feat(dashscope): add kimi-k3, MiniMax-M3, and deepseek-v4-pro-0813 to the cost map

[shivamrawat1]

TLDR

Problem this solves:

• DashScope's Kimi-K3, MiniMax-M3, deepseek-v4-pro-0813 missing from cost map
• Calls to them log $0 spend
• Model Hub shows no context limits for them
• Dual-mode models bill thinking-mode input at the cheaper rate

How it solves it:

• Adds all three with prices, limits, capability flags
• kimi-k3 added under both ids Model Studio documents
• New thinking-rate keys bill thinking-mode requests at their real prices

User Flow

Before: a platform admin routing DashScope's newest models through LiteLLM gets no spend tracking and no context limits

1. They add model: dashscope/kimi-k3 (or dashscope/deepseek-v4-pro-0813, dashscope/MiniMax/MiniMax-M3) to their config and start the proxy
2. They GET https://litellm-domain/v1/model/info and the entries come back with zero prices and null max input and output tokens
3. They POST https://litellm-domain/v1/chat/completions against the model; it answers, but the response carries x-litellm-response-cost-original: 0.0 and https://litellm-domain/ui/?page=logs shows the request at $0 spend
4. The Model Hub page shows "- / -" for the model's token limits, the same symptom this customer thread opened with

After: the same models bill correctly and expose their real limits

1. They add the same models to their config and start the proxy
2. They GET https://litellm-domain/v1/model/info and see input, output, and cache read prices plus real max input and output tokens
3. The same POST https://litellm-domain/v1/chat/completions returns x-litellm-response-cost matching the docs' rates to the digit; a deepseek-v4-pro-0813 response carrying reasoning tokens bills every token, input included, at the doubled thinking-mode prices from the pricing page
4. The Model Hub shows the real context limits

Relevant issues

Linear ticket

Pre-Submission checklist

Please complete all items before asking a LiteLLM maintainer to review your PR

• I have added meaningful tests
• The handful of test files covering my change pass locally, e.g. uv run pytest tests/test_litellm/<your_test_file>.py -v. Leave the suites (make test-unit-*, make test-unit) to CI: it finishes in ~15 minutes where a laptop takes an hour or more
• My PR passes all required CI/CD checks (e.g., lint, schema.d.ts sync check, etc.)
• My PR's scope is as isolated as possible; it only solves 1 specific problem
• I have received a Greptile Confidence Score of at least 4/5 before requesting a maintainer review (Greptile reviews automatically once the PR is opened; only comment @greptileai to re-request a review after pushing changes)

Delays in PR merge?

If you're seeing a delay in your PR being merged, ping the LiteLLM Team on Slack (#pr-review).

Screenshots / Proof of Fix

Captured live end to end with no mocks: real billed DashScope calls through a local proxy running LITELLM_LOCAL_MODEL_COST_MAP=True, hitting https://dashscope-intl.aliyuncs.com/compatible-mode/v1 with a funded key. Shared config for both runs:

model_list:
  - model_name: kimi-k3
    litellm_params:
      model: dashscope/kimi-k3
      api_key: os.environ/DASHSCOPE_API_KEY
      api_base: https://dashscope-intl.aliyuncs.com/compatible-mode/v1
  - model_name: deepseek-v4-pro-0813
    litellm_params:
      model: dashscope/deepseek-v4-pro-0813
      api_key: os.environ/DASHSCOPE_API_KEY
      api_base: https://dashscope-intl.aliyuncs.com/compatible-mode/v1

Each case ran the same two commands:

curl -s http://localhost:4010/v1/model/info -H "Authorization: Bearer sk-1234" | jq -r \
    '.data[] | [.model_name, .model_info.input_cost_per_token, .model_info.input_cost_per_token_thinking,
                .model_info.output_cost_per_token, .model_info.output_cost_per_token_thinking] | @tsv'

curl -s http://localhost:4010/v1/chat/completions -H "Authorization: Bearer sk-1234" \
    -H "Content-Type: application/json" -D - \
    -d '{"model": "<model>", "messages": [{"role": "user", "content": "Reply with one word: hello"}], "max_tokens": 500}'

Before (a28f075)

kimi-k3

1. /v1/model/info row: kimi-k3 0 0 null null
2. Completion answered "hello" with usage: {prompt_tokens: 91, completion_tokens: 64, reasoning_tokens: 48} but the headers carried x-litellm-response-cost-original: 0.0, so the request logged at $0

deepseek-v4-pro-0813

1. /v1/model/info row: deepseek-v4-pro-0813 0 0 null null
2. Completion answered "hello" with usage: {prompt_tokens: 89, completion_tokens: 26, reasoning_tokens: 24} and x-litellm-response-cost-original: 0.0

