fix(models): correct qwen3.6-flash metadata

[vicovaro]

Summary

Qwen3.6 Flash is a native vision-language model supporting image, text, and video input, but our catalog had incorrect metadata. This corrects the vision flag, pricing, context size, cache pricing, and description.

References:

• https://www.qwencloud.com/models/qwen3.6-flash
• https://modelstudio.console.alibabacloud.com/ap-southeast-1?tab=doc#/doc/?type=model&url=2840914_2&modelId=qwen3.6-flash&serviceSite=international

Changes

Field	Before	After (≤256K)	After (>256K)
vision	false	true	true
inputPrice	$0.15/M	$0.25/M	$1.00/M
outputPrice	$0.90/M	$1.50/M	$4.00/M
cachedInputPrice	$0.015/M	$0.025/M	$0.10/M
cacheReadInputPrice	—	$0.025/M	$0.10/M
cacheWriteInputPrice	—	$0.3125/M	$1.25/M
contextSize	262K	1M	1M
pricingTiers	—	Up to 256K	Over 256K
Description	262K context	1M context, native vision-language	1M context, native vision-language

🤖 Generated with Claude Code

Summary by CodeRabbit

• New Features

  • Added native vision-language support for the Alibaba Qwen 3.6 Flash model.
  • Expanded the model’s context window to 1 million tokens.
  • Added tiered pricing details for different token usage ranges.

• Updates

  • Updated input, output, and cached-input pricing information.
  • Added cache read and write pricing details.
  • Refined the model description to reflect its current capabilities.

[coderabbitai]

Warning

Review limit reached

@vicovaro, you've reached your PR review limit, so we couldn't start this review.

Next review available in: 2 minutes

Enable usage-based reviews in Billing to review now. Otherwise, wait until the next included review is available.
You're only billed for reviews past your plan's rate limits ($0.25/file).

How can I continue?

After more reviews become available, a review can be triggered using the @coderabbitai review command as a PR comment. Alternatively, push new commits to this PR.

To avoid repeated limits, reduce automatic review volume by pausing incremental auto-reviews earlier, using label-based review opt-in, excluding WIP or generated PR titles, or requesting reviews manually when the PR is ready. If your team needs uninterrupted high-volume reviews, an organization admin can enable usage-based reviews.

How do review limits work?

CodeRabbit enforces per-developer PR review limits for each organization. Most developers receive the normal plan review availability.

For paid Pro and Pro+ PR reviews, CodeRabbit uses adaptive limits for sustained high-volume activity. When a developer's recent PR review activity reaches the 95th percentile or higher among CodeRabbit users, additional reviews become available more gradually as earlier reviews age out of the rolling window.

Please refer docs for additional details.

Review details

⚙️ Run configuration

Configuration used: Repository UI

Review profile: CHILL

Plan: Pro Plus

Run ID: d25b446e-b4d5-47ee-ac8a-1789f7a78a00

📥 Commits

Reviewing files that changed from the base of the PR and between 23597fc and 930a6d6.

📒 Files selected for processing (1)

• packages/models/src/models/alibaba.ts

Walkthrough

Updates Alibaba’s qwen3.6-flash definition with native vision-language support, a 1M-token context window, revised pricing, cache pricing, and token-based pricing tiers.

Changes

Qwen Flash model update

Layer / File(s)	Summary
Model metadata and provider settings packages/models/src/models/alibaba.ts	The model description, pricing, cache pricing, pricing tiers, context size, and vision capability are updated for qwen3.6-flash.

Estimated code review effort: 2 (Simple) | ~10 minutes

Possibly related PRs

• theopenco/llmgateway#3201: Updates Alibaba Qwen model pricing fields in the same model file.
• theopenco/llmgateway#3285: Updates related Alibaba Qwen Flash provider configuration, context, vision, and pricing tiers.

🚥 Pre-merge checks | ✅ 5

✅ Passed checks (5 passed)

Check name	Status	Explanation
Description Check	✅ Passed	Check skipped - CodeRabbit’s high-level summary is enabled.
Title check	✅ Passed	The title is concise and accurately describes the main change: correcting qwen3.6-flash metadata.
Docstring Coverage	✅ Passed	No functions found in the changed files to evaluate docstring coverage. Skipping docstring coverage check.
Linked Issues check	✅ Passed	Check skipped because no linked issues were found for this pull request.
Out of Scope Changes check	✅ Passed	Check skipped because no linked issues were found for this pull request.

✨ Finishing Touches

🧪 Generate unit tests (beta)

• Create PR with unit tests

---

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.}

[coderabbitai]

Actionable comments posted: 2

🤖 Prompt for all review comments with AI agents

Verify each finding against current code. Fix only still-valid issues, skip the
rest with a brief reason, keep changes minimal, and validate.

Inline comments:
In `@packages/models/src/models/alibaba.ts`:
- Around line 2872-2875: Update the Alibaba model region entries in the regions
configuration to add explicit base prices and all matching token-tier price
overrides for us-virginia and cn-beijing, including the tiers above 256K tokens,
using the provider’s regional pricing. Do not leave these regions to inherit the
mapping-level Singapore/international prices; alternatively remove the region
entries if regional routing is not supported.
- Around line 2849-2869: Update the pricingTiers entries in the Alibaba model
definition so cachedInputPrice is 0.05e-6 for the “Up to 256K” tier and 0.2e-6
for the “Over 256K” tier. Leave both cacheReadInputPrice values unchanged.

