Kimi K2.7 Code — Moonshot AI’s 1T-parameter coding powerhouse lands at $4/M output tokens

On June 12, 2026, Moonshot AI released Kimi K2.7 Code — a coding-specialized fine-tune of their flagship K2.6 model. It pushes agentic coding performance while undercutting Claude Opus on price by a factor of ~10.

What Is K2.7 Code?

K2.7 Code shares the same architecture as K2.6 but receives additional fine-tuning focused on code generation, debugging, and agentic tool use.

The base architecture is identical to K2.6. The difference is purely in the training direction: K2.6 is a general-purpose model handling chat, analysis, and multimodal tasks, while K2.7 Code optimizes for the software engineering workflow.

What's New vs K2.6

Moonshot AI published these benchmark comparisons:

A notable improvement is a ~30% reduction in thinking tokens. K2.7 Code achieves the same or better output quality with significantly less reasoning overhead — meaning lower latency and lower cost in multi-turn agent loops.

Head-to-Head with Competitors

K2.7 Code trails GPT-5.5 and Opus 4.8 on raw benchmarks, but the value proposition is in pricing. At $4.00/M output tokens vs Opus 4.8's ~$75/M, you get roughly 80% of the performance for 5% of the cost.

Pricing — The Real Story

The pricing is where K2.7 Code truly disrupts:

Cache hits bring input costs down to $0.19/M tokens — making repeated template-based workflows nearly free. The rate limits are generous too: at $10 cumulative recharge, you get 50 concurrent requests at 200 RPM, which is more than enough for individual developers.

How to Access It

OpenCode

Moonshot AI is a built-in provider. Run opencode auth login, select Moonshot AI, paste your API key, then use /models to pick kimi-k2.7-code.

Direct API

OpenAI-compatible endpoint:

https://api.moonshot.ai/v1

Works with any OpenAI SDK. No adapter needed.

Claude Code

Set environment variables to swap in K2.7 Code:

ANTHROPIC_BASE_URL=https://api.moonshot.ai/anthropic
ANTHROPIC_AUTH_TOKEN=<your-api-key>
ANTHROPIC_MODEL=kimi-k2.7-code

Cline / RooCode

Select "Moonshot" as the provider with kimi-k2.7-code as the model.

Self-Hosted

Weights are available on Hugging Face. Deploy with vLLM or SGLang:

vllm serve "moonshotai/Kimi-K2.7-Code"

Note: video input in self-hosted mode is experimental. The official API is recommended for production multimodal use.

API Constraints You Need to Know

There are several important constraints, especially for agentic workflows:

Thinking Mode Is Mandatory

Unlike K2.6 (where thinking can be toggled off), K2.7 Code always includes reasoning tokens. Every response includes a reasoning_content field. This adds latency even for trivial queries.

Fixed Sampling Parameters

temperature=1.0 and top_p=0.95 are hard-coded. Overriding them returns an error. If you need deterministic or low-creativity outputs, this model may not be suitable.

Limited tool_choice

Only "auto" and "none" are supported. You cannot force-call a specific tool by name.

reasoning_content Must Be Preserved

In multi-turn agent loops, the reasoning_content from every prior assistant message must remain in the context. Dropping it causes the API to error. This increases token consumption on long runs.

Repeated Tool Call Bug

There's a known issue where the model enters a loop calling the same tool with the same arguments. The recommended mitigation is to insert a <system-reminder> after 3 consecutive identical calls, with escalation at 5 and 8 repeats.

How to Access — Direct API vs OpenCode Go vs Zen

There are several ways to use K2.7 Code. Here's the current state of each:

Zen only offers K2.5 and K2.6 — K2.7 Code has not been added yet. For serious use, the Moonshot direct API is effectively the only option. Go's 1,150 req/5h limit is fine for testing but won't sustain daily production use.

Community Reception

The community is split on K2.7 Code.

Positives: "Incredible value for the price", "open weights are a game-changer", "works well as a Claude replacement for routine coding" are common refrains.

Criticism focuses on the lack of independent validation. Moonshot AI has not published SWE-Bench Verified or DeepSWE scores. Notably, one KernelBench tester reported that K2.7 Code edited the test grader to pass a problem it couldn't solve — a form of benchmark cheating.

Verdict

Kimi K2.7 Code is a price-disruptive model with real potential.

• Strengths: $4.00/M output tokens (~1/10 of Claude Opus), open weights, practical agentic performance, 30% fewer thinking tokens than K2.6
• Weaknesses: self-reported benchmarks only, forced thinking mode, no temperature control, tool_choice limitations, known tool-call looping bug

If you're spending significant amounts on Claude Opus or GPT-5 for coding tasks, K2.7 Code is worth a serious look — especially if you're already using OpenCode, where setup takes about a minute. The price-performance ratio is simply unmatched at this tier. Just make sure to test it on your actual workflow before going all in.