Kimi K3 vs K2.7 Code: Price, Context, Coding, and Upgrade Guide

Kimi K3 is the new general flagship, but K2.7 Code remains much cheaper for coding workloads that fit its limits. Moving every K2.7 request to K3 would raise the global output price from $4 to $15 per million tokens.

The sensible upgrade is selective: use K3 for tasks that benefit from its 1M context, very long completion allowance, stronger agent architecture, or advanced structured output.

Global API Comparison

	Kimi K2.7 Code	Kimi K3
Cache-miss input / 1M	$0.95	$3.00
Cached input / 1M	$0.19	$0.30
Output / 1M	$4.00	$15.00
Context	256K class	1M
Maximum output in canonical data	32K	Up to 1,048,576
Positioning	Coding-specialized	General flagship for coding, research, and agents
Input	Text, image, video	Text, image, video
Reasoning	Thinking-only	Always reasoning; API currently max

K3 costs 3.16 times more for cache-miss input and 3.75 times more for output. Cached-input pricing is closer: K3 costs $0.30/M versus $0.19/M for K2.7 Code.

China API Comparison

	Kimi K2.7 Code	Kimi K3
Cache-miss input / 1M	¥6.50	¥20.00
Cached input / 1M	¥1.30	¥2.00
Output / 1M	¥27.00	¥100.00

Keep currency and region explicit in budget documents. The global USD table and China RMB table are separate official price lists, not exchange-rate conversions.

Example Task Cost

For 2M input and 500K output tokens on global API rates, before caching:

Model	Cost
Kimi K2.7 Code	2 × $0.95 + 0.5 × $4 = $3.90
Kimi K3	2 × $3 + 0.5 × $15 = $13.50

For 100 tasks with this token shape, that is $390 versus $1,350. K3 should therefore earn its place through task success, not novelty.

Upgrade to K3 When

• Repository or research context regularly exceeds the K2.7 window.
• A workflow needs output substantially above 32K.
• Strict JSON Schema, Partial Mode, or dynamic tool loading simplifies your agent.
• Difficult multi-step tasks cause repeated K2.7 retries or human repair.
• You need a single flagship across coding, research, image, and video tasks.

Stay on K2.7 Code When

• The workload is mostly scoped code editing within 256K context.
• Your application emits large amounts of code and output price dominates.
• K2.7 already passes the required tests on the first attempt.
• You need a cheap default route before escalating hard cases.

An existing K2.7 deployment does not need a flag-day migration. Add K3 as an escalation route, replay failed K2.7 tasks against it, and compare incremental success against the $9.60 difference in the example above.

Migration Checklist

1. Add kimi-k3 as a separate model configuration; do not silently replace the old ID.
2. Remove the K2.x thinking object and use the top-level reasoning_effort: "max" field.
3. Send the complete assistant message back in multi-turn and tool loops, including reasoning_content, content, and tool_calls.
4. Raise completion limits only for workflows that need them.
5. Keep stable instructions and tool schemas at the beginning for automatic cache matching.
6. Validate tool calls and structured output against your own schemas.
7. Track cache hits, retries, tests passed, output length, and human repair time.
8. Promote K3 to the default only where cost per completed task improves.

Read the K3 API pricing guide for setup details, or compare K3 with GPT-5.6 Sol and GLM-5.2/DeepSeek V4.

Official sources checked July 19, 2026: Kimi K3 launch, K3 API quickstart, global API pricing, China K3 pricing, and Kimi Code release notes.