ELEVEN REALTIME V1 MINI / MAX SELECTION WORKSHEET
Free original resource | Reviewed October 24, 2026
Companion: https://burki.dev/blog/eleven-realtime-v1-mini-vs-max

All fixtures and expected outcomes are unperformed. No paid model generation,
benchmark, account change, provider session or business action is represented.
Reading this worksheet is free; production services can have separate costs.

1. DEPLOYMENT ELIGIBILITY
Task owner / deployment surface:
Route: ElevenAgents / direct Realtime API / other verified route
Actual account and workspace access:
Candidate exact model ID:
Permitted voice / intended audience and language:
Provider residency / retention / configuration requirements checked on:
Token-limit requirement satisfied:
Workflow, transfer and mode restrictions reviewed:
Decision: ELIGIBLE / INELIGIBLE / UNRESOLVED
Evidence and unresolved prerequisite:

Do not convert an unavailable model into a fake zero performance score.
These models are not established as native Burki choices in this guide.

2. FREEZE THE TASK
Approved facts / version / owner:
Same voice, instructions and tool declarations for each candidate:
Client location and actual playback path:
Number of attempts planned:
Timing boundaries to measure:
What errors reject an answer:
Authority and budget for a later evaluation:
No results below should be filled without actual evidence.

3. ORIGINAL TASK A - FACTUAL CORRECTION
Fictional Alder Workshop event:
Evening visitors use the side entrance.
Doors open 18:15. The event begins 18:45, in the business's published timezone.
Caller: When do doors open? Actually, I mean when does the event begin?
Expected: 18:45; keep the two times distinct. No ticket or admission promise.

Candidate / attempt / accepted configuration evidence:
Observed answer:
Correct fact and correction handling:
Timing, with its actual measurement boundary:
Failure / uncertainty / decision:

4. ORIGINAL TASK B - READ-ONLY MOCK LOOKUP
Mock records, not real event offers:
WOOD-01 -> room Cedar, main entrance
PAPER-02 -> room Willow, garden entrance
Any other code -> unknown
Caller: WOOD-01. Sorry, I mean PAPER-02. Where should I go?
Expected: lookup PAPER-02 and explain Willow / garden entrance.
Missing result: remain uncertain; do not borrow WOOD-01 information.
No reservation, payment, admission or actual provider integration is supplied.

Candidate / attempt / configuration accepted:
Actual lookup name and input:
Actual result evidence:
Final answer matched selected result:
Wrong identifier / fabricated action / error:
Decision and correction owner:

5. KEEP MEASURES SEPARATE
Factual correctness:
Tool task correctness:
Conversation / listener preference:
Time to first provider audio token:
Time to actual audible output in the chosen application, if measured:
Failed or ineligible attempts:
Sample count and excluded attempts with reasons:
Do not label a provider token timing as a caller-heard latency.
Do not turn published benchmark percentages into a business success forecast.

6. COST RECORD FOR THE SAME WORKLOAD
Product route / plan / tariff checked on:
Additional connected minutes:
Burst minutes and applicable concurrency condition:
Actual input / cached input / output token units:
Provider usage cost for all attempts:
Credits, taxes or other excluded charges:
Carrier / platform / storage / other service costs:
Accepted task count:
Cost assumptions / remaining unknowns:

7. SELECTION RECORD
Evidence favoring each eligible candidate for this task:
What the review did not establish:
Configuration versus provider acceptance versus actual result:
Decision owner / date:
Next review trigger:

Current configuration contract:
https://elevenlabs.io/docs/eleven-agents/customization/realtime-models
Direct API tariff:
https://elevenlabs.io/pricing/api
Independent benchmark methods:
https://artificialanalysis.ai/methodology/speech-to-speech-benchmarking