Gemini 4 Carbon nine days on: top of the index, a third of Claude's task cost, and still no API

Google announced Gemini 4 Carbon on October 28, about four weeks after Gemini 4 Argon, calling it its strongest model yet for cybersecurity and long-running work. Nine days later there is enough independent data to judge the claim, and it largely holds, with exceptions Google’s launch did not stress. Carbon is at or near the frontier and much cheaper to run than its rivals on one new third-party benchmark. It is also unavailable to almost everyone: as of early November it was not on the Gemini API pricing page, and we found no announcement of wider access through November 6.

This post is a catch-up, since the launch fell inside a stretch when this site was not publishing. What is new since launch day is the third-party evidence.

Where Carbon leads, and where it does not

Carbon scores 59 on the Artificial Analysis Intelligence Index v5.0, ahead of Claude Opus 5.5 at 58 and GPT-6 Astra at 57. Argon sits at 50, so nine points came in one month. VentureBeat counts Carbon as highest in 15 of the 17 rows in Google’s comparison table, a table that mixes self-computed scores, public leaderboards and rivals’ self-reported figures.

Benchmark	Argon	Carbon	Best rival shown
AA Intelligence Index v5.0	50	59	Opus 5.5: 58
FrontierSWE 2	55.0%	61.2%	GPT-6 Astra: 65.5%
OSWorld-2.1 (offline)	74.7%	81.5%	Opus 5.5: 87.2%
GraphWalks	84.2%	96.2%	GPT-6 Astra: 71.8%
SWE-sweep	4.3%	9.6%	GPT-6 Astra: 9.5%
Vending-Bench 2	$13,718	$16,484	GPT-6 Astra: $15,515

The losses are in the rows many engineering teams care about most. On FrontierSWE 2, Carbon also trails Opus 5.5’s 62.3%. On Artificial Analysis’s Coding Agent Index, Claude Code with Opus 5.5 leads at 62 to Carbon’s 59. On WebDev Arena, Carbon is fourth at 1768 Elo, behind Opus 5.5, GPT-6.1 Astra and Sonnet 5.5.

Some leads are too thin to mean anything. Google’s table has Carbon at 70.4% on the Vals Index against Opus 5.5 at 70.2%, while Vals’ own page lists 70.45% with a margin of plus or minus 0.97, ranked second of 34. Harness choice moves results as much as that gap does. Terminal-Bench 4.0 comes out at 65.2% in Google’s table via Artificial Analysis and at 63.13% on Vals, where it ranks fourth of 61.

The cost story comes from GDPx

The most useful new number arrived on November 2. SemiAnalysis released GDPx v1.0, 206 tasks across finance, materials science, cybersecurity, software, law and biology, and Carbon came second to Opus 5.5.

Model	GDPx overall	Avg cost per task
Claude Opus 5.5	12.0%	$51.41
Gemini 4 Carbon	10.9%	$16.29
GPT-6 Astra	9.6%	$53.99

These figures average execution cost across evaluation runs, failures included, and the overall score is not a task success rate. Even so, roughly a third of the cost for about nine-tenths of the score is a strong commercial argument.

Two cautions apply before anyone rebuilds a budget on that ratio. First, Carbon’s token price is introductory: $2 input and $10 output per million, rising to $4 and $20 on an unannounced date. The higher rate matches Opus 5.5’s pricing. If the GDPx cost reflects the introductory rate, and it scales with token price, the gap narrows to roughly a third cheaper rather than two-thirds. SemiAnalysis’s coverage we read does not state which rate applied, so treat that as our arithmetic. Second, the ranking depends on what you count. Remove cybersecurity, biology and alignment and the order changes: GPT-6 Astra moves to first, with GPT-6.1 Sol and Sonnet 5.5 above Carbon. Reporting attributes the shift to safety guardrails holding back Claude and GPT in those domains, and SemiAnalysis says the reordered list better matches its subjective impressions.

That second point should temper the launch narrative. Part of Carbon’s overall strength appears to come from domains where rivals’ guardrails hold them back.

Accuracy went up, and so did hallucination

Google says unsupported claims roughly halved against Argon in internal testing. The independent picture is less tidy. On AA-Omniscience, Carbon has the top listed score at 54 and answers 58% of questions correctly, up from Argon’s 48%, but its hallucination rate rose from 4% to 11%.

Those two results can both be true, because they measure different things. A model that attempts more of what it does not know can gain correct answers and invented ones together. For retrieval-free question answering, the trade matters. If your application relied on Argon declining rather than guessing, Carbon needs re-evaluating and should not be treated as a drop-in upgrade.

Other alignment numbers are better. Google reports successful cheating on Terminal-Bench 4.0 at 0.0% with attempts at 0.4%, high-severity misalignment alerts in internal live coding down almost 80%, and 0.0% attack success on Gray Swan’s indirect prompt injection test, against 0.7% for Argon. Gray Swan’s is the only third-party evaluation among these, and Argon’s score there left little room to improve.

The cyber gate, and the missing paperwork

Access runs through cybersecurity. Google is starting with a subset of cyber defenders, then widening to trusted partners through its Fairwind Program, and promises broad availability “as soon as possible.” It says it is participating in US government voluntary pre-release testing. Varun Mohan, who heads Antigravity, told VentureBeat that Fairwind access will expand before wider Antigravity availability, without giving a date.

The cyber scores explain the caution. Google reports 88.2% on SEC-bench Pro, 54.8% on ExploitGym (up from Argon’s 42.0%) and 71.0% on CWE-bench, all run without guardrails, so customers will not see those numbers in production. The Chrome team’s claim of more than 10x better patches than “another lab’s large frontier model” comes with no named rival and no published data.

The gap is documentation. One close reading of the launch found no Carbon model card, and found that the linked safety report covers Argon rather than Carbon. The launch also gives no context window, output limit or parameter count. We could not confirm whether Google has determined Carbon reached a critical cyber capability level under its Frontier Safety Framework. That compares poorly with OpenAI, which this week published a Critical rating for GPT-6 Cyber along with a restricted access tier. A model restricted because of its offensive potential should come with a public account of what that potential is.

What to watch

The API listing and the end of introductory pricing. Those two dates decide whether the GDPx cost story survives contact with production. Argon’s own higher rates take effect December 1, which is a reasonable guide to how Google runs these windows.

A Carbon model card. A published card, with a Frontier Safety Framework determination, would answer the biggest open question about this release.

Reruns of GDPx at list price. SemiAnalysis or anyone else repeating the cost comparison at $4 and $20 would give buyers the number they actually need. Until then, teams on Argon can only plan: keep a private eval ready, and include a hallucination check rather than relying on index rank.