chore(alibaba): update Qwen model pricing per Alibaba Cloud

[vicovaro]

Summary

Sync Qwen model pricing with Alibaba Cloud Model Studio official list prices
(source: https://www.alibabacloud.com/help/en/model-studio/model-pricing —
"Model inference pricing", page updated 2026-08-12 03:12 UTC; region values
verified per-tab on the same page). Only permanent, lower changes applied.

Rule applied: repo price > official list price, and the change is
permanent (no "Limited-time X% off" promo). ✅ = change applied.
❌ = deliberately NOT changed (promo trap / already lower / deactivated / not in catalog).

✅ Changes applied

qwen3-max — flat top-tier → official base + tiers

	Input $/1M	Output $/1M
Before (repo)	3.00	15.00	❌ stale flat
After (repo)	1.20	6.00	✅
Tiers added	2.40 / 12.00 (32-128K), 3.00 / 15.00 (128-256K)		✅
Source	Alibaba list price, qwen3-max row (≤32K tier), page 2026-08-12		✅

qwen3-coder-plus — flat top-tier → official base + tiers

	Input $/1M	Output $/1M
Before (repo)	6.00	60.00	❌ charged top tier for every request
After (repo)	1.00	5.00	✅
Tiers added	1.80 / 9.00 (32-128K), 3.00 / 15.00 (128-256K), 6.00 / 60.00 (256K-1M)		✅
Source	Alibaba list price, qwen3-coder-plus row (≤32K tier), page 2026-08-12		✅

qwen3.7-flash — add non-SG regional overrides

Region	Before $/1M (in/out)	After $/1M (in/out)
Singapore	0.030 / 0.130	0.030 / 0.130 (unchanged)	✅ already correct
EU-Frankfurt	0.030 / 0.130	0.028 / 0.110	✅
US-Virginia	0.030 / 0.130	0.028 / 0.110	✅
CN-Beijing	0.030 / 0.130	0.028 / 0.110	✅
Tiers (non-SG)	—	0.083/0.330 (32-256K), 0.165/0.660 (256K-1M)	✅
Source	Alibaba region tabs: Frankfurt/US/Beijing rows, qwen3.7-flash (Global id), page 2026-08-12		✅

qwen3.7-plus — add missing us-virginia + cn-beijing overrides

Region	Before $/1M (in/out)	After $/1M (in/out)
Singapore	0.40 / 1.60	0.40 / 1.60 (unchanged)	✅ already correct
EU-Frankfurt	0.276 / 1.101	0.276 / 1.101 (unchanged)	✅ already correct
US-Virginia	0.40 / 1.60	0.276 / 1.101	✅
CN-Beijing	0.40 / 1.60	0.276 / 1.101	✅
Tier >256K (non-SG)	—	0.826 / 3.301	✅
Source	Alibaba region tabs: US/Beijing rows, qwen3.7-plus (Global id), page 2026-08-12		✅

qwen3.6-plus — add missing us-virginia + cn-beijing overrides

Region	Before $/1M (in/out)	After $/1M (in/out)
Singapore	0.50 / 3.00	0.50 / 3.00 (unchanged)	✅ already correct
EU-Frankfurt	0.276 / 1.651	0.276 / 1.651 (unchanged)	✅ already correct
US-Virginia	0.50 / 3.00	0.276 / 1.651	✅
CN-Beijing	0.50 / 3.00	0.276 / 1.651	✅
Tier >256K (non-SG)	—	1.101 / 6.602	✅
Source	Alibaba region tabs: US/Beijing rows, qwen3.6-plus (Global id), page 2026-08-12		✅

qwen3.6-flash — add missing eu-frankfurt override

Region	Before $/1M (in/out)	After $/1M (in/out)
Singapore	0.25 / 1.50	0.25 / 1.50 (unchanged)	✅ already correct
EU-Frankfurt	0.25 / 1.50	0.165 / 0.99	✅
US-Virginia	0.165 / 0.99	0.165 / 0.99 (unchanged)	✅ already correct
CN-Beijing	0.165 / 0.99	0.165 / 0.99 (unchanged)	✅ already correct
Tier >256K (eu-frankfurt)	—	0.66 / 3.961	✅
Source	Alibaba region tab: Frankfurt row, qwen3.6-flash (Global id), page 2026-08-12		✅

All overrides carry cache prices at the repo's existing 20%/10%/125%
(cachedInput/cacheRead/cacheWrite) pattern.

