Gemini Models Tested: Which One Is Right for Call Center Transcription?

Verdict: Recommended — best cost/accuracy balance

Gemini 3 Flash (Preview) delivers the best overall balance of speed, cost, and accuracy across the 7 models tested. At $0.0023/min and 31.9 seconds of latency, it correctly diarizes all speakers and redacts the account number without a secondary pass.

The one gap is name consistency: after initially capturing the agent name correctly, the model reverts to a generic label for subsequent turns. This is a consistent pattern across most models in this evaluation and is addressable with a lightweight post-processing step.

For most call center pipelines — regulated or otherwise — Gemini 3 Flash (Preview) is the primary recommendation. It avoids the diarization inversion that plagues both Flash-Lite variants, provides redaction without the cost premium of Pro-tier models, and processes a typical call in under 35 seconds.

Verdict: Recommended — best GA alternative

Gemini 2.5 Flash is the recommended generally-available fallback for pipelines that cannot use preview models. At $0.0013/min, it is the lowest-cost model with correct diarization. It captured the most utterances of the non-Pro models — 57 — and correctly redacted the account number.

Name consistency reverts with some transcription noise on reversion, which is a mild quality concern for coaching workflows that depend on clean speaker attribution strings. The diarization labels themselves are correct throughout, so the underlying speaker separation is sound.

If your procurement policy requires GA-only models, Gemini 2.5 Flash is the clear choice over 2.5 Pro (higher cost, misses redaction) and 2.5 Flash-Lite (diarization inversion, content loss).

Verdict: Recommended for best redaction — higher cost

Gemini 3.5 Flash is the only model that correctly redacted both the account number and the card number in this evaluation. If maximum PII coverage is the primary requirement, it is the right pick.

The cost is $0.0076/min — significantly higher than Gemini 3 Flash (Preview) at $0.0023/min — and latency is 46.5 seconds. Diarization is correct throughout. Name consistency reverts in the same pattern as other models.

For regulated industries where missing a card number redaction carries compliance or legal exposure, the cost premium is likely justified. For general call center transcription, Gemini 3 Flash (Preview) covers the account number case at a fraction of the price.

Verdict: Not recommended — high cost, misses redaction

Gemini 2.5 Pro produced the highest utterance count in the evaluation at 67 — indicating it captures the most granular turn-level detail. Diarization is correct throughout. However, it misses account number redaction despite apparently capturing the correct numeric value in the transcript, which suggests a possible hallucination or redaction-logic gap rather than a transcription failure.

At $0.0070/min, it is the second most expensive model tested and costs more than five times Gemini 3 Flash (Preview), which outperforms it on redaction. The cost premium is not justified by accuracy gains in this evaluation.

If complete transcript granularity matters more than redaction coverage, 2.5 Pro is the strongest GA-only option. Add a secondary redaction layer for any PII-sensitive content.

Verdict: Not recommended — cost not justified

Gemini 3.1 Pro (Preview) is the only model in this evaluation with correct name consistency throughout — agent names are correctly attributed from first mention to last. It also achieves correct diarization. However, it misses account number redaction, which is a material gap for payment-related call center use cases.

At $0.0096/min, it is the most expensive model tested — more than four times the cost of Gemini 3 Flash (Preview). Given that it fails on redaction while costing more, it does not justify the price premium for typical call center transcription workloads.

The correct name consistency behavior is a genuine differentiator if your pipeline does downstream entity resolution against agent names. Outside that specific requirement, the cost-to-accuracy ratio is unfavorable.

Verdict: Not recommended — diarization inversion

Gemini 3.1 Flash-Lite exhibits the same systemic diarization inversion as Gemini 2.5 Flash-Lite: after the opening speaker turn, labels flip and persist for the remainder of the call. This is not a one-off mislabel — it is a consistent failure that affects every subsequent coaching note, compliance flag, and scored output derived from the transcript.

It also misses account number redaction. At $0.0011/min, it is the second cheapest model, but the diarization failure alone disqualifies it from any production deployment where speaker attribution matters.

The speed advantage — 10.0 seconds latency — is real, but it does not compensate for the reliability gap. Use Gemini 3 Flash (Preview) instead at twice the cost but with correct diarization.

Verdict: Not recommended — diarization inversion and content loss

Gemini 2.5 Flash-Lite is the fastest and cheapest model in this evaluation at 8.7 seconds and $0.0004/min, but it carries the most failures. Diarization is inverted — speaker labels flip after the first turn and persist throughout. It misses account number redaction and partially leaks the card number. It also produced the fewest utterances (38), indicating content is being dropped.

The combination of diarization inversion, redaction failure, and content loss makes 2.5 Flash-Lite unsuitable for any call center transcription pipeline where transcript quality, speaker attribution, or PII handling matter.

This is the model to avoid regardless of cost constraints. Even for non-regulated, low-PII-sensitivity workloads, the content-drop problem means you are working with an incomplete picture of the call.