fix(model_prices): registry audit 2026-10-09, together qwen3.7-max price, azure kimi-k2.7-code retirement

[devin-ai-integration]

TLDR

Problem this solves:

• together_ai/Qwen/Qwen3.7-Max is priced at $1.50/$4.50 per 1M (cached $0.30), but Together now lists it at $2.50/$7.50 (cached $0.50), so LiteLLM underbills it by 40%
• azure_ai/kimi-k2.7-code carries deprecation_date: 2027-04-16 (the Kimi-K2.6 date), but Azure's retirement schedule lists Kimi-K2.7-Code (2026-06-12, Preview) as retiring 2026-10-03

How it solves it:

• Field-level edits to those two entries in model_prices_and_context_window.json and the backup copy, nothing else

This is the rolling registry PR for the 2026-10-09 audit run (no open Devin registry-only PR existed to roll into)

Relevant issues

Supports #26900 (registry deprecation_date drives deprecation surfacing)

User Flow

Before: a team calling together_ai/Qwen/Qwen3.7-Max through the proxy is billed at the old rate, and the Kimi K2.7 Code retirement is reported six months late

1. Admin adds together_ai/Qwen/Qwen3.7-Max to config.yaml and restarts the proxy
2. Developer streams a completion: 16 prompt tokens, 320 completion tokens
3. Spend for the request is $0.001464 (16 x $1.50/M + 320 x $4.50/M)

After: the same call is billed at Together's listed price

1. Admin adds together_ai/Qwen/Qwen3.7-Max to config.yaml and restarts the proxy
2. Developer streams a completion: 16 prompt tokens, 86 completion tokens
3. Spend for the request is $0.000685 (16 x $2.50/M + 86 x $7.50/M)

Sources (one per changed value)

Entry	Field	Old	New	Source
together_ai/Qwen/Qwen3.7-Max	input_cost_per_token	1.5e-06	2.5e-06	Together models API GET https://api.together.xyz/v1/models (pricing.input: 2.5), read 2026-10-09; https://www.together.ai/pricing
together_ai/Qwen/Qwen3.7-Max	output_cost_per_token	4.5e-06	7.5e-06	same (pricing.output: 7.5)
together_ai/Qwen/Qwen3.7-Max	cache_read_input_token_cost	3e-07	5e-07	same (pricing.cached_input: 0.5)
azure_ai/kimi-k2.7-code	deprecation_date	2027-04-16	2026-10-03	https://learn.microsoft.com/en-us/azure/ai-foundry/concepts/model-lifecycle-retirement (raw HTML row: Kimi-K2.7-Code / 2026-06-12 / Preview / 2026-10-03)

Absorbed from other open PRs

#45533 (together_ai registry sync): the Qwen3.7-Max price change, re-verified against the Together models API above. Its other change, deleting cache_read_input_token_cost from together_ai/together/Tev1-4B-experimental, is dropped: the API reports cached_input: 0 for that model, which does not establish that the existing 4.2e-08 row is wrong

Dropped as unverifiable or out of scope

• feat(scx-ai): add six SCX.ai models and update Qwen3.8-Max pricing #45502 (SCX): the PR's prices are the temporary sale prices on https://scx.ai/pricing. Each model's own page lists the regular price (Qwen3.8-Max $1.65/$0.21/$4.99, which is what main already has; Kimi-K3 $2.83/$14.13 and GLM-5.3 $0.98/$3.08, not the PR's $2.37/$11.87 and $0.93/$2.92). GLM-5.2-Fast and GLM-5.3-Fast have no model page. Nothing absorbed
• feat(model_prices): add deepseek-v3 to model price maps (#43660) #43753 (deepseek-v3): DeepSeek's API exposes deepseek-chat/deepseek-reasoner, not deepseek-v3, and deepseek/deepseek-v3 already exists with the same prices
• feat(models): add kimi-k2.5 to model cost map and backup #43879 (kimi-k2.5): duplicate of the existing moonshot/kimi-k2.5 row
• feat(models): sync together_ai model registry #41775: removes Together V4 deprecation dates; Together's deprecations page still does not support removal
• fix(cost): bill OpenRouter and Perplexity v1.1 decisions models #45476, chore(github_copilot): sync model catalog with Copilot's served models #45386, fix(bedrock-invoke): use native output_format for Claude 4.5+ models that support it #29898: code or behavior changes (or catalog syncs with code), left open

