• Root cause found: Together spent the entire output budget ...

Prompt

• Root cause found: Together spent the entire output budget on hidden reasoning and never reached the visible answer. Evidence: - Both requests returned HTTP 200 with valid usage and request IDs. - Text: 12 output tokens, 43 reasoning characters, finish_reason: length. - Image: 16 output tokens, 59 reasoning characters, finish_reason: length. - The stream parser correctly captured reasoning separately, so this is not a parsing failure. The current harness now disables reasoning and allows 64 tokens at /Users/pyro/Documents/OverBench/scripts/ vlm_benchmark.mjs:82. However, one bug remains: an HTTP-200 empty response is not retried. It should retry reasoning- only truncations with 128–256 tokens. Together confirms that max_tokens covers reasoning plus visible content and recommends generous limits. Together reasoning documentation (https://docs.together.ai/docs/inference/chat/reasoning) I couldn’t live-retest because the Together key is absent from the current shell. This was not a payment or access failure.