Gemini 4 Carbon vs Claude Opus 5.5: Benchmarks, Cost, and Access

Google announced Gemini 4 Carbon on October 28, 15 days after Gemini 4 Argon went public, and Artificial Analysis had it at #1 on its Intelligence Index the same day, one point ahead of Claude Opus 5.5. That makes Gemini 4 Carbon vs Claude Opus 5.5 the new Gemini vs Claude question, with one twist: only one of these models will take your prompt today.

I read Google's launch post, Artificial Analysis's full results and model page, Anthropic's Opus 5.5 pricing docs and launch page, Ars Technica's report, and the two-week-old Argon vs Opus 5.5 comparisons on October 28, the day Carbon was announced.

Here's which model to pick for coding, security work, research, long documents, and cost-sensitive API jobs, what each one costs per task at today's prices, and exactly who can get Gemini 4 Carbon right now.

Gemini 4 Carbon vs Claude Opus 5.5: the short verdict

Carbon wins overall, by a point. Artificial Analysis scores Gemini 4 Carbon (high) at 59 on its Intelligence Index v5, with Claude Opus 5.5 (max) at 58 and GPT-6 Astra at 57. Nine evaluations go into that number, and Carbon takes the top score on four of them. The two ran at different reasoning settings, so read the gap as close to a tie.

Opus 5.5 still wins coding agents. On Artificial Analysis's Coding Agent Index, Claude Code running Opus 5.5 scores 62 to Carbon's 59 in Google's Antigravity CLI, and Opus 5.5 ranks above Carbon on Terminal-Bench 4.0 and FrontierCode in that coding-agent run. In the plain-API Terminal-Bench run that feeds the Intelligence Index, Carbon is ahead, 65.2% to 59.6%.

Carbon is cheaper per job. At its standard $4 and $20 per million tokens, the same list price as Opus 5.5, Carbon costs $7.76 per Artificial Analysis task against $10.04 for Opus 5.5, mostly because it writes 29% fewer output tokens to get there.

Opus 5.5 is the one you can use. It's been on the Claude API, in Claude Code, and in the paid Claude plans since September 22. Carbon is going to a small group of cyber defenders first, and Google hasn't given a date for everyone else.

Choose Gemini 4 Carbon if you:

• Do security or vulnerability work: it posts 88.2% on SEC-bench Pro, 54.8% on ExploitGym, and 71.0% on CWE-bench in Google's table, all three the top scores Google published.
• Ask it things you'll quote: on Artificial Analysis's AA-Omniscience test it scores 54 to Opus 5.5's 48, with an 11% hallucination rate against 50% for Opus 5.5 at max effort.
• Feed it very long documents: it scores 96.2% on the 256K-to-1M-token GraphWalks test in Google's table, where Opus 5.5 scores 66.8%.
• Run business automations: it tops Zapier's AutomationBench at 53.9% to 42.5% in Google's table, and leads Artificial Analysis's version too.
• Care about the API bill: cheaper per task at list price, and half price for as long as Google's launch discount lasts.

Choose Claude Opus 5.5 if you:

• Need a flagship today: it's live on every major cloud and in Claude Code, and Carbon isn't live anywhere you can sign up.
• Run long coding sessions in a terminal: it leads Artificial Analysis's Coding Agent Index by three points and ranks above Carbon on FrontierCode in the same run.
• Need computer use: it scores 87.2% on OSWorld-2.1 to Carbon's 81.5% in Google's own table, one of the two rows in that table that Opus 5.5 wins.
• Want a predictable output limit: Opus 5.5 returns up to 128K tokens per request, per Anthropic's model docs, while Google hasn't published Carbon's output limit yet.
• Already pay for Claude: Pro, Max, Team, and Enterprise plans include it, and in Claude Code the opus alias now points to it.

My take: wait for Carbon if your work is research, security, or long documents, and use Opus 5.5 if you're building this week, since it's the only one of the two you can call.

What is Gemini 4 Carbon?

