feat(scx-ai): add six SCX.ai models and update Qwen3.8-Max pricing by bhuvan2134686 · Pull Request #45502 · BerriAI/litellm

bhuvan2134686 · 2026-10-09T04:02:21Z

TLDR

Problem this solves:

Six SCX.ai chat models have no cost map row
Tool calls to them fail with UnsupportedParamsError
Their spend is recorded as $0
scx-ai/Qwen3.8-Max still carries SCX.ai's old rates

How it solves it:

Adds cost map rows for the six models, root and backup
Moves scx-ai/Qwen3.8-Max to SCX.ai's current rates
Extends the existing root and backup sync test to the new rows

Intentional product change: scx-ai/Qwen3.8-Max spend is now reported at 1.408 / 4.235 / 0.176 USD per 1M input, output and cached tokens, down from 1.65 / 4.99 / 0.21

Rows

Rates are SCX.ai's API prices in USD per 1M tokens. Context windows come from SCX.ai's model listing. Output limits are the ones the endpoint enforces where it validates max_tokens (GLM-5.2-Fast and GLM-5.3 reject anything above 131072) and the published limit otherwise. Vision, JSON schema, tool calling and prompt caching were each checked with live requests to api.scx.ai on every model except GLM-5.3-Fast (see Caveats)

Model	Input / output / cached	Context	Max output	Vision	JSON schema
`scx-ai/DeepSeek-V4.1-flash`	0.15 / 0.60 / 0.01	1,048,576	384,000	yes	no
`scx-ai/GLM-5.2-Fast`	1.65 / 4.90 / 0.30	1,048,576	131,072	no	yes
`scx-ai/GLM-5.3`	0.93 / 2.92 / 0.164	1,048,576	131,072	no	no
`scx-ai/GLM-5.3-Fast`	2.80 / 8.80 / 0.56	1,048,576	131,072	no	no
`scx-ai/GLM-5.3-Flash`	0.09 / 0.30 / 0.018	1,048,576	131,072	yes	yes
`scx-ai/Kimi-K3`	2.37 / 11.87 / 0.237	1,048,576	1,048,576	yes	yes
`scx-ai/Qwen3.8-Max` (updated)	1.408 / 4.235 / 0.176	1,000,000	131,072	yes	yes

All of them support tool calling, reasoning and prompt caching. GLM-5.3 and DeepSeek-V4.1-flash are marked without JSON schema support because the endpoint does not honour response_format of type json_schema on them, while plain JSON mode works

The page listing these models on docs.litellm.ai is updated in BerriAI/litellm-docs#2204

I work at SCX.ai

User Flow

Before: a developer calling a new SCX.ai model with tools through the proxy gets a 400 and no spend is recorded

They send POST http://localhost:4000/v1/chat/completions with "model": "scx-ai/GLM-5.3" and a tools array
The proxy answers 400 UnsupportedParamsError: scx-ai does not support parameters: ['tools']
The x-litellm-response-cost header reads 0, and GET http://localhost:4000/model/info shows no prices for the model

After: the same request returns a tool call and is charged at SCX.ai's rates

They send POST http://localhost:4000/v1/chat/completions with "model": "scx-ai/GLM-5.3" and a tools array
The proxy answers 200 with a get_weather tool call
The x-litellm-response-cost header reads 0.0002656, which is 160 x 0.93 + 40 x 2.92 USD per 1M tokens, and GET http://localhost:4000/model/info shows those prices

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/unit/<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

