Stop putting the smartest model on the phone. Put it after the call.

When a frontier model drops, voice-AI teams do the same thing: swap it into the live phone loop and hope callers notice a smarter agent. When Gemini 4 Carbon came out on November 10, 2026, we skipped the blind swap. Instead, we ran Carbon and its sibling, Gemini 4 Argon, against our internal benchmarks in the two days after launch.

What we found goes against the usual instinct. Carbon is not a meaningfully better live listener than Gemini 3.8 Flash, and it is about twice as slow. Where it earns its keep is after the call ends: writing the transcript of record, admitting when audio is too garbled to hear, and reviewing long conversations.

Every number in this post comes from real phone calls to voice agents. Transcripts were labeled by hand by a person who listened to each turn; agent replies were graded by a two-judge model panel. We publish only aggregate scores: no caller audio, words, or anything that could identify a person or business.

We had Gemini 4 Carbon draft this post from our results, then checked every number against the runs.

1. On the live path, twice the wait for the same ears

Carbon heard hard caller turns about as well as Gemini 3.8 Flash, and took twice as long to reply.

To test live hearing, we pulled 59 caller turns that were hard to hear: background noise, or transcription services that disagreed about what was said. Each model listened to the raw audio and spoke next as the agent. Blind judges checked whether the reply showed the model heard the caller right.

Gemini 4 Carbon got 45 of the 59 turns right, compared to 42 for Gemini 3.8 Flash. On a sample that size, a three-turn gap is basically a wash.

The wait is not a wash. Carbon's median reply took 4.7 seconds, about twice Flash's 2.4 seconds. On the slowest 1 in 10 turns, Carbon took 10.0 seconds versus Flash's 4.0. Ten seconds is a long silence on a phone call.

2. More thinking is not more hearing

On turns the fast model got wrong, more thinking barely helps. On transcripts, high thinking is best when it answers, and returns nothing on about 1 turn in 5.

Reasoning models spend "thinking tokens" (hidden scratchpad words written before answering) to work through hard problems. It is tempting to assume more thinking will rescue muddy audio. We tested that on 36 real caller turns that Flash had previously missed. Run again, Flash got 8 right in 2.8 seconds. Carbon at low thinking got 8 right in 5.4 seconds. Medium thinking still got 8 right, in 13.4 seconds. High thinking nudged that to 10 right, with a median reply of 64 seconds.

On verbatim transcription, high thinking backfires. When Carbon at high thinking produces an answer, it is our most accurate setting: 90.5% of the 474 turns it answered were right for the call, versus 85.2% for Flash. But it returns nothing at all on about 1 in 5 turns (160 of 698). It spends 3,337 thinking tokens on average and, on the hardest tenth of turns, runs into its 9,216-token ceiling before writing a single word.

That failure hits hardest on numbers, where Carbon at high thinking drops to 48% right for the call versus 86% for Flash, pulling its overall score across all 594 graded turns down to 72.4%. It also costs about 240 times as much per turn: 3.41 cents, against 0.014 cents for Flash and 0.077 cents for Carbon at low thinking. Argon at high thinking avoids the cliff (588 thinking tokens on average, 5 empty turns, 0.66 cents per turn), though its overall accuracy ties Argon at low thinking.

3. After the call, it writes a cleaner transcript of record

At low thinking, with no stopwatch running, the Gemini 4 models beat 3.8 Flash as transcribers: fewer word errors, more turns right for the call.

Across 594 whole caller turns transcribed by hand, we graded two ways: word error rate (the share of words inserted, deleted, or swapped) and "right for the call" (whether the transcript keeps the meaning a phone agent needs).

Carbon at low thinking cut the word error rate from 11.7% on Flash to 8.7% (Argon high 8.4%, Argon low 9.7%). On exact word-for-word matches, Carbon low beat Flash 75.8% to 72.1%, better on 62 turns and worse on 40, a gap unlikely to be luck (p = 0.04). Argon low reached 76.1% and Argon high 77.3%. On right for the call, Carbon low led Flash 84.8% to 82.5% (better on 41 turns, worse on 27; p = 0.11), and Argon low reached 85.2%.

Two caveats. First, nearly all of Carbon's gain is on ordinary open-ended sentences, where it leads Flash 85% to 77% (15 better, 4 worse). On structured details it barely moves: short answers like "yes" or a date (92% vs 90%), numbers (87% vs 86%), names (79% vs 76%) and spelled-out letters (81% vs 81%). On email addresses it trails (52% vs 58%). Second, the gain sits in one half of the set: on the first 353 turns Carbon low beat Flash clearly, 85.0% to 80.2% (p = 0.01); on the other 241 they tied, 84.6% to 85.9%.