❌ Not changed (deliberate)

• qwen3.7-max / qwen3.7-plus (Singapore) — Alibaba shows "Limited-time 50%/20% off"
  promos; the gateway already carries the permanent list price. Promo prices
  intentionally NOT applied. ❌
• qwen3.6-35b-a3b — repo $0.248/$1.485 is already BELOW official $0.375/$2.25;
  lower-only rule → no raise. ❌
• Deactivated entries (qwen3-next-80b-a3b-, qwen3-vl-235b-a22b-, qwen-vl-plus)
  — deactivatedAt in the past, zero billing impact. ❌
• qwen3.7-plus-us / qwen3.6-flash-us — US-pinned SKUs are separate model ids
  (still $0.40 / $0.25 in US) NOT used by the gateway; do not confuse with the
  Global ids this PR fixes. ❌
• qwen3.5-plus — non-SG drops exist ($0.115/$0.688) but the model is not in
  the repo catalog. ❌

How regional pricing works here (DashScope multi-region)

Alibaba Cloud Model Studio is a separate OpenAI-compatible endpoint per
region, each with its own host, list price, and API key. The gateway models
this with three layers; this PR only touches the price-data layer:

Layer	File	Role
Region endpoints	packages/models/src/providers.ts regionConfig	region ids + endpointMap (singapore → dashscope-intl.aliyuncs.com, us-virginia → dashscope-us.aliyuncs.com, cn-beijing → dashscope.aliyuncs.com, eu-frankfurt → workspace-dedicated {workspaceId}.eu-central-1.maas.aliyuncs.com with trial.eu-central-1… fallback)
Per-region prices	packages/models/src/models/alibaba.ts regions[]	this PR — each region entry overrides inputPrice/outputPrice/cache and may carry its own pricingTiers
Region expansion	packages/models/src/region-helpers.ts expandProviderRegions	regions[] → flat entries ([root, singapore, eu-frankfurt, us-virginia, cn-beijing]); externalId stays constant, region is the disambiguator
Endpoint resolution	packages/actions/src/get-provider-endpoint.ts	endpointMap[region] → resolveWorkspaceScopedEndpoint() fills {workspaceId} for Frankfurt → appends /compatible-mode
Cheapest-region routing	apps/gateway/src/chat/chat.ts (PR #3533/#3540)	prices the expanded entries, picks cheapest eligible region per provider

Concretely: with this PR's overrides in place, expandProviderRegions exposes
the lower non-Singapore list prices, cheapest-region routing sends the request
to the winning region's real DashScope host, and billing matches the served
endpoint. No endpoint, credential, or routing code changes are needed — the
machinery already exists (eu-frankfurt overrides for qwen3.7-plus/qwen3.6-plus
and us-virginia/cn-beijing for qwen3.6-flash were already present and working).

⚠️ One pre-existing operational note (not changed by this PR): eu-frankfurt
requests need a workspace id on the Alibaba credential, otherwise they fall back
to the shared trial.eu-central-1.maas.aliyuncs.com host (1000 RPM cap, no SLA).

Verifications to run for reviewers

1. Confirm the official Alibaba list prices with curl

The pricing page is server-rendered — plain curl returns the numbers. Both
commands below fetch the SAME page used for this PR (updated 2026-08-12):

# a) Raw HTML contains the model rows + USD list prices (international default = Singapore)
curl -sL "https://www.alibabacloud.com/help/en/model-studio/model-pricing" \
  -H "User-Agent: Mozilla/5.0" | grep -oE 'qwen3\.7-plus.{0,600}' | head -1
# Expect: ... List price $0.4 ... List price $1.6 ... (Singapore, 0-256K)