Gemini 4 Carbon is Google's newest frontier model, announced on October 28, 2026, 15 days after Gemini 4 Argon went public. Google's launch post positions it as the company's strongest model to date, built for long-running engineering work and cybersecurity: finding vulnerabilities, testing exploit paths, and patching the root cause. Artificial Analysis lists it as a reasoning model with a 1M-token context window and text, image, video, and speech input.

It's the second Gemini 4 release within a month, and Google says it "illustrates the pace at which we're pushing the frontier forward." Google hasn't published a model ID, output limit, or knowledge cutoff yet, so some of the details below come from Artificial Analysis and not from Google.

Can you use Gemini 4 Carbon yet?

No, you can't use Gemini 4 Carbon yet unless you're one of the cyber defenders Google has picked (status as of October 28, 2026). Google's launch post files Carbon under "Coming soon to developers, businesses, and consumers," and says it will reach Google's products "as soon as possible," with no date attached.

Here's the ladder Google described, in order:

1. U.S. government pre-release testing. Google says it's "actively engaged" in the government's voluntary process for testing frontier models before release.
2. A subset of cyber defenders, starting this week. This is the first outside group to get Carbon, which fits where its biggest gains are: finding, exploiting, and patching vulnerabilities.
3. Trusted partners through the Fairwind Program. Google DeepMind's Fairwind Program gives vetted defenders early model access, prioritizing governments, national cyber authorities, critical infrastructure operators, and core technology platforms. It currently works with more than 650 partners.
4. Developers, businesses, and consumers. Google's wording is "as soon as possible," after it has continued to "gather feedback and iterate on guardrails."

Ars Technica expects Carbon to probably arrive first for AI Ultra subscribers and API developers, possibly within a couple of weeks if the rollout matches Argon's, but Google hasn't confirmed any of that.

How to get Claude Opus 5.5 today

Opus 5.5 has been available since September 22, and Anthropic splits access by surface:

1. The Claude API. Call claude-opus-5-5 directly. In Claude Code, the opus alias now resolves to it on the Anthropic API, Claude Platform on AWS, Amazon Bedrock, and Google Cloud's Agent Platform. On Microsoft Foundry the alias still points to Opus 4.6, so select claude-opus-5-5 by name.
2. The Claude apps. Opus models are included in the Pro, Max, Team, and Enterprise plans, according to Claude's pricing page, and not in the Free plan.
3. Claude Code. It works from version 2.1.280 or later, with fast mode available in Claude Code, on the Claude API, and on Anthropic's managed endpoints on Google Cloud and Microsoft Foundry.
4. Cloud platforms. Amazon Bedrock, Claude Platform on AWS, Google Cloud, and Microsoft Foundry all list it.

Our Claude Opus 5.5 guide covers the effort levels and the things that break when you upgrade from Opus 5, like thinking you can no longer switch off.

Specs, pricing, and access side by side

The two models match on context window and, once Google's discount ends, on list price. They differ on who can use them, what they read, and how many tokens they burn per task.

How they compare at a glance:

Spec	Gemini 4 Carbon	Claude Opus 5.5
Made by	Google DeepMind	Anthropic
Announced	Oct 28, 2026	Sept 22, 2026
Who can use it today	Selected cyber defenders only	Anyone on a paid Claude plan or the API
API model ID	Not published yet	claude-opus-5-5
API price per 1M (in / out)	$2 / $10 at launch, $4 / $20 standard	$4 / $20
Cached input per 1M	$0.10 at launch, $0.20 standard (95% off)	$0.20 read, $5 write (5-min)
AA cost per task (list price)	$7.76	$10.04
Output tokens per AA task	123k	173k
Speed (AA)	92.4 tokens/s, 100.31 s to first token (high)	98.2 tokens/s, 663.66 s to first token (max)
Context window	1M tokens (per Artificial Analysis)	1M tokens
Max output	Not published	128K tokens
Inputs	Text, images, video, speech	Text, images
Runs on	Google surfaces, when it ships	Claude API, Claude Code, Bedrock, Google Cloud, Microsoft Foundry

Carbon's figures come from Google's launch post and Artificial Analysis's Carbon model page.

