Gemini 4 Carbon is Argon with more thinking per turn. Google documents native long decode up to 1,048,576 tokens minus input, while AIHubMix documents all four protocol surfaces and a 65,536-token max output; keep gateway requests at or below that cap. Google charges the same introductory rate ($2/M input, $10/M output through November 30, 2026; $4/$20 from December 1), but Carbon spends more tokens and waits longer before its first answer.
For queue-based coding agents, the trade can pay off. Artificial Analysis's 2026-11-11T05:23Z snapshot measured Carbon High at 60.03 (#4 of 104), $7.76 per task and 25.6 minutes, versus Argon High at 50.41, $9.95 and 29.0 minutes. First token was slower: 67.1 seconds versus 40.1.
What changes in an Argon migration
Carbon is based on Argon. The model ID is gemini-4-carbon. Google documents a 1,048,576-token context window, low/medium/high thinking with medium as the default, and a 65,536-token standard output cap. Google's native long-decode path can reach up to 1,048,576 tokens minus the input, with 131,072 and 262,144 as examples. AIHubMix documents all four protocol surfaces and a 65,536-token max output, so keep gateway requests at or below that cap until a larger path is documented. Google says upgrading from Argon requires no request schema changes.
Measurements that matter
Does Carbon deliver a higher AA index for less money per completed task? Plotting quality against task cost makes the four configurations easier to compare.
AA index versus USD per task. Source: Artificial Analysis, November 11, 2026. Costs use standard list prices of $4/M input and $20/M output; exact values follow in the table.
Carbon High sits above and to the left of Argon High: a higher index at lower task cost. Medium is the cheapest Carbon setting in this run; Low is a latency choice rather than the lowest-cost choice. These are AA task costs, not AIHubMix request bills.
| Model | AA index | $/task | Min/task | First token |
|---|---|---|---|---|
| Carbon High | 60.03 | 7.76 | 25.6 | 67.1 s |
| Carbon Medium | 58.26 | 6.59 | 23.3 | 23.2 s |
| Carbon Low | 54.22 | 7.19 | 22.6 | 12.9 s |
| Argon High | 50.41 | 9.95 | 29.0 | 40.1 s |
AA uses standard $4/$20 prices. Carbon Low costs more per task than Medium in this snapshot, so treat Low as a latency setting rather than a discount setting.
The Coding Agent Index v2 puts Antigravity CLI + Carbon High at #4 (59.4, $7.74/task, 29.8 minutes) and Argon High at #11 (53.2, $11.73/task, 34.7 minutes). Fable 5.2 is not on that index.
When to pick which
Carbon fits vendor-native batch or background agent jobs, quality-sensitive work and loops that can wait for a longer first token. Background mode is listed as a capability, but its native request fields are not documented. Argon fits interactive coding, autocomplete and any product with a tight latency budget. Test the same 20–50 real requests through both and record first-token time and completed-task cost.
AIHubMix request shape
AIHubMix documents Carbon on Chat Completions, Messages, Responses and native Gemini, with text, image, audio, video and PDF input. What remains unconfirmed is parameter-level: reasoning_effort mapping, forced tool calls and the native long-decode path. Keep max_tokens at or below 65,536; no live request was run for this post.
import os
from openai import OpenAI
client = OpenAI(base_url="https://aihubmix.com/v1", api_key=os.environ["AIHUBMIX_API_KEY"])
response = client.chat.completions.create(
model="gemini-4-carbon",
messages=[{"role": "user", "content": "Summarise this incident report and list follow-ups."}],
max_tokens=16384,
)
print(response.choices[0].message.content)
Keep reading: the Gemini 4 Carbon series
- Migration errors and timeout settings: Gemini 4 Carbon API Migration: 400 Errors, Timeouts and Output Limits
- Agentic coding comparison: Gemini 4 Carbon vs Fable 5.2: Which One for Agentic Coding?

Top comments (0)