Gemini 4 Argon is now on LLM Gateway: here’s how it scores on voice tasks

What’s available on LLM Gateway

Gemini 4 Argon is served through Google Vertex, with US, EU, and global routing. Here’s what the Models endpoint reports for it:

• Model string: gemini-4-argon
• Context window: 1,048,576 tokens, with up to 32,768 completion tokens
• Parameters: tools, tool_choice, response_format, stream, max_tokens, and reasoning_effort. There’s no temperature.
• Price: $2.00 per million prompt tokens and $10.00 per million completion tokens with global routing, at provider rates with 0% markup. In-region US or EU routing is 10% higher.

For comparison, Gemini 3.8 Flash lists at $0.75/$3.75 and GPT-6 Luna at $0.10/$0.50.

How Gemini 4 Argon scores on the Voice Standard Bench

The Voice Standard Bench scores each model on four tasks developers run on transcripts: dictation cleanup, information extraction, summarization, and text correction, on real transcripts from 30 seconds to an hour.¹ We compare Gemini 4 Argon with Gemini 3.8 Flash, the previous Gemini Flash release; GPT-6 Luna, a low-cost model that scores near the top of the run; and Qwen3.5 4B Fast, the small open model AssemblyAI hosts for fast rewrite tasks.

Overall score by transcript length

Score and cost per hour of audio, by task

Median response time

Argon leads on dictation, stays within a few points of Gemini 3.8 Flash everywhere else, and trails GPT-6 Luna from 15 minutes up. Qwen3.5 4B Fast is the fastest and cheapest of the four at every length, but it’s built for short rewrites, and its scores drop sharply on longer transcripts.

Where it wins: dictation

Dictation is the task where a user speaks and expects clean, formatted text back: filler words removed, punctuation added, self-corrections resolved. Gemini 4 Argon had a performance score of 100 on 30-second clips and 99 on 2-minute clips, a 99.5 mean and the highest performance score of the 50+ models on the bench.

The models just behind it cost far less. Gemma 4 31B, GPT-6 Luna, and GLM-5.3 Flash all cost less than a tenth of Argon per hour of audio. At high volume that’s a lot to pay for two or three points, but where a single wrong correction matters, such as medical or legal dictation, Argon is worth testing first. It’s quick too: a 2.0-second median response on 30-second clips, against 4.1 seconds for Gemini 3.8 Flash.

If you’d rather not wire up speech-to-text and a cleanup pass yourself, the Dictation API does both in one call: a spoken clip in, finished, formatted text out. Pass an llm_instruction when you need the output in a specific shape, such as a bulleted task list or a chat message.

Gemini 4 Argon vs Gemini 3.8 Flash

Close on performance, faster at every length, and pricier the longer the transcript.

Green favors Argon, pink favors Gemini 3.8 Flash.

The speed gap matters for features where a user waits on the response, and much less for batch jobs.

Gemini 4 Argon vs GPT-6 Luna

Even on short transcripts, behind on long ones, and 14–19× the cost per hour of audio.

Green favors Argon, pink favors GPT-6 Luna.

Meetings, sales calls, and interviews often run 15 minutes or more, and that’s where Luna pulls ahead. Argon’s faster response at those lengths matters more for a live meeting assistant than for a summary sent after the call.

Which model to use

• Highest dictation accuracy: Gemini 4 Argon. It has the top score on the bench and a 2-second median on short clips.
• Dictation at volume: GPT-6 Luna or Gemma 4 31B, within four points of Argon for less than a tenth of the cost.
• Summarization and extraction on long transcripts: GPT-6 Luna. It scores higher than Argon at 15 minutes and an hour, at about a sixteenth of the cost.
• Already on Gemini 3.8 Flash: stay, unless you need responses about twice as fast and can pay 28–45% more on non-dictation tasks.

Results on your own prompts and transcripts can differ, and switching models on LLM Gateway means changing one string.

Get started

Gemini 4 Argon works like every other model on LLM Gateway: one OpenAI-compatible endpoint, your AssemblyAI API key, and the model string.

import requests

API_KEY = "YOUR_API_KEY"

transcript = open("call-transcript.txt").read()

response = requests.post(
    "https://llm-gateway.assemblyai.com/v1/chat/completions",
    headers={"Authorization": API_KEY, "Content-Type": "application/json"},
    json={
        "model": "gemini-4-argon",
        "max_tokens": 1000,
        "messages": [
            {
                "role": "system",
                "content": "Rewrite this dictation as clean, formatted text. "
                           "Drop filler words and resolve self-corrections "
                           "to the speaker's final intent.",
            },
            {"role": "user", "content": transcript},
        ],
    },
)

print(response.json()["choices"][0]["message"]["content"])

Before you switch, run your own prompt over your own transcript on all three models. This loop prints the time and token counts for each, so you can weigh the answers against what they cost:

import time
import requests

API_KEY = "YOUR_API_KEY"
MODELS = ["gemini-4-argon", "gemini-3.8-flash", "gpt-6-luna"]

transcript = open("call-transcript.txt").read()

for model in MODELS:
    start = time.perf_counter()
    r = requests.post(
        "https://llm-gateway.assemblyai.com/v1/chat/completions",
        headers={"Authorization": API_KEY, "Content-Type": "application/json"},
        json={
            "model": model,
            "messages": [
                {"role": "system", "content": "Summarize this call in five bullet points."},
                {"role": "user", "content": transcript},
            ],
        },
    )
    elapsed = time.perf_counter() - start
    usage = r.json()["usage"]
    print(f"{model}: {elapsed:.1f}s, "
          f"{usage['prompt_tokens']} in / {usage['completion_tokens']} out")
    print(r.json()["choices"][0]["message"]["content"], "\n")

Prefer not to write code yet? Upload a file in the Playground, run a prompt over the transcript, and switch models to see how the answers differ.

Frequently asked questions