Both runs use this config, with SCX_API_KEY and a random LITELLM_MASTER_KEY exported, and LITELLM_LOCAL_MODEL_COST_MAP=True so the proxy reads the cost map from the checkout

model_list:
  - model_name: scx-ai/DeepSeek-V4.1-flash
    litellm_params: {model: scx-ai/DeepSeek-V4.1-flash, api_key: os.environ/SCX_API_KEY}
  - model_name: scx-ai/GLM-5.2-Fast
    litellm_params: {model: scx-ai/GLM-5.2-Fast, api_key: os.environ/SCX_API_KEY}
  - model_name: scx-ai/GLM-5.3
    litellm_params: {model: scx-ai/GLM-5.3, api_key: os.environ/SCX_API_KEY}
  - model_name: scx-ai/GLM-5.3-Fast
    litellm_params: {model: scx-ai/GLM-5.3-Fast, api_key: os.environ/SCX_API_KEY}
  - model_name: scx-ai/GLM-5.3-Flash
    litellm_params: {model: scx-ai/GLM-5.3-Flash, api_key: os.environ/SCX_API_KEY}
  - model_name: scx-ai/Kimi-K3
    litellm_params: {model: scx-ai/Kimi-K3, api_key: os.environ/SCX_API_KEY}
  - model_name: scx-ai/Qwen3.8-Max
    litellm_params: {model: scx-ai/Qwen3.8-Max, api_key: os.environ/SCX_API_KEY}

general_settings:
  master_key: os.environ/LITELLM_MASTER_KEY

python litellm/proxy/proxy_cli.py --config config.yaml --port 4000

Before (`4df5006`)

Model info

Ask the proxy what it knows about each model (model, input, output and cached cost per token, function calling, vision)

curl -s localhost:4000/model/info -H "Authorization: Bearer $LITELLM_MASTER_KEY" \
  | jq -r '.data[] | [.model_name, .model_info.input_cost_per_token, .model_info.output_cost_per_token, .model_info.cache_read_input_token_cost, .model_info.supports_function_calling, .model_info.supports_vision] | @tsv'

Only the existing Qwen3.8-Max row has prices, at the old rates

scx-ai/DeepSeek-V4.1-flash					
scx-ai/GLM-5.2-Fast					
scx-ai/GLM-5.3					
scx-ai/GLM-5.3-Fast					
scx-ai/GLM-5.3-Flash					
scx-ai/Kimi-K3					
scx-ai/Qwen3.8-Max	0.00000165	0.00000499	2.1E-7	true	true

Tool call and spend

Send one tool-calling request per model and print the status, the tool call or error, the token counts and the cost header

h=$(mktemp); b=$(mktemp)
for m in DeepSeek-V4.1-flash GLM-5.2-Fast GLM-5.3 GLM-5.3-Flash Kimi-K3 Qwen3.8-Max; do
  curl -s -D $h -o $b localhost:4000/v1/chat/completions -H "Authorization: Bearer $LITELLM_MASTER_KEY" -H "Content-Type: application/json" \
    -d "{\"model\": \"scx-ai/$m\", \"messages\": [{\"role\": \"user\", \"content\": \"What is the weather in Perth? Use the tool.\"}],
         \"tools\": [{\"type\": \"function\", \"function\": {\"name\": \"get_weather\", \"parameters\": {\"type\": \"object\", \"properties\": {\"city\": {\"type\": \"string\"}}, \"required\": [\"city\"]}}}]}"
  printf '%-28s HTTP %s  ' "scx-ai/$m" "$(head -1 $h | awk '{print $2}')"
  jq -r 'if .error then "error: \(.error.message[0:110])" else "tool_call=\(.choices[0].message.tool_calls[0].function.name)(\(.choices[0].message.tool_calls[0].function.arguments))  prompt=\(.usage.prompt_tokens) completion=\(.usage.completion_tokens)" end' $b
  printf '%-28s x-litellm-response-cost: %s\n' "" "$(grep -i '^x-litellm-response-cost:' $h | awk '{print $2}' | tr -d '\r')"
done

