CekuraBench
All comparisons
GoogleCascade baseline

Gemini 3.8 Live Extended Thinking vs Flux → GPT-4.1 → ElevenLabs Flash

Flux → GPT-4.1 → ElevenLabs Flash is ahead of Gemini 3.8 Live Extended Thinking on reliability, success rate, agent responses, data accuracy, stalled calls, interruption and response time.

Compare

Success rate by suite

Appointments59 scenariosMedicare23 scenarios
  • GPT Realtime 2.1OpenAI
    92%
    81%
  • GPT-Live 1 + Sol lowOpenAI
    97%
    49%
  • Gemini 3.8 Live ThinkingGoogle
    82%
    61%
  • Grok Think Fast 2.0xAI
    96%
    25%
  • Gemini 3.1 Flash LiveGoogle
    93%
    32%
  • Phonic v1Phonic
    97%
    14%
  • Realtime 2.1 MiniOpenAI
    84%
    4%
  • Nova 2 SonicAmazon
    80%
    6%
  • Cascade, GPT-4.1Not ranked
    97%
    83%
Each model's share of calls that passed every check, on each suite.

Every metric, side by side

Reliability

Scenarios that passed on all three runs.

Gemini 3.8 Live Thinking
56.1%6th
Cascade, GPT-4.1
82.9%

Cascade, GPT-4.1, 26.8 points higher

Success rate

Calls that passed every check.

Gemini 3.8 Live Thinking
76.4%3rd
Cascade, GPT-4.1
93.1%

Cascade, GPT-4.1, 16.7 points higher

Agent responses

Calls where the agent's replies reached the expected outcome.

Gemini 3.8 Live Thinking
94.7%6th
Cascade, GPT-4.1
99.2%

Cascade, GPT-4.1, 4.5 points higher

Data accuracy

Calls where everything the agent saved matched exactly.

Gemini 3.8 Live Thinking
80.2%3rd
Cascade, GPT-4.1
94.1%

Cascade, GPT-4.1, 13.9 points higher

Stalled calls

Calls where the agent went silent for 10 seconds or more.

Gemini 3.8 Live Thinking
3.7%6th
Cascade, GPT-4.1
0.8%

Cascade, GPT-4.1, 2.8 points lower

Interruption

How well it handled being talked over, 0 to 5.

Gemini 3.8 Live Thinking
4.906th
Cascade, GPT-4.1
4.96

Cascade, GPT-4.1, 0.06 higher

Response time

Median time to start replying after the caller finishes.

Gemini 3.8 Live Thinking
2.32s7th
Cascade, GPT-4.1
2.17s

Cascade, GPT-4.1, 0.15s sooner

Cost / min

One minute of call at the vendor's published rates.

Gemini 3.8 Live Thinking
No verified rate
Cascade, GPT-4.1
No verified rate

Not comparable

The small figure is the place among ranked models. The cascade is not ranked.

Cascade, GPT-4.1 answers first

Cascade, GPT-4.1 2.17s, Gemini 3.8 Live Thinking 2.32s at the median.

Response time, median to p90

MedianTo p90
Medianp90
  • Phonic v1Phonic
    Phonic v1: median 1.59s, p90 1.86s
    1.59s1.86s
  • Grok Think Fast 2.0xAI
    Grok Think Fast 2.0: median 1.62s, p90 1.94s
    1.62s1.94s
  • GPT Realtime 2.1OpenAI
    GPT Realtime 2.1: median 1.94s, p90 2.46s
    1.94s2.46s
  • GPT-Live 1 + Sol lowOpenAI
    GPT-Live 1 + Sol low: median 2.02s, p90 2.42s
    2.02s2.42s
  • Realtime 2.1 MiniOpenAI
    Realtime 2.1 Mini: median 2.08s, p90 2.75s
    2.08s2.75s
  • Nova 2 SonicAmazon
    Nova 2 Sonic: median 2.18s, p90 2.76s
    2.18s2.76s
  • Gemini 3.8 Live ThinkingGoogle
    Gemini 3.8 Live Thinking: median 2.32s, p90 3.06s
    2.32s3.06s
  • Gemini 3.1 Flash LiveGoogle
    Gemini 3.1 Flash Live: median 2.88s, p90 4.00s
    2.88s4.00s
  • Cascade, GPT-4.1Not ranked
    Cascade, GPT-4.1: median 2.17s, p90 2.51s
    2.17s2.51s
2.0s3.0s4.0s
Seconds from the caller finishing to the agent starting to reply, per model.

