Make the transcription model configurable; default to gpt-transcribe

[sadasant-ghost]

Stacked on #4.

whisper-1 was hardcoded in conductor_api.py. It is the cheapest thing to get wrong and the most expensive to trust, because it does not fail silently — it fills silence with fluent invention.

Over one week of Daniel's voice notes it produced a Korean news sign-off, a YouTube outro, and turned "billing" into "building" in a message he was drafting for executives.

Measured, not assumed

One second of a 440 Hz sine tone, against the live API:

model	returned	billing
gpt-transcribe	""	duration, $0.0045/min
gpt-4o-mini-transcribe	""	tokens
whisper-1	"Oh"	duration, $0.006/min

Whisper invented a word from a pure tone, on demand, which is the failure mode above reproduced in one request.

The change

Default is now gpt-transcribe: quieter on non-speech, cheaper ($0.0045 vs $0.006), and billed by duration — which this endpoint needs, because /api/transcribe reports a per-minute cost and a token-priced model would make that number a guess.

CONDUCTOR_TRANSCRIBE_MODEL overrides it, matching how every other knob in this file already works.

Two smaller things that follow from being honest about cost:

• A model with no known per-minute price reports cost: null rather than a plausible wrong number. gpt-4o-mini-transcribe is better than Whisper but token-billed, and this endpoint cannot know the token count.
• The response now names the model that produced the text, so a surprising transcript can be attributed rather than guessed at.

Verified live

Deployed and restarted before opening this:

POST /api/transcribe
{"text": "", "seconds": 1.0, "cost": 0.0001, "model": "gpt-transcribe"}

32 tests, including one asserting the default is cheaper than what it replaced and one asserting a token-billed model reports no cost rather than fiction.

---

Summary by cubic

Makes the transcription model configurable and defaults to gpt-transcribe instead of hardcoded whisper-1, which hallucinated speech on silence. The endpoint now reports cost based on the model's per-minute rate, returning null when the model is token-billed, and includes the model name in responses.

Bug Fixes

• whisper-1 inventing words on pure tones is fixed by switching the default to gpt-transcribe.
• Unknown or token-billed models now report cost: null instead of a guessed number.
• Adds CONDUCTOR_TRANSCRIBE_MODEL to override the model.
• Adds model to the /api/transcribe response so output can be attributed.
• Keeps whisper-1 price constant for compatibility.

^{Written for commit 7646944. Summary will update on new commits.}

[cubic-dev-ai]

2 issues found across 3 files

Prompt for AI agents (unresolved issues)

Check if these issues are valid — if so, understand the root cause of each and fix them. If appropriate, use sub-agents to investigate and fix each issue separately.


<file name="conductor_api.py">

<violation number="1" location="conductor_api.py:181">
P3: `transcribe_cost`'s docstring says None is returned only "when the model is token-billed", but the implementation returns None for any unknown model not present in `USD_PER_MIN` (the added test even asserts this for "something-new"). Widen the docstring to cover the unknown-model case so it matches the actual contract.</violation>

<violation number="2" location="conductor_api.py:189">
P3: `WHISPER_USD_PER_MIN` is defined on the added line with the comment "kept: referenced elsewhere", but nothing in the repository references it anymore — the only consumers were the lines this change replaced with `transcribe_cost(secs)`. It is dead code and the comment is misleading. Remove the line.</violation>
</file>

<sub>Reply with feedback, questions, or to request a fix.

Fix all with cubic | Re-trigger cubic</sub>

[cubic-dev-ai]

P3: transcribe_cost's docstring says None is returned only "when the model is token-billed", but the implementation returns None for any unknown model not present in USD_PER_MIN (the added test even asserts this for "something-new"). Widen the docstring to cover the unknown-model case so it matches the actual contract.

Prompt for AI agents

Check if this issue is valid — if so, understand the root cause and fix it. At conductor_api.py, line 181:

<comment>`transcribe_cost`'s docstring says None is returned only "when the model is token-billed", but the implementation returns None for any unknown model not present in `USD_PER_MIN` (the added test even asserts this for "something-new"). Widen the docstring to cover the unknown-model case so it matches the actual contract.</comment>

<file context>
@@ -146,7 +146,47 @@ def translate_command(description, cwd):
+def transcribe_cost(seconds, model=None):
+    """Dollars for `seconds` of audio, or None when the model is token-billed.
+
+    None is the honest answer for a token-priced model: the caller knows the
+    duration and nothing about the token count, and a number invented here
+    would be indistinguishable from a real one.
</file context>

[cubic-dev-ai]

P3: WHISPER_USD_PER_MIN is defined on the added line with the comment "kept: referenced elsewhere", but nothing in the repository references it anymore — the only consumers were the lines this change replaced with transcribe_cost(secs). It is dead code and the comment is misleading. Remove the line.

Prompt for AI agents

Check if this issue is valid — if so, understand the root cause and fix it. At conductor_api.py, line 189:

<comment>`WHISPER_USD_PER_MIN` is defined on the added line with the comment "kept: referenced elsewhere", but nothing in the repository references it anymore — the only consumers were the lines this change replaced with `transcribe_cost(secs)`. It is dead code and the comment is misleading. Remove the line.</comment>

<file context>
@@ -146,7 +146,47 @@ def translate_command(description, cwd):
+    return None if rate is None else round(seconds / 60 * rate, 4)
+
+
+WHISPER_USD_PER_MIN = USD_PER_MIN["whisper-1"]  # kept: referenced elsewhere
 
 # Voice input reaches an agent as ordinary text, and speech-to-text mishears in ways
</file context>