- NewNew Published in the last 14 days
Google’s Gemini 4 Carbon leads 15 of 17 benchmarks on its own table, but most organisations cannot buy it yet. The metrics, the differences from GPT-6 Astra, Claude and Argon, the caveats, and what to decide now.
- 59
- Artificial Analysis Intelligence Index v5.0, ranked #1
- 15
- of 17 benchmarks led in Google’s comparison table
- 96.2%
- long-context score (GraphWalks); next non-Gemini model 71.8%
- $4 / $20
- per million input and output tokens after the introductory period
Every new frontier model arrives with a table in which it wins. The useful question for a leadership team is narrower: where is the lead big enough to change a decision, how was it measured, and can you buy the thing yet? This guide answers those questions for Gemini 4 Carbon.
A note on sources. All model facts and scores here come from Google: its launch announcement, its Gemini model page and its evaluation methodology document. Google published these figures and they should be read as vendor claims, including the scores it lists for competitors. Differences between scores and the recommendations are my own reading of those figures.
What Google says Gemini 4 Carbon is #
Google describes Carbon as its strongest model, with state-of-the-art cybersecurity and long-horizon capabilities. It follows Gemini 4 Argon, which is the Gemini model you can use through the API today. Three claims define its positioning:
- Long-running work. Google says its internal teams have seen Carbon run for days on complex engineering tasks, such as root-causing and patching subtle bugs across large codebases.
- Cyber defence. Google reports leading scores for finding, exploiting and patching vulnerabilities, and more than 2x uplift over Argon in generating secure code on internal benchmarks.
- More dependable behaviour. Google reports fewer shortcuts, fewer unsupported claims and stronger resistance toprompt injection than Argon.
Gemini 4 Carbon vs GPT-6 Astra, Claude and Argon: the full table #
These are the scores on Google’s model page, as of October 2026. A dash means Google lists no score. The highest score in each row is marked in bold on Google’s page; here, the last column names the leader.
| Benchmark | Carbon | Argon | GPT-6 Astra | Claude Opus 5.5 | Claude Fable 5.1 | Leader |
|---|---|---|---|---|---|---|
| AA Index v5.0 | 59 | 50 | 57 | 58 | 49 | Carbon |
| Vals Index v2.2 | 70.4% | 69.2% | 65.0% | 70.2% | — | Carbon |
| SWE-sweep | 9.6% | 4.3% | 9.5% | 3.6% | 2.9% | Carbon |
| FrontierSWE 2 | 61.2% | 55.0% | 65.5% | 62.3% | 56.3% | GPT-6 Astra |
| Terminal-Bench 4.0 (AA) | 65.2% | 57.1% | 59.1% | 59.6% | 52.0% | Carbon |
| Agents’ Last Exam v1.1 | 77.4% | 69.0% | 73.1% | 71.6% | — | Carbon |
| OSWorld-2.1 (offline subset) | 81.5% | 74.7% | 80.9% | 87.2% | 76.8% | Opus 5.5 |
| AutomationBench (Zapier) | 53.9% | 51.3% | 41.4% | 42.5% | 31.4% | Carbon |
| Vending-Bench 2 (bank balance) | $16,484 | $13,718 | $15,515 | $9,235 | $5,422 | Carbon |
| Chartography | 72.5% | 71.6% | 71.0% | 66.3% | 46.2% | Carbon |
| Blueprint-Bench 2 | 58.6% | 54.4% | 49.7% | 51.2% | 41.9% | Carbon |
| GraphWalks (256k to 1M, BFS F1) | 96.2% | 84.2% | 71.8% | 66.8% | 65.0% | Carbon |
| RiemannBench | 80.0% | 76.0% | 72.0% | 69.6% | 65.6% | Carbon |
| Terminal-Bench-Science 0.1 (AA) | 64.8% | 52.9% | 63.3% | 59.0% | 43.3% | Carbon |
| CWE-bench | 71.0% | 68.0% | 68.0% | 67.0% | 58.0% | Carbon |
| ExploitGym | 54.8% | 42.0% | 42.4% | 34.5% | — | Carbon |
| SEC-bench Pro | 88.2% | 83.4% | 85.4% | — | — | Carbon |
Key differences: where the lead is large, narrow or missing #
Counting wins hides the most useful signal. Sorting by the gap to the best non-Gemini model on the same row gives a clearer picture.
