GPT-LIVE 1 / GEMINI 3.8 LIVE BACKGROUND-WORK EVALUATION
Reviewed October 9, 2026. Free worksheet, not free production voice usage.
Capability comparison only. No matched model benchmark or provider calls run.
Do not put secrets/customer data in this worksheet. Start with fictional read-only
stubs; a stub checks app logic and is not a voice-model measurement.

1. EXACT CONFIGURATION
Candidate A voice ID: gpt-live-1
Backend model/service, mode and prompt/version:
Candidate B voice ID: gemini-3.8-live (base, not Extended Thinking)
Tool declaration behavior and result scheduling:
Current Google model/tools docs disagree on legacy defaults/interrupt enum
spelling. Verify actual accepted SDK/API contract before configuring.
Worked script is English only. Verify exact model language contracts; evaluate
caller languages/accents in separate labeled sets. Modalities are not accuracy.
Each: SDK/API version, transport, region, language, microphone/telephone path,
prompt/policy version, tool implementation, network conditions, test date.
Burki currently exposes GPT Live, not a verified Gemini Live selector.
Current source uses client delegation; selector is not execution proof.

2. TASK STATE DESIGN TO REVIEW
Task ID:
Input revision and latest caller-confirmed reference:
Tool call ID and input revision used:
Pending/completed/cancelled/superseded state:
Result provenance and external operation ID:
Who can authorize a write:
Duplicate prevention:
How stale result is excluded from current spoken answer:
Cancellation before start / while pending / after commit:
Fallback when outcome is unknown:
This is an engineering design worksheet, not an exposed Burki control.

3. FICTIONAL DELAYED READ
Caller asks delivery request R204.
Lookup starts with fixed local delay chosen before evaluation: ____
Caller corrects reference to R240 while pending.
Old lookup returns: R204 awaiting collection.
Expected: never attribute R204 status to R240; preserve correction; explain pending.
Stub actual observed result (not model benchmark):
Later authorized voice session/result evidence (if performed):

4. CASES AND PREDECLARED CRITERIA
Case | Proposed pass | Observed result | Evidence | Reviewer
Correction during read | no stale-reference claim:
Cancel before write starts | write prevented:
Cancel after external commit | state actual outcome/remedy, no false undo:
Duplicate result/reconnect | no duplicate action:
Tool timeout | outcome unconfirmed, agreed fallback:
Playback interruption | stops audio as intended, task state accurate:
Ambiguous 'stop' | clarify speaking vs lookup vs business action when needed:
These are criteria, not measured scores.

5. COMPARABLE MEASUREMENT LOG
Attempt | Candidate exact version/backend | Case | Delay/conditions
Caller audio and transcript | provider acceptance/session ID | tool/task receipt
Acknowledgment delay | time to correct useful answer | correction retained?
Stale-result exposure? | duplicate action? | cancellation actual outcome
Audio heard vs generated | failure/timeout | usage record | reviewer
Keep all attempted cases, including failed connections and timeouts.
If backend/config differs, disclose; don't call it voice-model-only benchmark.
Report raw counts/sample size. Use repeated trials for tail latency; record the
uncertainty method. No common current exact-pair benchmark verified here.

6. COST, AS LISTED OCTOBER 9
GPT-Live 1: USD 0.05 per session minute, billed per second; backend/tools separate.
Illustrative six-minute voice-session component: USD 0.30, not an observed bill.
https://developers.openai.com/api/docs/models/gpt-live-1
Gemini 3.8 Live paid per 1 million tokens:
audio input USD 3, audio output USD 12; text input USD 0.75 / output USD 4.50.
Provider also lists USD 0.005 per input audio minute and USD 0.018 per output audio minute.
https://ai.google.dev/gemini-api/docs/pricing
Session duration != audio output duration. Use recorded billable units.
Add backend/tool/grounding/platform/telephone/transport costs as applicable.
Candidate | actual billable units | rate/date | subtotal | other costs
No savings ratio/winner inferred from unlike units.

7. BURKI INSPECTION AND DECISION
Assistants -> draft -> Voice -> Model and provider choices if collapsed
Conversation engine -> GPT Live option
Record actual live model and separate reasoning configuration shown:
Instructions/actions/readiness/credentials/Usage & billing review:
If authorized testing occurs, record actual accepted session + caller-heard result
+ business receipt. Neither source nor dropdown proves acceptance.
Selected configuration / evidence / unresolved limits / owner / review date:

Primary contracts:
https://developers.openai.com/api/docs/guides/live
https://developers.openai.com/api/docs/guides/live-delegation
https://ai.google.dev/gemini-api/docs/models/gemini-3.8-live
https://ai.google.dev/gemini-api/docs/live-api/tools