Now live New models for transcription and text-to-speech. AssemblyAI-STT-3.7 AssemblyAI-TTS-1
Insights & Use Cases

Claude Opus 5.5 vs GPT-6 Astra vs Gemini 4 Argon: which LLM wins on voice data?

Claude Opus 5.5 vs GPT-6 Astra vs Gemini 4 Argon, scored on transcript summaries, extraction, correction and dictation, with cost per audio hour and latency.

Abstract green cone illustration

Written by

Kelsey Foster

Published on

18 November 2026

A year ago we compared Gemini 3 Pro, GPT-5 and Claude 4.5 by reading their meeting notes side by side. Gemini 3 Pro wrote the most executive-ready summaries, GPT-5 dug deepest into follow-ups, and Claude stayed literal and conservative. Those were impressions from a handful of transcripts.

All three families have shipped new flagships since then, and we now score every model on LLM Gateway the same way, on real call transcripts. So this update swaps impressions for numbers: Claude Opus 5.5, GPT-6 Astra and Gemini 4 Argon, scored on summarization, information extraction, text correction and dictation cleanup, with cost per hour of audio and response time.

The short answer, from the Voice Standard Bench run of October 29, 2026:

  • GPT-6 Astra has the highest or tied-highest overall score of the three at every transcript length. It's also the most expensive per hour of audio and runs on the US endpoint only.
  • Claude Opus 5.5 is the best summarizer of the three, ranked #1 of every model measured on 2-minute and 15-minute summaries, at about two-thirds of Astra's cost per audio hour.
  • Gemini 4 Argon has the top dictation score of any model measured and the fastest responses of the three, at about a fifth of Astra's cost. It falls behind on transcripts of 15 minutes and longer.

Most "Claude vs GPT vs Gemini" comparisons rate the chat apps on writing and coding. This one rates the API models on the work developers run on transcripts, so treat it as a shortlist and test the top pick on your own audio.

What changed since Gemini 3 Pro, GPT-5 and Claude 4.5

Each family moved up a full generation. Here's how the models in the original post map to the flagships we benchmarked on LLM Gateway, with release dates from each vendor's own pages:

Family In our November 2025 post Flagship in this comparison Released Newer top model, not in the Oct 29 benchmark run
Anthropic Claude Claude Sonnet 4.5 Claude Opus 5.5 (claude-opus-5-5) Sep 22, 2026 Claude Fable 5.2 (not on LLM Gateway)
OpenAI GPT GPT-5 GPT-6 Astra (gpt-6-astra) Sep 3, 2026 GPT-6.2 Astra (released Nov 17, 2026)
Google Gemini Gemini 3 Pro Gemini 4 Argon (gemini-4-argon) Oct 13, 2026 (GA) Gemini 4 Carbon (gemini-4-carbon, on LLM Gateway since Nov 19, 2026)

Two of the original three are still on LLM Gateway: GPT-5 (gpt-5) and Claude Sonnet 4.5 (claude-sonnet-4-5-20250929). Gemini 3 Pro isn't. Each vendor also has a newer top model than the one compared here. Claude Fable 5.2 isn't on LLM Gateway. GPT-6.2 Astra was released on November 17, 2026, and Gemini 4 Carbon was added to LLM Gateway on November 19, 2026, both after the October 29 benchmark run. So this comparison covers the strongest model from each family with published benchmark scores. The model catalog lists all 55+ models on LLM Gateway with live prices.

Specs: context window, price and regions

Context windows below come from Anthropic, OpenAI and Google. LLM Gateway's own limits come from its public models endpoint and differ in one place: GPT-6 Astra is listed at 1,050,000 tokens by OpenAI and 270,000 tokens on LLM Gateway. An hour-long meeting transcript is far shorter than either limit, so this only matters if you send very long context alongside the transcript.

Spec Claude Opus 5.5 GPT-6 Astra Gemini 4 Argon
Model ID on LLM Gateway claude-opus-5-5 gpt-6-astra gemini-4-argon
Context window (vendor) 1M tokens 1,050,000 tokens 1,048,576 tokens
Context window on LLM Gateway 1,000,000 tokens 270,000 tokens 1,048,576 tokens
Max output on LLM Gateway 128,000 tokens 128,000 tokens 32,768 tokens
Price per 1M tokens (input / output) $4 / $20 $10 / $50 $2 / $10
Cached input read, per 1M tokens $0.20 $1.00 $0.10
LLM Gateway endpoints US and EU US only US and EU
Serving provider AWS Bedrock OpenAI, AWS Bedrock Google Vertex
reasoning_effort supported Yes Yes Yes

Prices are LLM Gateway's catalog rates per million tokens. In-region processing on the US or EU endpoint is 10% higher, per the cloud endpoints and data residency docs. All three models are served by providers that LLM Gateway lists with zero data retention, and none of LLM Gateway's providers train on your data.