4. It admits when it can't hear instead of making words up

On turns nobody could make out, the Gemini 4 models leave them blank far more often, and when they mark something [inaudible], a person usually couldn't hear it either.

The hardest habit to break in speech models is confident invention: clean text out of static or crosstalk. Our transcription set has 17 turns so garbled that the person labeling them could not make out the caller at all.

Gemini 3.8 Flash wrote words on 16 of those 17 turns. The Gemini 4 models held back far more often: Carbon low left 7 of 17 blank or marked them [inaudible], Argon low 6, Carbon high 8, and Argon high 11.

Just as important, when a Gemini 4 model marks part of a turn [inaudible], the flag is trustworthy: the person labeling agreed 5 of 6 times for Carbon low and 4 of 4 for Argon high, against 3 of 7 for Flash. For a team that reviews calls, a model that raises its hand is worth more than one that always fills the box.

5. Instruction following: a strong second opinion on long calls

Carbon is behind the best current models overall, but holds up deep into long conversations, and slips on wrapping up calls and using tools.

Hearing is half of a phone agent's job; following a business's rules is the other half. We tested 853 moments from real and synthetic agent conversations. Each model read the agent's full instructions, its tools (actions like looking up an order) and the conversation so far, then wrote the next reply. A good reply has no critical or major mistake.

Overall, Gemini 4 Carbon wrote a good reply 88.7% of the time (26 critical mistakes) and Argon 87.9% (35). That beats GPT-5.6 Terra, an earlier-generation model, at 83.4% (43 critical mistakes). Carbon trails GPT-6.1 Terra (93.2%) and GPT-6.1 Sol (94.5%), with 13 critical mistakes each.

Split by conversation length, the picture changes. On the 71 moments deep in long conversations, Carbon reached 91.5%, just ahead of GPT-6.1 Terra (90.1%) and well ahead of GPT-5.6 Terra (78.9%). It also held up on prompts over 8,000 tokens (89.8% vs 92.9% for GPT-6.1 Terra). Where it stumbled was wrapping up calls (79.6% vs 97.1% for GPT-6.1 Terra, 103 moments) and using tools (79.3% vs 92.7%, 164 moments), with a median of 3.95 seconds to its first word (0.82 seconds for GPT-6.1 Terra). That makes Carbon a poor live speaker and a good offline reviewer of whether a long call drifted off-script.

Where we would use Gemini 4, and where we wouldn't

If you build phone agents, think twice before wiring Gemini 4 Carbon into the live turn-by-turn loop. Put it to work after hang-up instead:

• Carbon low or Argon low for the post-call transcript of record. At about 0.08 cents per turn you get a 3-point lower word error rate and an 8-point lift on conversational sentences. (Carbon is $2 in / $10 out per million tokens until November 30, then $4 / $20; Gemini 3.8 Flash is $0.75 / $3.75.)
• Argon high or Carbon low to flag unhearable audio for a person to check. When these models leave a turn blank or write [inaudible], a person usually can't hear it either.
• Carbon as an offline second opinion on long calls. Its 91.5% on long conversations makes it a good check on whether the live agent lost the plot twenty turns in.
• Not on the live reply path. A 4.7-second median audio reply (10.0 seconds for the slowest 1 in 10), nearly 4 seconds to its first word on text, and under 80% on tool use and wrap-ups would frustrate callers.
• Not at high thinking for transcripts. Paying 3.41 cents a turn for blank output on 23% of turns, and 48% on numbers, is worse than Carbon at low thinking.

How we tested

We ran four tests in the two days after Gemini 4 Carbon launched on November 10, 2026:

• Live hard turns (59 turns): real caller audio that was noisy or that transcription services disagreed on. Each model heard the audio and replied as the agent; blind judges graded whether the reply showed it heard the caller right. We also reran 36 turns Flash had missed at each thinking level.
• Verbatim transcripts (594 graded turns; 698 in the high-thinking empty-rate check): whole caller turns from real calls across many businesses, each transcribed by hand by a person who listened to it. Each model got the turn audio plus the agent's previous line.
• Unhearable audio (17 turns): turns in the transcription set the person labeling could not make out at all, plus a check on how often [inaudible] marks agree with the label.
• Instruction following (853 moments): real and synthetic conversations with the agent's full instructions and tools, graded by two independent model judges with a defense pass (a second look before a mistake is confirmed).

Caveats: every score is one sample per item. Speed was measured on launch day, and serving often gets faster after a release. The instruction-following judges are other companies' models, which can carry style preferences. And while Carbon low's exact-match transcript gain holds overall (p = 0.04), its right-for-the-call gain came from one half of the set (p = 0.01), with a tie on the other half.