Opus 5.5's come from Anthropic's pricing docs and its Opus 5.5 launch page, plus Artificial Analysis's Opus 5.5 page for the speed row.

The discount has no end date, so budget at the list price. Google's post calls $2 and $10 the "introductory price" and footnotes that $4 and $20 "will apply" after it expires. Artificial Analysis describes it as a 50% launch discount "with no confirmed end date," and Google declined to give Ars Technica one.

Every Artificial Analysis cost figure in this article uses the $4 and $20 list price, which is also what Argon charges from December 1. If you're comparing plans, our Gemini pricing breakdown has the subscription side.

Two spec details to check before you build:

1. Google hasn't published Carbon's output limit or model ID. Argon returns 65,536 tokens per request by default and needs long decode to go further, and Carbon may inherit that, but nothing in the launch post says so. Opus 5.5's 128K limit is documented.
2. Opus 5.5 benchmarks at max effort, but defaults to medium. Anthropic's scores (and Artificial Analysis's "Opus 5.5 (max)" runs) use the top effort settings, max or xhigh. A request that doesn't set effort runs at medium, and thinking can't be switched off, so pin the setting you tested.

What a real bill looks like

Say your app sends 1,000 requests, each with 5,000 input tokens and 2,000 output tokens.

The bill at each price:

At list price the two cost exactly the same per token, so the per-task gap is all about how much each model writes. Artificial Analysis measured about 123,000 output tokens per index task for Carbon and 173,000 for Opus 5.5, which is why Carbon lands at $7.76 per task and Opus 5.5 at $10.04.

Anthropic has two levers Google hasn't announced for Carbon yet (Argon already gets Batch at half price on the Gemini API): Batch API requests at 50% off for work that can wait, and fast mode at $8 and $40 for up to 2.5x faster output, both per Anthropic's pricing docs. Google's 95% cache discount is the bigger saver if your prompts repeat a lot of context.

Gemini 4 Carbon benchmarks vs Claude Opus 5.5

In Google's published table, Gemini 4 Carbon beats Claude Opus 5.5 on 14 of the 16 rows where both models have a score. Artificial Analysis, which runs every model through the same harness itself, calls it a one-point lead overall and a three-point loss on coding agents. Both are right.

Google's own numbers

These come from Google's Carbon evaluation document. Google runs the Gemini models at their highest thinking setting and takes the Opus 5.5 scores from Anthropic's reports, public leaderboards, and Artificial Analysis, at max effort where available, so treat the table as the vendor's best case.

Google's numbers, Carbon vs Opus 5.5:

Benchmark	Gemini 4 Carbon	Claude Opus 5.5	Leader
AA Intelligence Index v5.0	59	58	Carbon
Vals Index v2.2 (knowledge work)	70.4%	70.2%	Carbon
SWE-sweep (coding)	9.6%	3.6%	Carbon
FrontierSWE 2 (project-scale coding)	61.2%	62.3%	Opus 5.5
Terminal-Bench 4.0, AA run	65.2%	59.6%	Carbon
Agents' Last Exam v1.1	77.4%	71.6%	Carbon
OSWorld-2.1 (computer use)	81.5%	87.2%	Opus 5.5
AutomationBench (Zapier)	53.9%	42.5%	Carbon
Vending-Bench 2 (bank balance)	$16,484	$9,235	Carbon
Chartography (charts)	72.5%	66.3%	Carbon
Blueprint-Bench 2 (floor plans)	58.6%	51.2%	Carbon
GraphWalks 256K to 1M (long context)	96.2%	66.8%	Carbon
RiemannBench (math)	80.0%	69.6%	Carbon
Terminal-Bench-Science 0.1, AA run	64.8%	59.0%	Carbon
CWE-bench (patching)	71.0%	67.0%	Carbon
ExploitGym (exploitation)	54.8%	34.5%	Carbon
SEC-bench Pro (finding vulnerabilities)	88.2%	not published	n/a

Carbon's biggest leads are on long documents (29.4 points on GraphWalks), exploitation (20.3 points on ExploitGym), and business automation (11.4 points on AutomationBench). Opus 5.5's two wins are computer use and project-scale coding, by 5.7 and 1.1 points.