Token prices don't tell you what a summary costs, because each model writes a different number of tokens for the same transcript. The benchmark below converts everything to cost per hour of audio, which is the number that maps to your bill.

Compare Claude, GPT and Gemini on your own transcripts

Upload a recording in the Playground, then send the transcript to Claude Opus 5.5, GPT-6 Astra and Gemini 4 Argon with the same prompt and compare the answers side by side.

Try playground

How the benchmark works

Voice Standard Bench is AssemblyAI's open-source (MIT) benchmark for LLMs on transcripts. It runs four tasks on call transcripts of 30 seconds, 2 minutes, 15 minutes and 1 hour:

  • Summarization: a summary of the call or meeting.
  • Information extraction: action items, decisions, participants and entities.
  • Text correction: fixing transcription errors without rewriting what was said.
  • Dictation: applying spoken edits and formatting to a short dictated clip. It's measured on 30-second and 2-minute audio only.

LLM judges score summaries, extraction and correction from 1 to 100 against the source transcript, and dictation is scored against an answer key. Cost per hour of audio is what the task costs on one hour of audio at global pricing. Latency is the end-to-end request time. The prompts, rubrics, full results and harness are in the voice-standard-bench repo, and the Choose a model docs explain how to read the results.

Two caveats from the docs apply here. A gap of 1 or 2 points isn't meaningful, so treat close scores as a tie. And results are in beta, so values can change between runs. Every number in this post is from the October 29, 2026 run.

Overall score by transcript length

The overall score averages the tasks measured at each length. Bold marks the top score of the three.

Transcript length Claude Opus 5.5 GPT-6 Astra Gemini 4 Argon
30 seconds 95 96 94
2 minutes 93 94 94
15 minutes 88 88 82
1 hour 87 88 79

The pattern from the original post still holds in one way: no family wins everywhere. On short audio all three sit within 2 points of each other, which the benchmark treats as a tie. The split opens at 15 minutes. Opus 5.5 and Astra hold 87 to 88 on hour-long meetings, while Argon drops to 79.

If your audio is mostly short calls and dictated notes, pick on cost and speed. If it's meetings, the long-transcript scores decide it.

Scores by task

Ranks are among all models measured in the run, with ties sharing a rank. Scores of the three flagships, at a short and a long length:

Task and length Claude Opus 5.5 GPT-6 Astra Gemini 4 Argon
Summarization, 2 min 99 (#1) 98 (#2) 95 (#10)
Summarization, 1 hour 95 (#2) 95 (#2) 89 (#12)
Information extraction, 2 min 85 (#34) 91 (#3) 91 (#3)
Information extraction, 1 hour 84 (#2) 83 (#4) 75 (#24)
Text correction, 2 min 93 (#2) 93 (#2) 89 (#11)
Text correction, 1 hour 81 (#10) 86 (#2) 75 (#28)
Dictation, 30 sec 99 (#8) 99 (#8) 100 (#1)
Dictation, 2 min 96 (#8) 96 (#8) 99 (#1)

Summarization: Claude Opus 5.5

Opus 5.5 is the strongest summarizer of the three. It ranks #1 of every model measured on 2-minute and 15-minute summaries (99 and 97) and stays at 95 on hour-long meetings, tied with Astra. Argon trails by 4 to 7 points from 2 minutes up. If summaries are the product, as in an AI notetaker, start with Opus 5.5. For the full field, including smaller models that come close, see our best LLM for summarizing transcripts picks.

Information extraction: depends on length

On 2-minute calls, Astra and Argon tie at 91 and Opus 5.5 trails at 85, ranked #34. On hour-long meetings the order flips: Opus 5.5 scores 84 (#2), Astra 83, and Argon falls to 75. Extraction is the task where the three disagree most, so test on your own schema. Our guide to extracting structured data from transcripts covers JSON output and the models that lead at each length.

