add moonshot/kimi-k2.6 to model registry

[ishaan-berri]

Relevant Issues / Related PRs

Adds Kimi K2.6 (released April 20, 2026) to the LiteLLM model registry.

Pre-Submission Checklist

• I have added tests in tests/litellm/
• make test-unit passes for the modified tests

Changes

• Added moonshot/kimi-k2.6 to model_prices_and_context_window.json and backup
• Pricing: $0.60/M input, $2.80/M output (from https://platform.kimi.ai/docs/guide/kimi-k2-6-quickstart)
• 262K context window, supports function calling, vision, and video input
• Tests for model registry entry

Usage

import litellm

response = litellm.completion(
    model="moonshot/kimi-k2.6",
    messages=[{"role": "user", "content": "Hello!"}],
)

[CLAassistant]

Thank you for your submission! We really appreciate it. Like many open source projects, we ask that you sign our Contributor License Agreement before we can accept your contribution.
_{You have signed the CLA already but the status is still pending? Let us recheck it.}

[codspeed-hq]

Merging this PR will not alter performance

✅ 16 untouched benchmarks

---

_{Comparing worktree-golden-yawning-hammock (240c061) with main (26fcbc9)}

[greptile-apps]

Greptile Summary

This PR adds moonshot/kimi-k2.6 to the LiteLLM model registry with pricing ($0.60/M input, $2.80/M output), a 262K context window, and capability flags for function calling, vision, and video input. Both JSON files are kept in sync and the accompanying tests correctly read from the in-memory model cost map without making real network calls.

Confidence Score: 5/5

Safe to merge; all findings are P2 clarification items on optional registry fields

The core change is a well-formed JSON model registry addition with correct pricing math and a thorough test suite. The two open questions (missing cache_read_input_token_cost and supports_reasoning compared to kimi-k2.5) are P2 — they won't break routing but could silently under-report capabilities or miss a caching discount. No logic errors, security issues, or test-integrity problems detected.

model_prices_and_context_window.json — verify cache pricing and reasoning support against official kimi-k2.6 docs

Important Files Changed

Filename	Overview
model_prices_and_context_window.json	Adds moonshot/kimi-k2.6 entry; missing cache_read_input_token_cost and supports_reasoning compared to sibling kimi-k2.5 model
litellm/model_prices_and_context_window_backup.json	Backup JSON mirrors main JSON — same entry with same omissions as above
tests/test_litellm/llms/moonshot/test_moonshot_chat_transformation.py	Adds TestKimiK26ModelRegistry with 5 unit tests verifying pricing, context window, capabilities, and provider — no real network calls, complies with test rules

Flowchart

%%{init: {'theme': 'neutral'}}%%
flowchart TD
    A[litellm.completion\nmodel='moonshot/kimi-k2.6'] --> B[model_prices_and_context_window.json\nlookup]
    B --> C{Entry found?}
    C -- Yes --> D[Resolve litellm_provider\n= 'moonshot']
    D --> E[Apply pricing\n$0.60/M input · $2.80/M output]
    D --> F[Apply limits\n262K ctx · 262K output]
    D --> G[Enable capabilities\nfunction calling · vision · video]
    E & F & G --> H[Route to Moonshot API]
    C -- No --> I[KeyError / fallback]

Loading

_{Reviews (1): Last reviewed commit: "add tests for moonshot/kimi-k2.6 model r..." | Re-trigger Greptile}

[greptile-apps]

Missing cache_read_input_token_cost

Every other moonshot/kimi-k2* model in the registry includes cache_read_input_token_cost (e.g., kimi-k2.5 has 1e-07, kimi-k2-0905-preview has 1.5e-07). If prompt caching is supported for kimi-k2.6, omitting this field means LiteLLM will never apply the discounted rate, silently over-billing users on cached tokens. If the model genuinely doesn't support caching yet, a brief source note explaining that would help future readers.

[greptile-apps]

supports_reasoning not set — verify vs kimi-k2.5

moonshot/kimi-k2.5 (the predecessor in the registry) has "supports_reasoning": true, but this entry omits it. If kimi-k2.6 also surfaces a <thinking> block or extended reasoning output, leaving this flag out will prevent LiteLLM from routing reasoning-aware handling. Please verify against the official docs and add the flag if applicable.

[codecov]

Codecov Report

✅ All modified and coverable lines are covered by tests.

📢 Thoughts on this report? Let us know!

[ishaan-berri]

Refiling with correct base branch (litellm_internal_staging) and branch naming convention.