🪄 Autofix (Beta)

Fix all unresolved CodeRabbit comments on this PR:

• Push a commit to this branch (recommended)
• Create a new PR with the fixes

---

ℹ️ Review info

⚙️ Run configuration

Configuration used: Repository UI

Review profile: CHILL

Plan: Pro Plus

Run ID: 2b5b8a6c-43dd-424a-ad05-4bd3cd0a7d15

📥 Commits

Reviewing files that changed from the base of the PR and between 46e1aab and d9e0da9.

📒 Files selected for processing (1)

• packages/models/src/models/alibaba.ts

[vicovaro]

The only thing that I am still not sure, is the pricing for input output cache exactly, as I don't know if you get special deals. I get different pricing too according to different region https://modelstudio.console.alibabacloud.com/us-east-1?tab=doc#/doc/?type=model&url=2840914_2&modelId=qwen3.6-flash&serviceSite=global

Review again the pricing, the rest I am sure is good, especially the vision capability and the 1m context size.

[steebchen]

E2E results — alibaba/qwen3.6-flash

Ran the full gateway e2e suite scoped to this mapping (TEST_MODELS="alibaba/qwen3.6-flash" FULL_MODE=true pnpm test:e2e) on this branch.

Singapore (the only region reachable with our local key) passes everything — basic, vision (image input, confirming the vision: true flip), streaming, reasoning, reasoning+streaming, tool calls, streaming tool calls, tool-call results, reasoning+tool calls, empty-response/DONE handling.

Two things that are not caused by this PR, for the record:

• :us-virginia / :cn-beijing failures are all invalid_api_key — the seeded e2e BYOK key has keyRegion: "singapore". I probed dashscope-us.aliyuncs.com directly with the real us-virginia key and qwen3.6-flash responds fine there, so the region set is correct.
• JSON output flakes: the model returns a bare {} for the e2e prompt roughly 2 out of 3 times. Reproduced directly against DashScope, so it's upstream behaviour, unrelated to this change.

---

One thing worth a second look before merge

Every region-override value in this PR is exactly 0.66× the Singapore price — except one:

field (Over 256K tier)	Singapore	us-virginia / cn-beijing	ratio
inputPrice	1.0e-6	0.66e-6	0.66 ✅
outputPrice	4e-6	3.961e-6	0.990 ❌
cachedInputPrice	0.2e-6	0.132e-6	0.66 ✅
cacheReadInputPrice	0.1e-6	0.066e-6	0.66 ✅
cacheWriteInputPrice	1.25e-6	0.825e-6	0.66 ✅

(the Up to 256K tier is 0.66× on all five fields: 0.25→0.165, 1.5→0.99, 0.05→0.033, 0.025→0.0165, 0.3125→0.20625)

0.66 × 4 = 2.64, not 3.961. Looks like the 0.99 factor from the ≤256K output price leaked into the >256K row. If 0.66× is the intended conversion, this should be 2.64e-6; if the number came from an actual US/China price list, could you link it? Existing us-virginia entries in alibaba.ts are not a fixed multiple of Singapore, so a real price-list source would be the better justification for the whole block.

Two smaller notes:

• Adding cheaper us-virginia/cn-beijing prices means cheapest-region routing will pull unpinned alibaba/qwen3.6-flash traffic off Singapore. That works (verified above), just flagging it as an intended consequence.
• The PR description table lists cachedInputPrice as $0.025/M, but the code sets 0.05e-6. The code is right per repo semantics (cachedInputPrice = implicit 20% rate, cacheReadInputPrice = explicit 10% rate) — the table just needs fixing.

Holding off on auto-merge until the 3.961e-6 question is resolved.

[steebchen]

Correction — my 3.961e-6 concern was wrong, please disregard

I went and checked Alibaba's pricing pages directly. 3.961e-6 is exactly what Alibaba publishes. Both the English and Chinese Model Studio pricing pages list, for the Beijing region in USD:

tier	input	output
0 < Token ≤ 256K	$0.165/M	$0.99/M
256K < Token ≤ 1M	$0.66/M	$3.961/M

So the 0.66× pattern I spotted is a coincidence of Alibaba's own CNY→USD conversion, not a rule — and the >256K output row simply doesn't follow it. Every value in this PR matches the published figures verbatim.

I also verified the us-virginia block against repo convention rather than assuming. On main, every us-virginia entry in alibaba.ts has an identical cn-beijing twin (0.115/0.287, 0.144/0.574, 0.143/1.434, 0.022/0.215, 0.359/1.434) — and qwen-plus's 0.115/0.287 is exactly Alibaba's published Beijing USD price. So mirroring Beijing prices onto us-virginia is the established convention here, and this PR follows it. (Alibaba's US Virginia tab has no qwen3.6-flash row of its own, so there's nothing more authoritative to use.)

The cache fields are internally consistent with Alibaba's documented context-cache rules — implicit hit 20%, explicit hit 10%, write 125% — on all four tiers:

• SG ≤256K: 0.25 → 0.05 / 0.025 / 0.3125 ✅
• SG >256K: 1.0 → 0.2 / 0.1 / 1.25 ✅
• BJ+US ≤256K: 0.165 → 0.033 / 0.0165 / 0.20625 ✅
• BJ+US >256K: 0.66 → 0.132 / 0.066 / 0.825 ✅

Combined with the e2e run in my previous comment (Singapore green across the board, including image input confirming the vision: true flip), this is good to go. Enabling auto-merge.

The only thing still worth a quick fix is cosmetic: the description table lists cachedInputPrice as $0.025/M, but the code correctly sets 0.05e-6 (that field is the implicit 20% rate; cacheReadInputPrice is the explicit 10% one). Code is right, table is off.

Sorry for the noise.