Every new model is rejected with a 400 and costs 0. Qwen3.8-Max is charged at the old rates: 320 x 1.65 + 56 x 4.99 per 1M is 0.00080744

scx-ai/DeepSeek-V4.1-flash   HTTP 400  error: litellm.UnsupportedParamsError: scx-ai does not support parameters: ['tools'], for model=DeepSeek-V4.1-flash. 
                             x-litellm-response-cost: 0
scx-ai/GLM-5.2-Fast          HTTP 400  error: litellm.UnsupportedParamsError: scx-ai does not support parameters: ['tools'], for model=GLM-5.2-Fast. To drop
                             x-litellm-response-cost: 0
scx-ai/GLM-5.3               HTTP 400  error: litellm.UnsupportedParamsError: scx-ai does not support parameters: ['tools'], for model=GLM-5.3. To drop thes
                             x-litellm-response-cost: 0
scx-ai/GLM-5.3-Flash         HTTP 400  error: litellm.UnsupportedParamsError: scx-ai does not support parameters: ['tools'], for model=GLM-5.3-Flash. To dro
                             x-litellm-response-cost: 0
scx-ai/Kimi-K3               HTTP 400  error: litellm.UnsupportedParamsError: scx-ai does not support parameters: ['tools'], for model=Kimi-K3. To drop thes
                             x-litellm-response-cost: 0
scx-ai/Qwen3.8-Max           HTTP 200  tool_call=get_weather({"city": "Perth"})  prompt=320 completion=56
                             x-litellm-response-cost: 0.00080744

After (`5c4d4e8`)

Model info

Same command as Before

curl -s localhost:4000/model/info -H "Authorization: Bearer $LITELLM_MASTER_KEY" \
  | jq -r '.data[] | [.model_name, .model_info.input_cost_per_token, .model_info.output_cost_per_token, .model_info.cache_read_input_token_cost, .model_info.supports_function_calling, .model_info.supports_vision] | @tsv'

Every model reports the rates from the Rows table

scx-ai/DeepSeek-V4.1-flash	1.5E-7	6E-7	1E-8	true	true
scx-ai/GLM-5.2-Fast	0.00000165	0.0000049	3E-7	true	false
scx-ai/GLM-5.3	9.3E-7	0.00000292	1.64E-7	true	false
scx-ai/GLM-5.3-Fast	0.0000028	0.0000088	5.6E-7	true	false
scx-ai/GLM-5.3-Flash	9E-8	3E-7	1.8E-8	true	true
scx-ai/Kimi-K3	0.00000237	0.00001187	2.37E-7	true	true
scx-ai/Qwen3.8-Max	0.000001408	0.000004235	1.76E-7	true	true

Tool call and spend

Same command as Before

h=$(mktemp); b=$(mktemp)
for m in DeepSeek-V4.1-flash GLM-5.2-Fast GLM-5.3 GLM-5.3-Flash Kimi-K3 Qwen3.8-Max; do
  curl -s -D $h -o $b localhost:4000/v1/chat/completions -H "Authorization: Bearer $LITELLM_MASTER_KEY" -H "Content-Type: application/json" \
    -d "{\"model\": \"scx-ai/$m\", \"messages\": [{\"role\": \"user\", \"content\": \"What is the weather in Perth? Use the tool.\"}],
         \"tools\": [{\"type\": \"function\", \"function\": {\"name\": \"get_weather\", \"parameters\": {\"type\": \"object\", \"properties\": {\"city\": {\"type\": \"string\"}}, \"required\": [\"city\"]}}}]}"
  printf '%-28s HTTP %s  ' "scx-ai/$m" "$(head -1 $h | awk '{print $2}')"
  jq -r 'if .error then "error: \(.error.message[0:110])" else "tool_call=\(.choices[0].message.tool_calls[0].function.name)(\(.choices[0].message.tool_calls[0].function.arguments))  prompt=\(.usage.prompt_tokens) completion=\(.usage.completion_tokens)" end' $b
  printf '%-28s x-litellm-response-cost: %s\n' "" "$(grep -i '^x-litellm-response-cost:' $h | awk '{print $2}' | tr -d '\r')"
done