The cybersecurity rows are the ones Google built the launch around. In its post, Google says Carbon more than doubled Argon's success rate at generating secure code on internal tests, and that the Chrome team saw it produce high-quality patches "more than 10X better than another lab's large frontier model."

Those two claims are internal and can't be checked from outside, so weigh them accordingly.

On prompt injection, Carbon posts a 0.0% attack success rate on Gray Swan's Indirect Prompt Injection benchmark, down from Argon's 0.7%. Google's own chart puts Opus 5.5 at 1% and GPT-6 Astra at 8.5%. Carbon also climbed 89 Elo to 1768 and fourth place on WebDev Arena.

What Artificial Analysis's independent index says

Artificial Analysis puts Gemini 4 Carbon at 59 on its Intelligence Index v5, ahead of Claude Opus 5.5 at 58 and GPT-6 Astra at 57. The lab reported the results on October 28, running Carbon at its high setting and Opus 5.5 at max. Those settings differ and I found no published margin of error for the index, so a one-point lead is better read as near parity until someone reruns both at matching settings.

Inside that index, the split looks like this:

• Carbon takes the top score on four of the nine tests: HLE-Diamond, Terminal-Bench-Science, AA-Omniscience, and AutomationBench-AA.
• Opus 5.5 keeps coding: in Artificial Analysis's separate Coding Agent Index, Carbon is second to Opus 5.5 on Terminal-Bench 4.0 at 64% and on FrontierCode v1.1 at 51%. Opus 5.5 scores higher on both (Anthropic reports 54.4% at max on FrontierCode).
• Coding agents go to Claude Code: on the Coding Agent Index, Opus 5.5 (max) in Claude Code scores 62, and Carbon (high) in Google's Antigravity CLI scores 59. Artificial Analysis notes Google "has historically underperformed in this area," so a three-point loss is a big step up for Gemini.
• Carbon makes fewer things up: on AA-Omniscience v1.1, it scores 54 to Opus 5.5's 48, answers 58% of questions correctly, and when it doesn't know, it guesses wrong 11% of the time instead of declining. Opus 5.5 (max) and Astra both guess wrong 50% of the time.

Vals AI's independent index agrees that Carbon edges Opus 5.5. On the Vals Index, Carbon ranks second of 29 models at 70.36%, just ahead of Opus 5.5 at 70.18% (a gap within its ±0.97 error bar) and behind Claude Sonnet 5.5 at 70.85%.

Where Carbon still trails

Opus 5.5 beats Carbon clearly at office-style agent work. On AA-Briefcase, Artificial Analysis's test of document, spreadsheet, and presentation tasks, Carbon scores 1609 Elo, a 126-point jump over Argon. Opus 5.5 (max) scores 1800, and Carbon also sits behind Kimi K3.1 Pro (1678), Muse Wonder (1659), and Grok 4.7 (1638).

Neither model is quick to start. Artificial Analysis's Carbon model page lists 100.31 seconds to first token at the high setting, and its Opus 5.5 page lists 663.66 seconds at max, against a comparable-model median of 3.72 seconds. Once they start, Opus 5.5 is a little faster, at 98.2 tokens per second to Carbon's 92.4. Opus 5.5's wait is about 6.6 times longer, though the reasoning settings differ, so for anything interactive, time the effort level you'll actually run.

It hallucinates more than Argon did. Carbon's 11% hallucination rate on AA-Omniscience is still the second lowest of any model Artificial Analysis tracks, but it's up from Argon's 4%. The trade is a higher share of correct answers, and whether that trade is worth it depends on whether a confident wrong answer or a blank costs you more.

Two benchmark numbers that need an asterisk.

