Gemini 4 Carbon vs Argon vs Claude Opus 5.5 vs GPT-6 Astra

Gemini 4 Carbon is not the best model at everything. It is first on Artificial Analysis's v5.0 index with 59 points, but a closer look at the tables shows this:

• On the coding agent index, Claude Code running Opus 5.5 scores 62 and Carbon 59.
• On FrontierSWE 2, GPT-6 Astra scores 65.5%, Opus 5.5 62.3% and Carbon 61.2%.
• On OSWorld-2.1, Opus 5.5 scores 87.2% and Carbon 81.5%.
• In Artificial Analysis's new measurement, the hallucination rate is 4% for Argon and 11% for Carbon.
• You cannot try any of this with Carbon today; access is open only to selected cyber defenders. I covered when it might arrive in a separate post.

For the model itself, start with What Is Gemini 4 Carbon?.

This post is a compilation; I have not tried all four models on the same tasks, and I cannot reach Carbon at all. The numbers come from Google's announcement table, the Artificial Analysis and Vals pages, and news sites; independent reproduction is limited. Sources are at the end.

Side-by-side table

The table in Google's announcement is an image; since I could not reach it as text, I used the version relayed by Eden AI. The Artificial Analysis index, FrontierSWE 2, OSWorld-2.1 and the cybersecurity rows match Google's text and other write-ups. I took the other rows only from Eden AI, so read them as values relayed by one write-up.

Benchmark	Carbon	Argon	Claude Opus 5.5	GPT-6 Astra
Artificial Analysis Intelligence Index v5.0 (points)	59	50	58	57
Vals Index v2.2	70.4%	69.2%	70.2%	65.0%
FrontierSWE 2	61.2%	55.0%	62.3%	65.5%
Terminal-Bench 4.0	65.2%	57.1%	59.6%	59.1%
OSWorld-2.1 (offline subset)	81.5%	74.7%	87.2%	80.9%
AutomationBench	53.9%	51.3%	42.5%	41.4%
Blueprint-Bench 2	58.6%	54.4%	51.2%	49.7%
GraphWalks (256K - 1M tokens)	96.2%	84.2%	66.8%	71.8%
RiemannBench	80.0%	76.0%	69.6%	72.0%
CWE-bench (patching)	71.0%	68.0%	67.0%	68.0%
ExploitGym	54.8%	42.0%	34.5%	42.4%
SEC-bench Pro	88.2%	83.4%	n/a	85.4%

"n/a" means the table has no value for that model.

Four things to keep in mind when reading the table:

• Do not mix it with the old Argon table. In the Argon comparison I wrote 66.4% for Opus 5.5 on Terminal-Bench and 67.0% on Vals. In the Carbon-day table the same benchmarks show 59.6% and 70.2%. Benchmark versions and run conditions differ; read each table on its own.
• According to Eden AI, Carbon leads on 15 of the table's 17 benchmarks. The other two are FrontierSWE 2 (Astra) and OSWorld-2.1 (Opus 5.5).
• The cyber benchmarks ran without guardrails. Google says so explicitly. The version that eventually ships to everyone need not produce the same results; that is my interpretation.
• ExploitGym has a time-limit difference. According to Handy's review, Carbon's headline result used a 10-hour cap while 6 hours was used for Opus 5.5; at 6 hours Carbon scores 53.5%. Read this as a value reported by a single source (relayed by NeuralTrust).

Independent measurements

Measure	Carbon	Note
Artificial Analysis Intelligence Index v5.0	59, first	Opus 5.5 58, Astra 57, Argon 50 (Dealroom, iWeaver)
Artificial Analysis Coding Agent Index	59	Claude Code with Opus 5.5 scores 62; Carbon was tested in only one harness (Antigravity CLI)
Vals Index	70.36% ± 0.97, second	Opus 5.5 70.18%; Claude Sonnet 5.5 at 70.85% is above Carbon
Vals Finance Index	66.35%, first
Vals Legal Index	75.83%, third
Terminal-Bench 4.0 on Vals	63.13%, fourth	Lower than the 65.2% in Google's table

On Vals the gap between Carbon and Opus 5.5 (70.36% and 70.18%) is far smaller than the margin of error (±0.97); in practice they are level. On agentic coding the only independent data is Artificial Analysis's index, and Claude Code leads there; but they tested Carbon in a single harness.

Hallucination, speed and verbosity

These are the least talked-about sides of Carbon.

Hallucination rate. In October Artificial Analysis updated its AA-Omniscience benchmark to v1.1. On its site Carbon's net score is 54 (Opus 5.5: 48), but the hallucination rate is 11% for Carbon and 4% for Argon. This rate is how often the model gives a wrong answer instead of saying "I don't know" on a question it does not know; lower is better. According to iWeaver, Carbon's accuracy went up, so it answers more questions correctly but goes to a wrong answer more often than Argon when it does not know. I could not find this rate for Opus 5.5 and Astra on the page.