Text correction: GPT-6 Astra

Astra is the most consistent corrector. It ties Opus 5.5 at 93 on 2-minute transcripts and leads the three on long audio, with 91 at 15 minutes (#1 of all models) and 86 at an hour. Opus 5.5 drops to 81 on hour-long transcripts and Argon to 75. Correction can go wrong by rewriting what the speaker said, which best LLM for transcript correction covers in detail.

Dictation: Gemini 4 Argon

Argon has the top dictation score of every model measured: 100 on 30-second clips and 99 on 2-minute clips. Opus 5.5 and Astra both score 99 and 96. Argon is also the fastest of the three, which matters most here because a user is waiting on the result. Our Gemini 4 Argon launch post has the full dictation breakdown.

See every model on Voice Standard Bench

Compare the LLM Gateway models on each task and transcript length, with score, cost per hour of audio and latency for each.

See the benchmark

Cost per hour of audio and response time

Metric Claude Opus 5.5 GPT-6 Astra Gemini 4 Argon
Cost per hour of audio, 2-min calls $0.27 $0.37 $0.073
Cost per hour of audio, 1-hour meetings $0.24 $0.35 $0.073
Median response, 2-min calls 5.4 s 3.8 s 2.6 s
p95 response, 2-min calls 36.4 s 10.7 s 4.7 s
Median response, 1-hour meetings 31.5 s 24.1 s 10.9 s
p95 response, 1-hour meetings 284.5 s 206.4 s 108.8 s

Argon runs at about $0.07 per hour of audio at both lengths, roughly a fifth of Astra's cost and under a third of Opus 5.5's. Astra is the most expensive of the three. Opus 5.5 sits between them, at about two-thirds of Astra's cost.

Response time splits the same way. Argon has the lowest median and the lowest p95 at both lengths. Opus 5.5 has the slowest tail: one in 20 requests on 2-minute calls took 36 seconds or longer, and on hour-long meetings the p95 is close to five minutes. That works for a summary emailed after a meeting and is too slow for a live assistant. The docs recommend checking p95 as well as the median for anything real-time.

How much better are the new models?

GPT-5 and Claude Sonnet 4.5 are still on LLM Gateway, so they ran in the same benchmark. Gemini 3 Pro isn't, so Gemini 2.5 Pro stands in for the older Gemini generation:

Model Overall, 2 min Overall, 1 hour Median response, 2 min Cost per audio hour, 1 hour
Claude Sonnet 4.5 86 75 3.5 s $0.11
Claude Opus 5.5 93 87 5.4 s $0.24
GPT-5 93 81 17.8 s $0.11
GPT-6 Astra 94 88 3.8 s $0.35
Gemini 2.5 Pro (Gemini 3 Pro is no longer on LLM Gateway) 89 77 11.5 s $0.11
Gemini 4 Argon 94 79 2.6 s $0.073

The biggest gains are on long audio. Claude went from 75 to 87 on hour-long meetings and GPT from 81 to 88. GPT-6 Astra also answers 2-minute calls in a median 3.8 seconds against 17.8 seconds for GPT-5. Opus 5.5 and Astra cost more per audio hour than the models they replaced, while Argon has a lower cost per audio hour than Gemini 2.5 Pro. The next section shows that a smaller model from the same family often gets close to its flagship.

The rest of each family

A flagship isn't always the right pick. These are the other current models from each family on LLM Gateway, from the same run:

Model Family tier Overall, 2 min Overall, 1 hour Cost per audio hour, 1 hour
Claude Sonnet 5.5 Claude mid 90 85 $0.15
Claude Haiku 5.5 Claude small 87 79 $0.0053
GPT-6.1 Sol GPT mid 93 87 $0.090
GPT-6 Luna GPT small 92 85 $0.0039
Gemini 3.8 Flash Gemini Flash 92 75 $0.040
Gemini 4 Hydrogen Gemini everyday Not in the Oct 29 run Not in the Oct 29 run Not measured ($1.50 / $7.50 per 1M tokens)
Gemini 4 Neon Gemini 4 Not in the Oct 29 run Not in the Oct 29 run Not measured ($3 / $15 per 1M tokens)
Gemini 4 Carbon Gemini top model Not in the Oct 29 run Not in the Oct 29 run Not measured ($4 / $20 per 1M tokens)

GPT-6.1 Sol scores within 1 point of Astra on hour-long meetings at about a quarter of the cost per audio hour, and GPT-6 Luna is within 3 points at about 1% of it. Claude Sonnet 5.5 is 2 points behind Opus 5.5 on hour-long audio at about 60% of the cost. On Gemini's side, Gemini 4 Hydrogen (added to LLM Gateway on November 13) and Gemini 4 Neon and Gemini 4 Carbon (both added on November 19) weren't in the October 29 run, so check the Voice Standard Bench results for their scores once they're published.