Opus 5.5's Terminal-Bench 4.0 score depends on who ran it and how. Google's table takes its 59.6% from Artificial Analysis's plain API run, where Carbon scores 65.2%. Anthropic's launch page reports 66.4% at xhigh effort.

Artificial Analysis's coding-agent harness flips the order again: Carbon in Antigravity CLI scores 64% and finishes second to Opus 5.5 running in Claude Code. The fair reading is that the two are close and the order depends on the setup.

FrontierCode moves the same way. Artificial Analysis ranks Opus 5.5 above Carbon's 51% without a figure I could confirm, and Anthropic reports 54.4% at max effort on the benchmark's main set. Compare each source's numbers only against its own, and run your own tasks before you trust any single row.

‍

{{templates}}

‍

What changed from Gemini 4 Argon

Carbon arrived 15 days after Argon went public, so the practical question for anyone who just started paying for Argon is whether to wait. Our Gemini 4 Argon explainer covers Argon's access and pricing in detail. Here's what Carbon adds, in numbers:

• Nine more points on the Intelligence Index: 59 against 50 in Google's table, at the same $4 and $20 standard price. Artificial Analysis says cost per task is level too, $7.76 for Carbon against $7.68 for Argon.
• A jump from 57.1% to 65.2% on Terminal-Bench 4.0 in Google's table, and up 6 points on Artificial Analysis's Coding Agent Index, from 53 to 59.
• A 126-point gain on AA-Briefcase, from 1483 to 1609 Elo, with the biggest improvement in presentation tasks (up 293 Elo).
• Better accuracy, worse hallucination rate: correct answers on AA-Omniscience rose from 48% to 58%, and the hallucination rate rose from 4% to 11%.
• Stronger alignment claims from Google: unsupported claims cut by about half, overreach down 26%, under-service down 29%, successful cheating on Terminal-Bench 4.0 at 0.0%, and high-severity misalignment alerts in Google's own coding environments down by almost 80%. All of these are internal measurements.

Ars Technica's verdict is that Carbon is "essentially a slightly better version of Argon," and that for all anyone outside Google knows, it could be a post-training upgrade of the same model. Google calls it a new frontier model. The scores say the gains are real and spread across security, long context, and science, with smaller steps elsewhere.

Argon's advantage over Carbon is availability. Argon is on the paid Gemini API and in the Gemini app today, at $2 and $10 through November 30.

Which is better for what

1. Agentic coding in a terminal

Opus 5.5 wins, for now. It leads Artificial Analysis's Coding Agent Index at 62 to Carbon's 59, and it ranks above Carbon's 51% on FrontierCode v1.1 in the same harness. In Google's own table, Opus 5.5 also edges Carbon on FrontierSWE 2, the test of long, project-scale engineering tasks, at 62.3% to 61.2%.

Carbon's coding scores are a big step up for Gemini, and in the plain-API Terminal-Bench 4.0 run Google's table cites, it leads 65.2% to 59.6%. But the tool many people use for this job is Claude Code, which only runs Claude models, and Carbon isn't available in anything yet.

Pick: Opus 5.5 in Claude Code today. Re-test Carbon in Antigravity once it ships.

2. Long-running engineering and security work

Carbon wins. Google built it for "sustained agentic workflows" and says its teams have watched Carbon run for days on multi-step engineering tasks like root-causing bugs across large codebases. On the security benchmarks Google published, it leads Opus 5.5 by 20 points on ExploitGym and 4 points on CWE-bench, and Anthropic didn't publish an Opus 5.5 score for SEC-bench Pro at all.

There's a policy angle too. Per Anthropic's Opus 5.5 launch post, most cybersecurity tasks on Opus 5.5 are re-routed to Opus 4.8 under its safeguards, so Opus 5.5 is often not the model you'd get for this work anyway. Google, for its part, is sending Carbon to cyber defenders before anyone else.

Pick: Carbon, once you can get it. Waiting is clearly worth it for this job.

