Summarize with AI
Open this article in your favorite assistant
The US export-control ban on Claude Fable 5 and Mythos 5 created an immediate gap for developers outside the United States who need frontier-class coding assistance. Kimi K2.7 Code, released by Moonshot AI on June 12, 2026, is the strongest open-weight response so far. It is not a direct replacement for Fable 5's general reasoning, but for agentic coding it is significantly cheaper, globally available, and easier to self-host.
What is Kimi K2.7 Code?
Kimi K2.7 Code is a coding-specialized, open-weight agentic model built on the same Mixture-of-Experts architecture as Kimi K2.6. It refines the base model for long-horizon software-engineering tasks, reduces reasoning-token waste, and adds an Anthropic-compatible API endpoint that lets it drop into Claude Code, Cline, and RooCode without changes.
| Spec | Kimi K2.7 Code |
|---|---|
| Release date | June 12, 2026 |
| Architecture | 1T parameter MoE, 32B active per token |
| Context window | 256,000 tokens |
| Max output | 66,000 tokens |
| Input modalities | Text |
| Output modalities | Text |
| API model ID | kimi-k2.7-code |
| API endpoints | OpenAI-compatible and Anthropic-compatible |
| Pricing | $0.95 / 1M input, $4.00 / 1M output, $0.16 cached input |
| License | Modified MIT (commercial use allowed) |
| Weights | Available on Hugging Face |
Benchmarks: where K2.7 Code wins
Moonshot's launch benchmarks focus on coding and agentic execution. Independent SWE-bench numbers are not yet available, but the company-reported improvements over K2.6 are substantial, and K2.7 Code beats Claude Opus 4.8 on the MCPMark Verified tool-calling benchmark.
| Benchmark | Kimi K2.7 Code | Kimi K2.6 | Claude Opus 4.8 | GPT-5.5 |
|---|---|---|---|---|
| Kimi Code Bench v2 | 62.0% | 50.9% | 67.4% | 69.0% |
| Program Bench | 53.6% | 48.3% | 63.8% | 69.1% |
| MLS Bench Lite | 35.1% | 26.7% | 42.8% | 35.5% |
| MCPMark Verified | 81.1% | 72.8% | 76.4% | 92.9% |
| MCP Atlas | 76.0% | 69.4% | 81.3% | 79.4% |
| Kimi Claw 24/7 Bench | 46.9% | 42.9% | 50.4% | 52.8% |
| Thinking-token reduction | ~30% lower vs K2.6 | - | - | - |
The most important row for cost-conscious teams is the last one. K2.7 Code uses roughly 30% fewer thinking tokens than K2.6 while scoring higher. For agentic coding workflows where reasoning tokens dominate the bill, this efficiency gain compounds quickly.
External sourceMarkTechPost - Moonshot AI Releases Kimi K2.7-CodeMarkTechPost, June 12, 2026Cost comparison: Kimi K2.7 vs Claude Fable 5
Price is where K2.7 Code changes the economics. At current API rates, Kimi is roughly five times cheaper on input and more than twelve times cheaper on output than Claude Fable 5. For a team processing 100 million input tokens and 20 million output tokens per month, the difference is tens of thousands of dollars.
| Model | Input / 1M | Output / 1M | Monthly cost* |
|---|---|---|---|
| Claude Fable 5 | $10.00 | $50.00 | $2,000,000 |
| Claude Opus 4.8 | $5.00 | $25.00 | $1,000,000 |
| GPT-5.5 | $5.00 | $30.00 | $1,100,000 |
| Kimi K2.7 Code | $0.95 | $4.00 | $175,000 |
| DeepSeek V4 | $0.50 | $2.00 | $90,000 |
*Estimated monthly cost for 100M input tokens and 20M output tokens, excluding cache discounts. Actual costs depend on provider, context compression, and reasoning-token ratios.
Global availability and access paths
Unlike Fable 5, Kimi K2.7 Code is available globally through multiple channels. The open weights mean you are not locked to a single vendor, and you can run the model on your own infrastructure if data residency or sovereignty is a concern.
