DeepSeek: a retired model id is listed with no context length and no price, so the panel reads 128K and $0.00

Cline Surface

VS Code Extension

Cline Version

4.1.21

Beta version

• I am using a beta version of Cline

What happened?

On the DeepSeek provider with my own key, the model id deepseek-v4-flash - still accepted by api.deepseek.com and served by DeepSeek-V4.1-Flash - is listed with no context length and no price, so the panel reports a 128,000-token window and $0.00 cost while the API serves 1,048,576 tokens and bills at the Flash price.

• GET https://api.deepseek.com/models returns exactly two ids: deepseek-flash (name DeepSeek-V4.1-Flash, context_window 1048576, max_output_tokens 393216) and deepseek-v4-pro. The retired names (deepseek-v4-flash, deepseek-v4-flash-vision-exp) are no longer listed, but requests to them are still accepted and served by V4.1-Flash at the Flash price.
• The models.dev record for the DeepSeek provider's deepseek-v4-flash carries status = "deprecated", a limit.output of 393216 and no limit.context at all; the V4.1-Flash entry beside it reads $0.00/$0.00.
• With no context length in the record, Cline falls back to a built-in 128,000 default (the ... ?? 128e3 beside Met(r?.contextWindow) in dist/extension.js), and with no price reaching it the run shows as free.

Steps to reproduce

1. API Provider -> DeepSeek, paste a key, then choose the model id deepseek-v4-flash (it is the DeepSeek defaultModelId in some builds, which is how it arrives configured).
2. Open the model list: that id shows a 128K context and no prices, while deepseek-flash in the same list shows 1M and $0.15/$0.60.
3. Run a task: the cost reads $0.00 each turn, and the context bar reads nearly full at a conversation size the real window holds ten times over - so long sessions are compacted far earlier than the model requires.

Provider/Model

DeepSeek (https://api.deepseek.com), model id deepseek-v4-flash (retired alias for DeepSeek-V4.1-Flash)

Diagnostics

Measured on one machine, with no change of provider or model id: session costs of $0.6765, $0.3177 and $0.1537 through 2026-09-25, then $0 at the next two sessions, and 1M with real prices the moment deepseek-flash was selected. Three compactions followed the early-full bar, at 141,923, 239,936 and 317,686 tokens before the summary; the one genuine limit compaction that day sits at 1,111,504 tokens, past the real window.

Workaround: select deepseek-flash, the id DeepSeek's own model list returns. What I would expect instead is a retired id either carrying its base model's real context length and price, or being refused rather than quietly mis-described.

Related: #13548 ("the context will be limited to 128K") reports the same reading for the vision-exp id on an OpenAI-compatible provider.

Filed from the operator's own Cline session, which took the measurements above on this machine.