# HTTP 200, empty answer

> The reasoning pass spent the whole completion budget before the first answer token, so the server reported success with nothing to show; raise the budget or turn reasoning off, and count answer tokens instead of status codes.

- Platform: Any local stack
- Runtime: llama.cpp server (Vulkan)
- Published: 2026-09-02
- Sources: https://agmind.ai/reports/answerless-http-200/, https://agmind.ai/glossary/#answerless-request
- Canonical: https://agmind.ai/symptoms/llama-server-http-200-empty-content/

## What you see

The request completes. The server streams tokens and returns a success
status. The client gets no answer text at all. The report names the
symptom in its title:

```text
HTTP 200, empty answer
```

In the per-request quality records every such request has the same shape:

```text
nonzero reasoning characters and zero answer characters
```

Every conventional serving metric looks healthy while it happens. Status is
200. Time to first token is excellent, because the first token is reasoning.
Tokens generated is high, because the model did work. The only counter that
catches it is answer tokens, the ones after the reasoning block closes, and
almost nobody collects it. A `usage` field that merges the two hides the
failure.

## Where we saw it

Strix Halo, llama.cpp Vulkan pinned by digest, single stream, frozen
corpora. Two model families reproduced it: Qwen3.6-35B-A3B with reasoning
enabled at a 1024-token completion budget, and Gemma-4-26B in its default
operating mode, on a chat template that does not even expose a reasoning
toggle. Report published 2026-08-14.
The exact runtime build and model digest are on the claim pages linked below.

## Cause

A thinking pass and a token budget share one completion window. When the
model emits a reasoning block, it writes that block before the answer. When
the budget runs out before the reasoning block closes, generation stops and
no answer token is ever produced. The transport succeeded, so the server
reports success. The report calls the failure rate a budget-starvation curve
rather than a model defect: the rate is a property of the model × budget ×
corpus cell.

## Fix

The report gives two settings and one check, no command line:

- **Budget for the reasoning, not just the answer.** "A 1024-token window
  that would fit any answer in our corpus fits almost no
  reasoning-plus-answer pair." Size the completion budget for the pair.
- **Or turn reasoning off for everyday traffic.** On the Strix Halo box,
  disabling the reasoning block cut the first-answer wait from tens of
  seconds to the value on
  [the claim page](https://agmind.ai/claims/strix.qwen36.interactive2.c1.ttfa-nothink/), with
  task success unchanged on these workloads. The paired cells are in
  [the reasoning cost comparison](https://agmind.ai/compare/thinking-vs-no-thinking-qwen36/).
- **Monitor answer tokens.** Count tokens after the reasoning block closes
  and treat an empty answer as a failed request even when the transport
  succeeded. That is why our harness counts [TTFA](https://agmind.ai/glossary/#ttfa)
  separately from TTFT and keeps answerless requests in the denominator.

## Still open upstream?

Not tracked by us. Neither source names an upstream issue; the report files
the failure rate under budget starvation, not a runtime or model defect.

## Evidence

- Report: [HTTP 200, empty answer: the failure mode your monitoring cannot see](https://agmind.ai/reports/answerless-http-200/)
- Glossary: [answerless request](https://agmind.ai/glossary/#answerless-request)
- Claim pages: [Qwen3.6 answerless share at the tested budget](https://agmind.ai/claims/strix.qwen36.interactive2.c1.answerless-1k/), [Gemma-4 answerless share in default mode](https://agmind.ai/claims/strix.gemma4.interactive2.c1.answerless-default/)
- Per-request quality records: [botAGI/agmind-lab](https://github.com/botAGI/agmind-lab)
- Related configuration: [Strix Halo × llama.cpp Vulkan × Qwen3.6-35B-A3B](https://agmind.ai/tested-configurations/strix-halo-qwen36-llamacpp-vulkan-interactive/)
- Related symptom: [thinking stream starts immediately, the answer comes late or never](https://agmind.ai/symptoms/thinking-model-first-token-instant-answer-late/)

---

Machine-readable claim registry: https://agmind.ai/claims.json · llms.txt: https://agmind.ai/llms.txt