DeepSeek V4 Flash: xhigh reasoning effort silently clamped to high

[gojack10]

What happened?

supportsXhigh() in packages/ai/src/models.ts does not recognize deepseek-v4-flash, so xhigh is silently clamped to high before reaching the provider. This mirrors #3662 (Pro), fixed in b4e93feb. Empirical testing shows Flash supports the higher tier too.

Steps to reproduce

pi --provider openrouter --model deepseek/deepseek-v4-flash --thinking xhigh -p "Say: ok"

Session JSONL shows the silent downgrade:

{"type": "thinking_level_change", "thinkingLevel": "xhigh"}
{"type": "thinking_level_change", "thinkingLevel": "high"}

Direct OpenRouter API comparison on an identical prompt to deepseek/deepseek-v4-flash, varying only reasoning.effort:

effort	HTTP	reasoning_tokens
minimal	200	226
low	200	266
medium	200	297
high	200	293
xhigh	200	1238 (4.2× the high baseline)
max	400	rejected — Invalid option: expected one of xhigh|high|medium|low|minimal|none

The bimodal distribution (minimal–high cluster within ±15%, xhigh jumps 4×) indicates DeepSeek upstream coerces minimal/low/medium to high and treats xhigh as a strictly higher tier. The 400 on max confirms xhigh is the canonical OpenRouter name.

Expected behavior

supportsXhigh() returns true for deepseek-v4-flash, mirroring the existing deepseek-v4-pro check.

Version

0.70.2

[github-actions]

This issue was auto-closed. All issues from new contributors are auto-closed by default.

Maintainers review auto-closed issues daily and reopen worthwhile ones. Issues that do not meet the quality bar in CONTRIBUTING.md will not be reopened or receive a reply.

If a maintainer replies lgtmi on one of your issues, your future issues will stay open. If a maintainer replies lgtm, your future issues and PRs will stay open.

See CONTRIBUTING.md.

[badlogic]

fixed on main

[coekfung]

fixed on main

it seems that deepseek-v4-pro has been fixed but deepseek-v4-flash not
https://github.com/badlogic/pi-mono/blob/2a7e82d3ef5ff887a0de1334bad26faf9ce73def/packages/ai/src/models.ts#L56-L65

[ppowo]

@badlogic Maybe I'm missing something but I think the fix that landed on main only covers deepseek-v4-pro. The affected model, deepseek-v4-flash won't be caught by the substring check.