3. Knowledge and research questions

Carbon wins on the independent numbers. It takes the top score on HLE-Diamond and AA-Omniscience in Artificial Analysis's index, and its 11% hallucination rate on AA-Omniscience is a fraction of Opus 5.5's 50%. In Google's table it also leads RiemannBench (80.0% vs 69.6%) and Terminal-Bench-Science (64.8% vs 59.0%).

Opus 5.5 has one counterpoint from Anthropic's own testing: asked to write reports from a copy of the web with a grader that fails any invented figure, Opus 5.5 cleared the bar on 16 of 18 reports across effort settings, where Fable 5.1 and Opus 5 cleared it on none. That's a vendor test, but it's a hard one.

Pick: Carbon for questions you'll quote, because a model that answers "I don't know" is cheaper than one that guesses.

4. Long documents and spreadsheets

Split decision, by input length. For very long inputs, Carbon is far ahead: 96.2% on GraphWalks at 256K to 1M tokens, against 66.8% for Opus 5.5, with the same 1M-token window, per Artificial Analysis. It also takes video and speech input, which Opus 5.5 doesn't.

For the office-work side of documents, Opus 5.5 leads. On AA-Briefcase, Artificial Analysis's spreadsheet-and-presentation test, Opus 5.5 (max) scores 1800 Elo to Carbon's 1609, and Kimi K3.1 Pro, Muse Wonder, and Grok 4.7 also beat Carbon.

Charts are a coin flip: Opus 5.5 leads Chartography in Anthropic's own reporting (89.0% with tools) but trails on the no-tools Surge leaderboard figure Google cites (66.3% vs 72.5%).

Pick: Carbon for contract libraries, transcripts, and anything over 256K tokens. Opus 5.5 for building decks and spreadsheets.

5. Computer use

Opus 5.5 wins, in Google's own table. It scores 87.2% on the OSWorld-2.1 offline subset to Carbon's 81.5%, and Anthropic says Opus 5.5 leads its own benchmarks on computer use. Carbon does lead Agents' Last Exam (77.4% vs 71.6%), a computer-use agent test, so the two trade wins.

Pick: Opus 5.5 for agents that click through apps with no API. Carbon for the kind of computer-use agent task Agents' Last Exam measures, once it's out.

6. Cost-sensitive API workloads

Carbon wins on paper, Opus 5.5 wins on the invoice you can get this month. At list price, Carbon costs $7.76 per Artificial Analysis task to Opus 5.5's $10.04, a 23% saving, because it uses 29% fewer output tokens. Google's launch discount halves Carbon's token prices on top of that, for as long as it lasts.

Opus 5.5 answers with Batch API requests at 50% off and cache reads at 5% of the input price, so a well-cached, batched Opus workload can land below Carbon's list-price figure.

Pick: Carbon when it ships, if your prompts don't cache well. Opus 5.5 with batch and caching if you need to run the workload this month.

Where GPT-6 Astra and Gemini 4 Argon fit

Google's launch chart compares five models, and two of the others deserve a line before you choose.

GPT-6 Astra scores 57 on Artificial Analysis's index, two points behind Carbon, and at $6.25 per task it's the cheapest of the three flagships to run. It also leads FrontierSWE 2 in Google's table at 65.5%. Our Opus 5.5 vs Astra breakdown covers that matchup, and the Argon vs Astra comparison covers how Google's previous flagship stacks up against OpenAI's.

Gemini 4 Argon scores 50 in Google's table and is the Gemini model you can use today, at $2 and $10 per million tokens through November 30. If you're weighing the subscription side, our Claude pricing guide covers every Claude plan and API rate.

Pick by job, then recheck when Carbon's access widens

Gemini 4 Carbon vs Claude Opus 5.5 comes down to whether you can wait: Carbon for security work, research, and long documents once you can get it, Opus 5.5 for anything you're shipping this week.