What we saw on the calls

Gemini 3.8 Live Extended Thinking

  • Strong on the long form; when it slips, it skips a save it said it made.
  • Weak on the short form: phone numbers saved wrong, lookups never made.
  • Acts on half an answer instead of asking; weakest when the caller corrects themselves.
  • Least predictable: the same scenario passes one run and fails the next.
  • Replies fall behind or stop; callers hang up waiting.
Model
gemini-3.8-live-extended-thinking
Setting
Thinking level high
Voice
Charon
Turn taking
Shared local turn detection
Audio
16 kHz
Cost
Some calls used cached input the vendor meters at a rate we have not verified.

Flux → GPT-4.1 → ElevenLabs Flash

  • Most scenarios pass every run; none fails every run.
  • Rare long-form slips: a reference from an earlier step left out.
  • Not the quickest to reply, but almost never leaves the caller waiting.
Model
flux-general-en → gpt-4.1 → eleven_flash_v2_5
Setting
No reasoning setting
Voice
ElevenLabs 21m00Tcm4TlvDq8ikWAM
Turn taking
Speech-to-text turn detection
Audio
16 kHz
Cost
Not published: no verified combined rate for the three services.
See the full leaderboardReference agent

Related comparisons

Frequently asked questions

Is Gemini 3.8 Live Extended Thinking or Flux → GPT-4.1 → ElevenLabs Flash more reliable?

Gemini 3.8 Live Extended Thinking passed 46 of 82 scenarios on all three runs, and Flux → GPT-4.1 → ElevenLabs Flash passed 68. Flux → GPT-4.1 → ElevenLabs Flash is ahead, by 26.8 points.

Which saves caller data more accurately, Gemini 3.8 Live Extended Thinking or Flux → GPT-4.1 → ElevenLabs Flash?

Gemini 3.8 Live Extended Thinking saved everything exactly right on 80.2% of the calls that save data, and Flux → GPT-4.1 → ElevenLabs Flash on 94.1%. On Appointments, 86.5% against 98.8%. On Medicare, 63.6% against 81.8%. A call counts only when nothing is missing, wrong, extra or saved twice.

Which responds faster, Gemini 3.8 Live Extended Thinking or Flux → GPT-4.1 → ElevenLabs Flash?

Flux → GPT-4.1 → ElevenLabs Flash starts replying 0.15s sooner at the median. Gemini 3.8 Live Extended Thinking takes 2.32s at the median and 3.06s at p90; Flux → GPT-4.1 → ElevenLabs Flash takes 2.17s and 2.51s. Both are measured on the call audio, from the caller finishing to the agent starting to reply.

Which costs less, Gemini 3.8 Live Extended Thinking or Flux → GPT-4.1 → ElevenLabs Flash?

There is no verified per-minute rate for Gemini 3.8 Live Extended Thinking and Flux → GPT-4.1 → ElevenLabs Flash, so cost cannot be compared yet.

How were Gemini 3.8 Live Extended Thinking and Flux → GPT-4.1 → ElevenLabs Flash compared?

Both ran as the whole agent in the same open-source Pipecat pipeline, with the same prompt, tools and phone line. A simulated caller worked through all 82 scenarios on live calls, three times each, 2,214 calls across the benchmark. A call passes only when the agent said the right things and saved the right data. The cascade runs the same agent for reference and is never ranked.