- Kimi Code CLI and web interface at kimi.com/code.
- Official Moonshot API at platform.kimi.ai with OpenAI-compatible and Anthropic-compatible endpoints.
- Third-party routers such as OpenRouter, Together AI, Fireworks, and DeepInfra.
- Self-hosted deployment via vLLM, SGLang, or KTransformers using Hugging Face weights.
How to migrate from Claude Code to Kimi K2.7 Code
Moonshot's Anthropic-compatible endpoint makes the migration almost frictionless. You can keep your existing Claude Code, Cline, or RooCode workflow and only change the base URL, auth token, and model name.
| Step | Action |
|---|---|
| 1. Get a Moonshot API key | Create a key at platform.kimi.ai/console/api-keys. |
| 2. Set environment variables | Point ANTHROPIC_BASE_URL to https://api.moonshot.ai/anthropic and ANTHROPIC_MODEL to kimi-k2.7-code. |
| 3. Launch Claude Code | Run claude and verify with /status. |
| 4. Adjust for thinking mode | K2.7 Code always runs with thinking enabled; preserve reasoning_content across multi-turn sessions. |
| 5. Benchmark your workflow | Run a representative coding task and compare cost, latency, and success rate against your previous model. |
Limitations to know before switching
- K2.7 Code is text-only. For image or video inputs, keep using K2.6 or another multimodal model.
- All launch benchmarks are first-party. Independent SWE-bench and Terminal-Bench scores are not yet published.
- Thinking mode cannot be disabled, so very low-latency use cases are not a good fit.
- The model is optimized for coding; for general writing, analysis, or long-form reasoning, K2.6 or Claude Opus 4.8 may still be better.
- Self-hosting the full 595 GB weights requires multiple high-memory GPUs.
“The ban on Fable 5 is a reminder that frontier AI access can change overnight. Open-weight models like Kimi K2.7 Code do not eliminate vendor risk entirely, but they give you the option to self-host, route across providers, and keep shipping even when a single API goes dark.”
Who should use Kimi K2.7 Code?
- Teams outside the US that lost access to Claude Fable and need a high-performance coding alternative.
- Startups and scale-ups optimizing AI spend without dropping below frontier coding capability.
- Engineering teams that already run agentic coding workflows in Claude Code, Cline, or RooCode and want a drop-in replacement.
- Organizations with data-sovereignty requirements that prefer self-hosted or multi-provider inference.
Frequently asked questions
Is Kimi K2.7 Code really cheaper than Claude?
Yes. At official API rates, K2.7 Code is roughly 5× cheaper on input and 12× cheaper on output than Claude Fable 5, and roughly 5× cheaper than Claude Opus 4.8 on output. Third-party providers can reduce the cost further.
Can I run Kimi K2.7 Code inside Claude Code?
Yes. Moonshot provides an Anthropic-compatible endpoint at https://api.moonshot.ai/anthropic. Set ANTHROPIC_MODEL to kimi-k2.7-code and use your Moonshot API key as the auth token.
Is Kimi K2.7 Code better than Claude Opus 4.8 for coding?
On Moonshot's reported agentic coding and tool-use benchmarks, K2.7 Code beats Opus 4.8 on MCPMark Verified. On pure reasoning and general knowledge, Opus 4.8 remains competitive. The best choice depends on whether your workload is coding-heavy and cost-sensitive.
Related readingClaude Fable 5: Capabilities, Specs, Benchmarks & Why the US Locked It DownEngineering insightsNeed help integrating Kimi K2.7 Code into your Proptech or Fintech engineering workflow? Contact SelectCursor and we'll help you benchmark, migrate, and run agentic coding pipelines that stay within your budget and compliance boundaries.
About the author
Ali Amin
Partner · Delivery & engineering
Part of the selectcursor core team. We build lending platforms, property marketplaces and fintech infrastructure for European companies. 200+ engineers placed, 50+ platforms shipped.
Published · Updated
Building something similar?
Our team has shipped 50+ Proptech and Fintech platforms. Book a 25-minute call to discuss your architecture, team structure, or product roadmap.
Book a call