18% in long-context stability versus K2.5. For teams choosing between K2.5 and K2.6, the Terminal-Bench and SWE-Bench jumps are the most decision-relevant — both reflect harder, multi-turn tool-use scenarios rather than single-shot completion tasks, and align with the long-horizon execution pitch.
Availability & Access
K2.6 is live day-of-release on every Kimi surface:
- kimi.com — web chat, free tier
- Kimi App — iOS and Android
- API — platform.kimi.ai (OpenAI-compatible endpoints)
- Kimi Code — official CLI coding assistant
Claw Groups remains a research preview in Moonshot's launch materials. K2.6 is officially available on kimi.com, the Kimi App, the API, and Kimi Code. API users should call the documented model identifier kimi-k2.6; default web/app routing is not asserted here without separate verification.
Pricing & Plans
Kimi offers free access on kimi.com and the mobile app, while the broader product also has paid membership plans. API pricing on platform.kimi.ai is usage-based rather than flat-rate.
- kimi.com web chat — Free access to K2.6 with standard rate limits.
- Kimi App (iOS / Android) — Free access to K2.6.
- API (platform.kimi.ai) — Usage-based; current per-token prices are published on Moonshot's pricing page and should be confirmed there before budgeting, because API rates can shift between releases.
- Kimi Code (CLI) — Available through Kimi Code membership plans; official docs describe it as a premium subscription benefit, with login or API-key access managed in the Kimi Code console.
This release does not publicly announce a new paid plan or a change in free-tier policy; cost comparisons against other LLMs should pull the latest per-token numbers from the official pricing page.
Best For
- Teams running production coding agents where K2.5's Agent Swarm ceiling (100 sub-agents) was the bottleneck.
- Developers on Kimi Code who want longer uninterrupted autonomous sessions for refactors and migrations.
- Research teams wanting a strong open-weight-style alternative to closed frontier models on BrowseComp and DeepSearchQA-style tasks.
- Platforms building 24/7 proactive agents (OpenClaw, Hermes, or custom) where tool invocation reliability matters more than single-turn completion quality.
- Engineering orgs benchmarking Kimi against Claude Opus 4.7 on SWE-Bench-class workloads before committing infrastructure.
- Cost-sensitive teams who need frontier-grade capabilities with a free web/app tier for experimentation before paying for API usage.
FAQ
How is Kimi K2.6 different from K2.5?
K2.6 is an operational upgrade focused on long-horizon agentic work. Agent Swarm scales from 100 to 300 parallel sub-agents; autonomous sessions run 12+ hours across 4,000+ tool calls; Terminal-Bench 2.0 rises from 50.8% to 66.7%; SWE-Bench Pro rises from 50.7% to 58.6%; and BrowseComp rises from 74.9% to 83.2%. The 262K context window is unchanged. The headline deltas are endurance and tool-use reliability rather than a new modality.
Where can I use K2.6?
K2.6 is live on kimi.com (web), the Kimi App (iOS and Android), the API at platform.kimi.ai, and Kimi Code (CLI). Existing K2.5 users on the web and app are routed to K2.6 by default; API users can pin versions via the model parameter.
Is K2.6 free to use?
Yes, on kimi.com and the Kimi App. API usage through platform.kimi.ai is priced per token and is published on Moonshot's pricing page — confirm the current rate there before building cost projections.
What is Claw Groups?
Claw Groups is a research preview introduced with K2.6 that enables heterogeneous multi-agent collaboration, including human-AI teaming, inside a single task graph. At launch it is not exposed through the standard API and is framed as a preview of where the platform is heading rather than a shipping feature.
Do I need to change anything when upgrading from K2.5?
For most interactive web and app usage, no — K2.6 is a drop-in replacement. API consumers should confirm the model identifier for K2.6 on platform.kimi.ai, verify current per-token pricing, and retest long-running agent harnesses: the bigger Agent Swarm ceiling and longer autonomous sessions may change where your harness's own limits bind.
Are the third-party benchmark numbers trustworthy?
CodeBuddy, Factory.ai, and Vercel's numbers are vendor-reported and were published alongside the Moonshot launch. Terminal-Bench, SWE-Bench Pro, BrowseComp, and DeepSearchQA are standard public benchmarks. Teams making production decisions should cross-check with independent evaluations, especially on workloads that differ from benchmark distributions.