Every request returns a tool call, and each cost header equals prompt x input rate plus completion x output rate, for example GLM-5.3 at 160 x 0.93 + 40 x 2.92 per 1M is 0.0002656 and Qwen3.8-Max at 320 x 1.408 + 49 x 4.235 per 1M is 0.000658075

scx-ai/DeepSeek-V4.1-flash   HTTP 200  tool_call=get_weather({"city": "Perth"})  prompt=289 completion=94
                             x-litellm-response-cost: 9.974999999999999e-05
scx-ai/GLM-5.2-Fast          HTTP 200  tool_call=get_weather({"city": "Perth"})  prompt=160 completion=36
                             x-litellm-response-cost: 0.00044039999999999997
scx-ai/GLM-5.3               HTTP 200  tool_call=get_weather({"city": "Perth"})  prompt=160 completion=40
                             x-litellm-response-cost: 0.0002656
scx-ai/GLM-5.3-Flash         HTTP 200  tool_call=get_weather({"city": "Perth"})  prompt=160 completion=36
                             x-litellm-response-cost: 2.5199999999999996e-05
scx-ai/Kimi-K3               HTTP 200  tool_call=get_weather({"city":"Perth"})  prompt=165 completion=119
                             x-litellm-response-cost: 0.00180358
scx-ai/Qwen3.8-Max           HTTP 200  tool_call=get_weather({"city": "Perth"})  prompt=320 completion=49
                             x-litellm-response-cost: 0.0006580749999999999

Type

New Feature

Caveats (if any)

Low

GLM-5.3-Fast is missing from the live run: the test key lacks access
- Its limits and capability flags match GLM-5.3 and were not checked live

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

Add cost map rows for DeepSeek-V4.1-flash, GLM-5.2-Fast, GLM-5.3, GLM-5.3-Fast, GLM-5.3-Flash and Kimi-K3, and move scx-ai/Qwen3.8-Max to SCX.ai's current rates. Without a row, LiteLLM routes these models but rejects tool calls and records zero spend.

chatgpt-codex-connector · 2026-10-09T04:02:34Z

Codex Review Summary

This comment shows the latest Codex review activity on this pull request.

Review	Status	Commit	Review trigger
📝 Code Review	⚠️ Failed 2026-10-09T04:05:00.375455Z	`5c4d4e8`	PR opened

ℹ️ About Codex in GitHub

Your team has set up Codex to review pull requests in this repo. Reviews are triggered when you

Open a pull request for review
Mark a draft as ready
Comment "@codex review" or "@codex security review".

Codex reacts with 👀 while any review is running, comments if it has suggestions, and reacts with 👍 once all reviews finish with no findings.

greptile-apps · 2026-10-09T04:06:17Z

Confidence Score: 4/5

[Medium impact] Adds new AI models and updates pricing data.

The PR needs behavioral regression tests and live proof for GLM-5.3-Fast before merging to meet repository requirements

Findings

Tests miss tools and spend ▶
One model lacks live proof ▶

Summary

Adds six SCX.ai model rows to both cost maps and lowers scx-ai/Qwen3.8-Max prices

LiteLLM lists prices and capabilities for six more SCX.ai models.
Qwen3.8-Max requests use SCX.ai's updated token prices.

bhuvan2134686 explicitly describes the Qwen price reduction as intentional because the provider's rates changed

_{Reviews (1) · Last reviewed commit: "feat(scx-ai): add six SCX.ai models and ..." · Reviewed by Greptile}

greptile-apps · 2026-10-09T04:06:20Z

+        "scx-ai/DeepSeek-V4.1-flash",
        "scx-ai/GLM-5.2",
+        "scx-ai/GLM-5.2-Fast",
+        "scx-ai/GLM-5.3",
+        "scx-ai/GLM-5.3-Fast",
+        "scx-ai/GLM-5.3-Flash",
+        "scx-ai/Kimi-K3",


Tests miss tools and spend

The expanded test only compares maps. AGENTS.md requires regression coverage before merging; add tool-support and cached-cost tests using loaded rates

Context Used: AGENTS.md (source)

Note: If this suggestion doesn't match your team's coding style, reply to this and let me know. I'll remember it for next time!