Which one should you use?

If you need Start with Why
Summaries of calls and meetings Claude Opus 5.5 Highest or tied-highest summarization score of the three at every length, #1 of all measured models at 2 and 15 minutes, and it runs on the EU endpoint
One model for several tasks on long recordings GPT-6 Astra Highest or tied-highest overall score at every length, and the strongest long-transcript correction score of the three
Dictation cleanup or a user waiting on the answer Gemini 4 Argon Top dictation score of every model measured, the fastest median and p95 of the three, and about a fifth of GPT-6 Astra's cost per audio hour
EU data residency Claude Opus 5.5 or Gemini 4 Argon GPT-6 Astra is served from the US endpoint only
High volume at the lowest cost A smaller model from the same family GPT-6 Luna and GPT-6.1 Sol score within 3 points of Astra on 1-hour audio

Whichever you choose, set a runner-up as a fallback, and transcribe with a strong speech-to-text model first. An LLM can only work with the words in the transcript, so errors in speech-to-text carry through to the summary.

Run all three through one API

LLM Gateway is OpenAI-compatible, so you can call all three models with the OpenAI SDK by changing the base URL and the model string. The examples below send the same transcript prompt to each model, time the response, and set a fallback model so a provider error doesn't stop the run.

Python:

import os
import time

from openai import OpenAI

MODELS = ["claude-opus-5-5", "gpt-6-astra", "gemini-4-argon"]
FALLBACKS = {
    "claude-opus-5-5": "claude-sonnet-5-5",
    "gpt-6-astra": "gpt-6.1-sol",
    "gemini-4-argon": "gemini-3.8-flash",
}

client = OpenAI(
    base_url="https://llm-gateway.assemblyai.com/v1",
    api_key=os.environ["ASSEMBLYAI_API_KEY"],
)

transcript = open("sample_transcript.txt", encoding="utf-8").read()
prompt = (
    "Summarize this call in 5 bullet points, then list every action item "
    f"with its owner.\n\nTranscript:\n{transcript}"
)

for model in MODELS:
    start = time.perf_counter()
    response = client.chat.completions.create(
        model=model,
        messages=[{"role": "user", "content": prompt}],
        max_tokens=1500,
        # LLM Gateway field, not part of the OpenAI schema
        extra_body={"fallbacks": [{"model": FALLBACKS[model]}]},
    )
    seconds = time.perf_counter() - start
    print(f"=== {model}  {seconds:.2f}s  {response.usage.total_tokens} tokens")
    print(response.choices[0].message.content, "\n")

JavaScript:

import { readFile } from "node:fs/promises";
import OpenAI from "openai";

const MODELS = ["claude-opus-5-5", "gpt-6-astra", "gemini-4-argon"];
const FALLBACKS = {
  "claude-opus-5-5": "claude-sonnet-5-5",
  "gpt-6-astra": "gpt-6.1-sol",
  "gemini-4-argon": "gemini-3.8-flash",
};

const client = new OpenAI({
  baseURL: "https://llm-gateway.assemblyai.com/v1",
  apiKey: process.env.ASSEMBLYAI_API_KEY,
});

const transcript = await readFile("sample_transcript.txt", "utf8");
const prompt =
  "Summarize this call in 5 bullet points, then list every action item " +
  `with its owner.\n\nTranscript:\n${transcript}`;

for (const model of MODELS) {
  const start = performance.now();
  const response = await client.chat.completions.create({
    model,
    messages: [{ role: "user", content: prompt }],
    max_tokens: 1500,
    fallbacks: [{ model: FALLBACKS[model] }], // LLM Gateway field, sent in the request body
  });
  const seconds = ((performance.now() - start) / 1000).toFixed(2);
  console.log(`=== ${model}  ${seconds}s  ${response.usage?.total_tokens} tokens`);
  console.log(response.choices[0].message.content, "\n");
}

curl:

curl -s https://llm-gateway.assemblyai.com/v1/chat/completions \
  -H "Authorization: $ASSEMBLYAI_API_KEY" \
  -H "Content-Type: application/json" \
  -d '{
    "model": "claude-opus-5-5",
    "messages": [{"role": "user", "content": "Summarize this call in 5 bullet points:\n\nAgent: Thanks for calling. Customer: My invoice is wrong..."}],
    "max_tokens": 1500,
    "fallbacks": [{"model": "claude-sonnet-5-5"}]
  }'