After (a481ccf)

kimi-k3

1. /v1/model/info row: kimi-k3 0.000002827 - 0.000014133 - (single-rate model, no thinking keys)
2. Completion answered "Hello" with usage: {prompt_tokens: 91, completion_tokens: 52, reasoning_tokens: 35} and headers

x-litellm-response-cost: 0.0009921729999999998
x-litellm-response-cost-input: 0.000257257
x-litellm-response-cost-output: 0.000734916

3. Matches the entry exactly: 91 x 2.827e-06 = $0.000257257 in, 52 x 1.4133e-05 = $0.000734916 out

deepseek-v4-pro-0813

1. /v1/model/info row: deepseek-v4-pro-0813 6.36E-7 0.000001272 0.000001908 0.000003816
2. Completion answered "hello" with usage: {prompt_tokens: 89, completion_tokens: 25, reasoning_tokens: 23} and headers

x-litellm-response-cost: 0.000208608
x-litellm-response-cost-input: 0.000113208
x-litellm-response-cost-output: 9.54e-05
x-litellm-response-cost-reasoning: 8.7768e-05

3. The response carried reasoning tokens, so every token billed at the thinking rates: 89 x 1.272e-06 = $0.000113208 in (double the non-thinking rate) and 25 x 3.816e-06 = $0.0000954 out, matching the pricing page's CNY 9 / CNY 27 thinking column

Type

🆕 New Feature

Caveats (if any)

• MiniMax-M3 is Beijing-only, so not live verifiable from an international account: The model MiniMax/MiniMax-M3 does not exist or you do not have access to it on dashscope-intl. Entry follows its Model Studio doc page, CNY converted at the 0.14133 USD rate the docs themselves use for kimi-k3 and deepseek-v4-pro-0813; its max output (128000) comes from the vendor since Model Studio lists none
• kimi-k3 is registered under both dashscope/kimi-k3 and dashscope/kimi/kimi-k3: the plain id serves completions on dashscope-intl while the namespaced one (the id the docs' nav uses) is recognized but gated behind product activation, so both are real ids a customer may configure
• Thinking mode is detected from reasoning tokens in the usage, the only signal visible at billing time; a thinking-mode request that emits zero reasoning tokens would bill at base rates

Final Attestation

• The tests check the right things, including the edge cases, and regressions in the respective real-world customer use-cases are not possible after this PR

[greptile-apps]

Greptile Summary

The PR adds DashScope pricing and model metadata for Kimi-K3, MiniMax-M3, and deepseek-v4-pro-0813, plus thinking-mode rate selection

• Adds both documented Kimi-K3 identifiers and the MiniMax-M3 and deepseek-v4-pro-0813 catalog entries
• Extends model-info types and lookup propagation with thinking-mode input, cache-read, and output rates
• Selects alternate DashScope rates when returned usage contains reasoning tokens

Confidence Score: 5/5

The PR appears safe to merge

No blocking failure remains

Important Files Changed

Filename	Overview
litellm/llms/dashscope/cost_calculator.py	Adds immutable thinking-rate overrides selected when usage reports reasoning tokens
model_prices_and_context_window.json	Registers the three new DashScope models, including both Kimi-K3 identifiers, prices, limits, and capability flags
litellm/model_prices_and_context_window_backup.json	Mirrors the primary cost-map additions
litellm/types/utils.py	Adds typed fields for DashScope thinking-mode input, cache-read, and output rates
litellm/utils.py	Propagates the new thinking-rate fields through model-info lookup
tests/test_litellm/llms/dashscope/test_dashscope_cost_calculator.py	Adds calculator behavior tests for thinking, non-thinking, and explicit-zero rates

_{Reviews (2): Last reviewed commit: "feat(dashscope): bill thinking-mode requ..." | Re-trigger Greptile}

[veria-ai]

PR overview

All previously flagged issues have been addressed. No open security concerns remain on this pull request.

Security review

No open security issues remain on this pull request.

Fixed/addressed: 1 · PR risk: 0/10

[codecov]

Codecov Report

✅ All modified and coverable lines are covered by tests.

📢 Thoughts on this report? Let us know!

[shivamrawat1]

@greptileai

[shivamrawat1]

@greptile review again!

[shivamrawat1]

@BugBot

[codspeed-hq]

Merging this PR will not alter performance

✅ 31 untouched benchmarks

---

_{Comparing litellm_dashscope_kimi_minimax_dsv4pro0813 (a481ccf) with litellm_internal_staging (fde3075)¹}

Footnotes

1. No successful run was found on litellm_internal_staging (8f6de53) during the generation of this report, so fde3075 was used instead as the comparison base. There might be some changes unrelated to this pull request in this report. ↩