Google's claim that "unsupported claims roughly halved" is about internal measurement in coding environments; Artificial Analysis's is on knowledge questions. The two measurements need not contradict each other, but do not generalize to "Carbon is more honest than Argon in every respect." (The roughly 15% I cited for Argon in my Argon post belongs to the older version of the benchmark; in the new version Argon is at 4%.)

Speed and verbosity. According to Artificial Analysis, Carbon generates 84.6 tokens per second (94th of 196 models; the median for its price tier is 80.5). Time to first token on Google's API is 118.6 seconds, which the page describes as "at the higher end." It spends 200 million output tokens to complete the index, against a median of 70 million. Because output tokens are the expensive ones, that translates directly into cost.

Cost per task. Artificial Analysis's cost per index task, as NeuralTrust relays it: Carbon $7.76, Argon $7.68, Opus 5.5 $10.04, Astra $6.25. So Carbon costs no more than Argon, but more than Astra.

Price

According to Eden AI's price list (check the official pages): Carbon in the intro period is $2 per million input tokens and $10 for output, then $4 and $20; Argon $2 and $10 ($4 and $20 after December 1); Opus 5.5 $4 and $20; Astra $10 and $50. Carbon cannot be used at this price yet; the model is not in the general API.

Which to pick, and when

This section is my interpretation; general conclusions I draw from the numbers above.

• Models you can reach today: Argon, Opus 5.5 and Astra. Carbon is not on this list.
• Very long documents or codebases: Carbon is far ahead on GraphWalks (96.2%); Argon (84.2%) is also above Astra and Opus.
• Tool-using agent loops: Mixed. In Google's table Carbon leads on Terminal-Bench, in Vals's own run it is fourth, and on Artificial Analysis's agent index Opus 5.5 leads. Try your own task.
• Hard mixed science and coding: Astra.
• Computer use (OSWorld): Opus 5.5.
• If not making things up when it doesn't know matters: Argon (4% hallucination rate).
• Cyber defense: Carbon leads on the numbers, but they were measured without guardrails and there is no access.
• Cost: Per task Astra is cheapest and Opus 5.5 the most expensive; Carbon is at Argon's level.

How to read benchmarks

1. A maker's table is made of the maker's chosen benchmarks. The Carbon-day table is full of rows where Carbon leads; the two rows where it trails are in it too, but the hallucination rate, speed and verbosity are on third-party pages.
2. The agent framework changes the result. Artificial Analysis measured Carbon only in Antigravity CLI and Opus 5.5 in Claude Code. I explained this in What Is an AI Harness? The DeepSeek Example.
3. Measure your own task. Five tasks of your own are more instructive than fifty of someone else's.

Frequently Asked Questions

Is Gemini 4 Carbon better than Claude Opus 5.5?

One point ahead on the Artificial Analysis index (59 to 58), level within the margin of error on Vals. Opus 5.5 leads on the coding agent index and on OSWorld-2.1. "Better" depends on the type of work.

How much better is Carbon than Argon?

Ahead of Argon on all 12 rows in the table and nine points higher on the index (59 to 50). In return, according to Artificial Analysis, the hallucination rate rose from 4% to 11%.

Which is better for writing code?

No single answer: Carbon for long-context codebases, Opus 5.5 or Carbon for agents that work in the terminal (it varies by measurement), Astra for hard mixed tasks. Also, you cannot reach Carbon right now.

Does Carbon make things up more?

In Artificial Analysis's measurement on knowledge questions, its rate of answering wrongly when it does not know is higher than Argon's (11% versus 4%). Google, for its part, says unsupported claims halved in coding environments. The two measure different things.

Which is cheaper?

By Artificial Analysis's cost per index task, Astra ($6.25) is cheapest and Opus 5.5 ($10.04) the most expensive; Carbon ($7.76) and Argon ($7.68) are in the middle.

Sources

• Google, Gemini 4 Carbon: our next frontier model (October 28, 2026): the numbers given as text, the note on measurement without guardrails.
• Eden AI, Gemini 4 Carbon Beats Opus 5.5 and GPT-6 Astra: The Catch: relay of Google's table and the price list.
• Artificial Analysis, Gemini 4 Carbon: index, speed, verbosity, cost per task.
• Artificial Analysis, AA-Omniscience: net score and hallucination rates (v1.1).
• Vals AI, Gemini 4 Carbon: Vals Index, finance, legal and Terminal-Bench results.
• NeuralTrust, Gemini 4 Carbon Benchmarks, Pricing: coding agent index, cost per task, the ExploitGym time-limit note.
• iWeaver, Gemini 4 Carbon: Benchmarks, Price & Availability: index ranking, change in hallucination rate.
• Dealroom, Google DeepMind's Gemini 4 Carbon takes the #1 spot on Artificial Analysis: index scores.
• Tech My Money, Gemini 4 Carbon Is Google's New Frontier Model: FrontierSWE 2 and OSWorld-2.1 values.