# b) The US/Frankfurt/Beijing region tabs are also server-rendered — grep for the dropped prices:
curl -sL "https://www.alibabacloud.com/help/en/model-studio/model-pricing" \
  -H "User-Agent: Mozilla/5.0" > /tmp/pricing.html
python3 - <<'PY'
import re
html = open("/tmp/pricing.html", encoding="utf-8", errors="replace").read()
t = re.sub(r"<[^>]+>", "|", html); t = re.sub(r"\|+", "|", t)
for model in ["qwen3.7-plus", "qwen3.6-plus", "qwen3.6-flash"]:
    i = t.find(model)
    print(model, "->", t[i:i+220].replace("|", " ")[:220])
PY
# Expect: qwen3.7-plus -> ... $0.276 ... $1.101 ... (US/Frankfurt/Beijing rows)

# c) qwen3.7-flash is JS-rendered on alibabacloud.com; use the CN-hosted page (server-rendered, CNY).
#    CNY 0.2/0.8 (≤32K) ≈ $0.028/$0.110 at the official ~7.14 CNY/USD rate.
curl -sL "https://help.aliyun.com/en/model-studio/model-pricing" \
  -H "User-Agent: Mozilla/5.0" | grep -oE 'qwen3\.7-flash.{0,300}' | head -1

Optional stronger check — run the packaged verifier (fetches + normalizes both pages):

python3 /tmp/verify-alibaba-pricing.py   # prints per-model, per-region USD prices

2. E2E the pricing through the LLMGateway stack

From a worktree on the PR branch (see Task 7 for the /tmp scratch setup):

# a) Build + unit suite (canonical invocation, per AGENTS.md)
pnpm build
pnpm test:unit        # = pnpm build:core && vitest run --no-file-parallelism
#   apps/gateway: zero failures expected. apps/api Postgres-dependent specs are
#   environmental (pre-existing), not caused by this PR.

# b) Local gateway — verify the catalogue exposes the new prices via the models API
docker compose up -d db redis      # isolated test DB per AGENTS.md
pnpm dev:gateway &                 # or the repo's documented local gateway start
curl -s "http://localhost:8787/v1/models?mapped=true" \
  -H "Authorization: Bearer test-token" | python3 - <<'PY'
import json, sys
data = json.load(sys.stdin)["data"]
for m in data:
    if m["id"].startswith("alibaba/qwen3.7-") or m["id"].startswith("alibaba/qwen3.6-"):
        p = m["pricing"]
        print(f"{m['id']:32s} prompt=${p['prompt']:.6f} completion=${p['completion']:.6f}")
PY
#   Expect: alibaba/qwen3.7-plus   prompt=0.000276 completion=0.001101
#           alibaba/qwen3.7-flash  prompt=0.000028 completion=0.000110   (region-pinned variants)
#           alibaba/qwen3.6-plus   prompt=0.000276 completion=0.001651
#           alibaba/qwen3.6-flash  prompt=0.000165 completion=0.000990

# c) E2E routing — cheapest-region smoke test (needs an Alibaba/DashScope key; skip if none)
TEST_MODELS="alibaba/qwen3.7-plus:us-virginia,alibaba/qwen3.7-flash:eu-frankfurt" \
  FULL_MODE=true pnpm test:e2e

Validation (from this PR)

• pnpm build ✅ · pnpm test:unit ✅ · no price increased in the diff ✅
• Every value above was captured live from the Alibaba page on 2026-08-12
  (page header: "Updated at: 2026-08-12 03:12:30").

---

[coderabbitai]

Important

Review skipped

Draft detected.

Please check the settings in the CodeRabbit UI or the .coderabbit.yaml file in this repository. To trigger a single review, invoke the @coderabbitai review command.

⚙️ Run configuration

Configuration used: Repository UI

Review profile: CHILL

Plan: Pro Plus

Run ID: 9340d534-e6fe-4672-9e88-00721ec89441

You can disable this status message by setting the reviews.review_status to false in the CodeRabbit configuration file.

Use the checkbox below for a quick retry:

• 🔍 Trigger review

---

Thanks for using CodeRabbit! It's free for OSS, and your support helps us grow. If you like it, consider giving us a shout-out.

❤️ Share

• X
• Mastodon
• Reddit
• LinkedIn

_{Comment @coderabbitai help to get the list of available commands.}