Here's a quick way to decide on your own work:

1. Pick 10 real tasks. Use work you do every week: a bug fix, a vulnerability review, a research question, a long contract.
2. Run Opus 5.5 at medium and max effort now. Note the cost, time, and quality for each, so you have a baseline before Carbon opens up.
3. Rerun the same tasks on Carbon the day you get access. Price it at $4 and $20, not the launch discount, so a price change doesn't surprise you.
4. Route by result. Send each kind of task to the model that won it, and keep Opus 5.5 for terminal coding and computer use until Carbon proves otherwise on your own tasks.

The same routing works in any tool that lets you pick the model for each task. If your team uses Lindy, an AI teammate that works in Slack, check which models are on its list and send each job to the one that won your tests.

‍

{{cta}}

‍

Frequently asked questions about Gemini 4 Carbon vs Claude Opus 5.5

Is Gemini 4 Carbon better than Claude Opus 5.5?

The main difference between Gemini 4 Carbon and Claude Opus 5.5 is that Carbon wins on security, research, and long documents and costs less per task, while Opus 5.5 wins on coding agents and computer use and is the only one of the two you can use today. Artificial Analysis rates Carbon 59 and Opus 5.5 58 on its Intelligence Index, while Opus 5.5 leads its Coding Agent Index 62 to 59 and leads computer use in Google's own table.

Can I use Gemini 4 Carbon yet?

No, Gemini 4 Carbon isn't available to the public as of October 28, 2026. Google is giving it to a subset of cyber defenders first, then to trusted partners through its Fairwind Program, and says it will reach developers, businesses, and consumers "as soon as possible," with no date given.

When is the Gemini 4 Carbon release date?

Google hasn't announced a public release date for Gemini 4 Carbon. It announced the model on October 28, 2026, and says developers, businesses, and consumers will get access "as soon as possible." Argon opened to the public 13 days after its announcement, but Google hasn't said Carbon will follow that schedule.

How much does Gemini 4 Carbon cost?

Gemini 4 Carbon costs $2 per million input tokens and $10 per million output tokens at its launch price, and $4 and $20 at its standard price. Google describes the lower rate as introductory with no end date announced, and Artificial Analysis measures Carbon at $7.76 per index task at the standard price.

Is Gemini 4 Carbon better than Gemini 4 Argon?

Yes, Gemini 4 Carbon beats Gemini 4 Argon on every headline benchmark Google published, including a nine-point gain on the Artificial Analysis Intelligence Index (59 vs 50) at the same standard price. Its hallucination rate on AA-Omniscience is higher, though, at 11% against Argon's 4%.

What is the best Gemini model right now?

Gemini 4 Carbon is Google's best model on the benchmarks, and Gemini 4 Argon is the best one you can use. Carbon scores 59 to Argon's 50 on Artificial Analysis's index, but only a subset of cyber defenders have it. Argon is on the paid Gemini API and in the Gemini app today, at $2 and $10 through November 30.

What is the #1 AI model right now?

Gemini 4 Carbon is the top model on Artificial Analysis's Intelligence Index at 59, as of October 28, 2026, one point ahead of Claude Opus 5.5 at 58 and two ahead of GPT-6 Astra at 57. On Vals AI's index, Claude Sonnet 5.5 is first at 70.85%, ahead of Carbon (second of 29 models at 70.36%) and Opus 5.5 (70.18%).

Is Gemini 4 Carbon better than GPT-6 Astra?

Gemini 4 Carbon scores two points higher than GPT-6 Astra on Artificial Analysis's Intelligence Index, 59 to 57, and leads it on most rows in Google's table. Astra is cheaper per task at $6.25 to Carbon's $7.76 and leads project-scale coding on FrontierSWE 2, and unlike Carbon, you can use it today.

Is Gemini 4 Carbon better than ChatGPT?

On the independent index, yes, by two points. Carbon scores 59 against 57 for GPT-6 Astra, OpenAI's flagship, per Artificial Analysis. Astra is cheaper per task at $6.25, and Carbon isn't available yet, so for most ChatGPT users the practical comparison today is Gemini 4 Argon vs GPT-6 Astra.