Checked with no change needed

Anthropic deprecations page (all dated models match or are already removed), Groq deprecations page, Together deprecations page (serverless removal table; the 2026-07-29 rows are the fine-tuning table and were not applied), Azure Foundry retirement schedule (other mismatches are azure/sora-2, deliberately set to the later Models API date on main, and azure/gpt-realtime-mini-2025-10-06, which appears in two tables with different dates, so it was left alone). ci_cd/cost_map_pins.json is not present on main, and no capability flags change in this PR

Pre-Submission checklist

• I have added meaningful tests (registry data only; existing schema and alias tests cover it)
• The handful of test files covering my change pass locally
• My PR passes all required CI/CD checks
• 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

Screenshots / Proof of Fix

Local checks on 894ac03:

python3 ci_cd/check_files_match.py                          -> Passed! Files ... match.
uv run --no-sync python ci_cd/generate_model_prices_schema.py --check -> in sync and validates
duplicate-key scan of both JSON files                       -> none
LITELLM_LOCAL_MODEL_COST_MAP=True uv run --no-sync pytest -q tests/unit/test_model_prices_schema.py tests/unit/test_model_cost_aliases.py -> 86 passed
uv run --no-sync python ci_cd/cost_map_guard.py --base origin/main --head HEAD --head-ref litellm_registry_audit_2026_10_09 -> passed

Live call, real Together key, LITELLM_LOCAL_MODEL_COST_MAP=True, Before on main 63a4f3f (port 4011), After on this branch (port 4012). Qwen3.7-Max only accepts streaming, so the request streams with include_usage:

curl -s -i -N http://localhost:4012/v1/chat/completions -H "Authorization: Bearer sk-1234" -H "Content-Type: application/json" \
  -d '{"model":"qwen37max","messages":[{"role":"user","content":"Reply with the single word ok"}],"max_tokens":20,"stream":true,"stream_options":{"include_usage":true}}'

Before, final usage chunk: "completion_tokens":320,"prompt_tokens":16,...,"cost":0.001464 (old rates)
After, final usage chunk: "completion_tokens":86,"prompt_tokens":16,...,"cost":0.000685 (16 x 2.5e-06 + 86 x 7.5e-06 = 0.000685)

Type

Bug Fix

Caveats

Streaming responses report x-litellm-response-cost: 0.0 in headers on both legs because headers flush before usage is known, so the proof reads the cost from the final usage chunk instead

Link to Devin session: https://app.devin.ai/sessions/81e2e7f1f1934b738b6f459898f1b721
Open in Devin Desktop: https://app.devin.ai/desktop/session/81e2e7f1f1934b738b6f459898f1b721?variant=devin

[devin-ai-integration]

I'll fix CI failures and address comments from users with write access. I'll skip comments containing "(aside)".

• Disable automatic comment, CI, and merge conflict monitoring

[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.}

[greptile-apps]

Confidence Score: 5/5

[Critical impact] This PR appears safe to merge, with no actionable issues found in the registry edits

Summary

Updates together_ai/Qwen/Qwen3.7-Max to $2.50/M input tokens, $7.50/M output tokens, and $0.50/M cached input tokens. Moves azure_ai/kimi-k2.7-code retirement to October 3, 2026

• Qwen3.7-Max requests use Together’s updated token prices.
• Kimi-K2.7-Code’s registry date matches its Azure retirement schedule.

_{Reviews (1) · Last reviewed commit: "fix(model_prices): registry audit 2026-1..." · Reviewed by Greptile}

[codspeed]

Merging this PR will not alter performance

✅ 31 untouched benchmarks

---

_{Comparing litellm_registry_audit_2026_10_09 (894ac03) with main (63a4f3f)}

[codecov]

Codecov Report

✅ All modified and coverable lines are covered by tests.

📢 Thoughts on this report? Let us know!