A few things to know before you run it:

  • EU data residency. Point the base URL at https://llm-gateway.eu.assemblyai.com/v1. The EU endpoint serves Claude and Gemini models, so gpt-6-astra won't run there.
  • Transcripts by ID. If you transcribe with AssemblyAI, you can pass a transcript_id and put {{ transcript }} in your prompt instead of pasting the text. LLM Gateway fetches the transcript for you.
  • Reasoning effort. All three models accept reasoning_effort. The benchmark scores above were measured without it, so a lower setting may trade a little quality for speed.

The full script, a Node.js version and a sample transcript are in the companion repo. For a production version with chunking for long meetings and a quality check, see how to summarize meetings with AI in Python.

Conclusion

The 2025 version of this post ended with "different models serve different workflows," and the numbers now show where. GPT-6 Astra is the safest single pick if you can pay for it and process in the US. Claude Opus 5.5 writes the best summaries and runs in the EU. Gemini 4 Argon is the one to use when a person is waiting, and for dictation.

The bigger change since 2025 is that a smaller model from the same family often lands within a few points of its flagship. GPT-6 Luna scores within 3 points of Astra on hour-long meetings at about 1% of the cost per audio hour. So run the flagship and its smaller sibling on a sample of your own transcripts before you commit. Our transcripts aren't your transcripts, and a 2-point gap on the benchmark can go either way on your data.

Run all three models with one API key

Sign up free to get an AssemblyAI API key, then call Claude, GPT and Gemini models through LLM Gateway by changing one model string.

Sign up free

Frequently Asked Questions (FAQs)

Which is better for transcripts, Claude, GPT or Gemini?

It depends on the task and the audio length. In the October 29, 2026 Voice Standard Bench run, GPT-6 Astra had the highest or tied-highest overall score at every length, Claude Opus 5.5 wrote the best summaries, and Gemini 4 Argon led on dictation and speed. On short calls all three scored within 2 points of each other.

Which is more accurate, Gemini or GPT?

On short audio they're tied: Gemini 4 Argon and GPT-6 Astra both scored 94 overall on 2-minute calls. On hour-long meetings GPT-6 Astra scored 88 and Gemini 4 Argon 79, with the biggest gaps on extraction and text correction. Gemini 4 Argon scored higher on dictation, at 99 against 96 on 2-minute clips.

Is there one API for Claude, GPT and Gemini?

Yes. AssemblyAI's LLM Gateway serves Claude, GPT and Gemini models, along with open models, through one OpenAI-compatible endpoint and one API key. You switch models by changing the model string, and you can set a fallback model on each request. The model catalog lists all 55+ models with their prices.

What does each model cost per hour of audio?

In the October 29, 2026 run, the benchmark tasks on an hour-long meeting cost about $0.35 per audio hour with GPT-6 Astra, $0.24 with Claude Opus 5.5 and $0.07 with Gemini 4 Argon. These are benchmark estimates at global pricing. In-region processing on the US or EU endpoint is 10% higher, and your own prompts and output length will change the total.

What happened to Gemini 3 Pro, GPT-5 and Claude Sonnet 4.5?

GPT-5 and Claude Sonnet 4.5 are still available on LLM Gateway, and Gemini 3 Pro isn't. Their successors score higher on long audio: Claude Opus 5.5 scored 87 on hour-long meetings against 75 for Claude Sonnet 4.5, and GPT-6 Astra scored 88 against 81 for GPT-5. If you still run one of the older models, the newer one in the same family is the first thing to test.

Which providers offer these models with zero data retention?

On LLM Gateway, Claude Opus 5.5 runs on AWS Bedrock, GPT-6 Astra on OpenAI and AWS Bedrock, and Gemini 4 Argon on Google Vertex. LLM Gateway lists all three providers with zero data retention, on by default. None of LLM Gateway's providers train on your data.