greptile-apps · 2026-10-09T04:06:20Z

        "scx-ai/GLM-5.2",
+        "scx-ai/GLM-5.2-Fast",
+        "scx-ai/GLM-5.3",
+        "scx-ai/GLM-5.3-Fast",


One model lacks live proof

GLM-5.3-Fast lacks live QA because the key lacks access. Repository rules require real-provider proof before merging; add proxy results with an authorized key

Rule Used: Flag if the PR has no proof of e2e testing (if e2e testing is applicable here. If it's like an AGENTS.md change or something then probably no need to have e2e test proof) that either doesn't exist, claims that the box that it was trying to run in doe... (source)

Note: If this suggestion doesn't match your team's coding style, reply to this and let me know. I'll remember it for next time!

codspeed · 2026-10-09T04:11:59Z

Merging this PR will not alter performance

✅ 31 untouched benchmarks

_{Comparing SouthernCrossAI:litellm_scx_ai_new_models (5c4d4e8) with main (4df5006)}

codecov · 2026-10-09T04:16:06Z

Codecov Report

✅ All modified and coverable lines are covered by tests.

📢 Thoughts on this report? Let us know!

bhuvan2134686 requested review from kerrylu-berri, mateo-berri and ryan-crabbe-berri as code owners October 9, 2026 04:02

bhuvan2134686 mentioned this pull request Oct 9, 2026

docs(providers): add the new SCX.ai models to the SCX.ai page BerriAI/litellm-docs#2204

Open

greptile-apps Bot reviewed Oct 9, 2026

View reviewed changes

Navigation Menu

Repository navigation

feat(scx-ai): add six SCX.ai models and update Qwen3.8-Max pricing - #45502

feat(scx-ai): add six SCX.ai models and update Qwen3.8-Max pricing#45502
bhuvan2134686 wants to merge 1 commit into
BerriAI:mainfrom
SouthernCrossAI:litellm_scx_ai_new_models

bhuvan2134686 commented Oct 9, 2026

Uh oh!

chatgpt-codex-connector Bot commented Oct 9, 2026 •

edited

Loading

Uh oh!

greptile-apps Bot commented Oct 9, 2026 •

edited

Loading

Uh oh!

greptile-apps Bot Oct 9, 2026

Uh oh!

greptile-apps Bot Oct 9, 2026

Uh oh!

codspeed Bot commented Oct 9, 2026

Uh oh!

codecov Bot commented Oct 9, 2026

Uh oh!

This branch has not been deployed

Reviewers

Assignees

Labels

Projects

Milestone

Development

Uh oh!

1 participant

Conversation

bhuvan2134686 commented Oct 9, 2026

TLDR

Rows

User Flow

Pre-Submission checklist

Delays in PR merge?

Screenshots / Proof of Fix

Before (4df5006)

Model info

Tool call and spend

After (5c4d4e8)

Model info

Tool call and spend

Type

Caveats (if any)

Low

Final Attestation

Uh oh!

chatgpt-codex-connector Bot commented Oct 9, 2026 • edited Loading Uh oh! There was an error while loading. Please reload this page.

Uh oh!

Codex Review Summary

Uh oh!

greptile-apps Bot commented Oct 9, 2026 • edited Loading Uh oh! There was an error while loading. Please reload this page.

Uh oh!

Confidence Score: 4/5

Findings

Summary

Uh oh!

greptile-apps Bot Oct 9, 2026

Choose a reason for hiding this comment

Uh oh!

greptile-apps Bot Oct 9, 2026

Choose a reason for hiding this comment

Uh oh!

codspeed Bot commented Oct 9, 2026

Merging this PR will not alter performance

Uh oh!

codecov Bot commented Oct 9, 2026

Codecov Report

Uh oh!

This branch has not been deployed

Reviewers

Assignees

Labels

Projects

Milestone

Development

Uh oh!

1 participant

Before (`4df5006`)

After (`5c4d4e8`)

chatgpt-codex-connector Bot commented Oct 9, 2026 •

edited

Loading

greptile-apps Bot commented Oct 9, 2026 •

edited

Loading