API contract: /v1/chat/completions accepts reasoning_effort but the level has no effect (deepseek-v4.1-flash)

[linfeng961]

Cross-posting from Issues (#9232) because this is arguably an API-contract question as much as a bug: the Cloud API accepts reasoning_effort, validates it, and then behaves identically whatever you ask for.

Same prompt, deepseek-v4.1-flash, 3 samples per level, max_tokens: 3000, comparing usage.completion_tokens_details.reasoning_tokens:

reasoning_effort	OpenDesign Cloud API	OpenCode Go (same model, same prompt)
none	0 / 0 / 0	0 / 0 / 0
low	538 / 808 / 1271 (mean 872)	620 / 635 / 889 (mean 715)
high	614 / 1337 / 850 (mean 934)	1043 / 1171 / 1053 (mean 1089)
max	1147 / 862 / 1026 (mean 1012)	2596 / 1991 / 3000 (mean 2529)

On the Cloud API that is 1.16× from low to max, with intra-level spread larger than the gap between levels. On the other gateway the same model spreads 3.5×. A second control gateway (OpenCode Zen) gives 4.4×, so it is not a one-off.

Two supporting observations:

• An invalid value returns HTTP 400 with the accepted vocabulary, so the field is parsed, not ignored as unknown.
• With an effort supplied, prompt_tokens reports 74 for a prompt that reports 48 at reasoning_effort: "none" — the gateway appears to rewrite the request rather than pass it through.

The question I'd like a stance on: when a route cannot honour a graded level, is the intended behaviour to accept and ignore it, or to reject it? Accepting it makes the failure invisible to callers — you can't tell a downgraded request from a normal one without writing a benchmark. #8839 asks for a per-request usage log that would at least surface it; that has been in product-direction review since Oct 17.

Either fix the ladder, or reject the parameter, or document that this route ignores it. Full repro, raw numbers and controls are in #9232.

[linfeng961]

If you are on a Design Plan (Go/Plus/Pro/Max) key and have five minutes, an independent replication would be genuinely useful here — one report is easy to file as an outlier, three are not.

What to check: whether reasoning_effort changes the reasoning budget on /v1/chat/completions for deepseek-v4.1-flash.

KEY=...
for E in none low high max; do for i in 1 2 3; do
  curl -s https://amr-link.open-design.ai/v1/chat/completions \
    -H "Authorization: Bearer $KEY" -H "Content-Type: application/json" \
    -d "{\"model\":\"deepseek-v4.1-flash\",\"max_tokens\":3000,\"reasoning_effort\":\"$E\",\"messages\":[{\"role\":\"user\",\"content\":\"Integers 1..1000 divisible by none of 2, 3, 5: derive with inclusion-exclusion step by step, then verify the arithmetic.\"}]}" \
    | jq -r "\"$E \(.usage.completion_tokens_details.reasoning_tokens)\""
done; done

Two things matter or the numbers mean nothing: the prompt has to need real derivation, and max_tokens has to be at least 3000 (below that, max hits the ceiling and every level looks the same). Three samples per level, because single-sample spread on this prompt is around +/-40% — my own first pass, one sample per level, was unreadable.

What I get on the Cloud API (means of 3): none 0, low 872, high 934, max 1012 — a 1.16x spread. The identical prompt sent straight to api.deepseek.com gives low 483, high 1146, max 2255 — a 4.7x spread. So the ladder works upstream and not through this route.

If you get a materially different result, please say so — that would be more informative than a confirming run. Either way, #9232 is the issue to add it to.