| Size of lead | Benchmark | Carbon vs best non-Gemini model | What it means for a buyer |
|---|---|---|---|
| Large | GraphWalks | +24.4 points over GPT-6 Astra | The clearest gap. Relevant if you work across very large contexts. |
| Large | ExploitGym | +12.4 points over GPT-6 Astra | Strong offensive capability, which explains the restricted rollout. |
| Large | AutomationBench | +11.4 points over Opus 5.5 | Relevant to multi-step business automation. |
| RiemannBench | +8.0 points over GPT-6 Astra | Relevant to mathematical research work. | |
| Blueprint-Bench 2 | +7.4 points over Opus 5.5 | Relevant to reading plans and diagrams. | |
| Terminal-Bench 4.0 | +5.6 points over Opus 5.5 | Relevant to agentic work in the terminal. | |
| Narrow | AA Index, Vals Index, SWE-sweep | +1 point, +0.2 and +0.1 | Treat as level. Do not switch vendors on these alone. |
| Behind | FrontierSWE 2 | −4.3 points to GPT-6 Astra; also behind Opus 5.5 | Carbon places third of the five listed models. |
| Behind | OSWorld-2.1 | −5.7 points to Opus 5.5 | Computer use is not a clean win: Carbon leads Agents’ Last Exam but trails here. |
The pattern matters more than any single row. Carbon’s strongest case is long context, cybersecurity and sustained automation. For general capability, the top three are close enough that your own workload should decide. For agentic coding, the results are genuinely split across three benchmarks.
Carbon vs Argon: is the upgrade worth waiting for? #
For most teams this is the real comparison, since Argon is the model available now and Google says the two launch at the same introductory price.
| Measure | Gemini 4 Argon | Gemini 4 Carbon |
|---|---|---|
| Availability | Stable in the Gemini API | Subset of cyber defenders; trusted partners next |
| Introductory price per 1M tokens | $2 input, $10 output | $2 input, $10 output |
| Price after the introductory period | See Google’s pricing page | $4 input, $20 output |
| AA Index v5.0 | 50 | 59 |
| SWE-sweep | 4.3% | 9.6% |
| Terminal-Bench 4.0 | 57.1% | 65.2% |
| GraphWalks, 256k to 1M | 84.2% | 96.2% |
| ExploitGym | 42.0% | 54.8% |
| Gray Swan indirect prompt injection, attack success | 0.7% | 0.0% |
| Generating secure code (internal) | Baseline | More than 2x uplift in success rate |
| Unsupported claims (internal) | Baseline | Cut by about 2x |
| Context window | 1,048,576 tokens | Not stated in the announcement |
Carbon scores higher than Argon on all 17 benchmarks, though some gaps are small: 0.9 points on Chartography and 1.2 on the Vals Index. If Argon already clears your quality bar, there is little reason to hold a project for Carbon. Argon’s published specifications give you a firm base to design against today.
Safety and reliability claims #
Google says Carbon builds on Argon’s safety stack and reports these results:
- Cheating: successful cheating on Terminal-Bench 4.0 fell to 0.0%, and attempts fell to 0.4%.
- Misalignment: high-severity alerts dropped by almost 80% in live coding environments with internal teams.
- Staying in scope: overreach fell by 26% and under-service by 29%.
- Prompt injection: Carbon resisted every evaluated attempt on the Gray Swan indirect prompt injection benchmark, with a 0.0% attack success rate.
A 0.0% score means no evaluated attack succeeded on that benchmark. It does not mean an agent is safe to run unsupervised on your systems, so keep your own permissions, sandboxing and approval steps. Our AI governance framework covers the controls to keep in place whichever model you choose.
What the cost looks like #
Google states the introductory price as $2 per million input tokens and $10 per million output tokens, with cached input at 95% off the input price. After the introductory period, $4 and $20 apply. Google has not said when that period ends for Carbon. Here is an illustration for a workload of 100 million input tokens and 20 million output tokens a month:
| Scenario | Introductory | List price |
|---|---|---|
| No caching | $400 a month | $800 a month |
| 60% of input tokens cached | $286 a month | $572 a month* |
*Assumes the 95% cache discount continues at list price. Google states the discount alongside the introductory price and does not confirm it beyond that. Token prices are also only part of the cost: a model that runs for days can consume far more tokens than one that answers a single prompt, so measure cost per completed task. Our guide to evaluating AI ROI shows how to build that business case.
How to read Google’s numbers #
Google’s methodology document is worth reading before anyone quotes this table in a board paper. Five details affect how far you can lean on it:
- Sources differ by row. Some scores come from third parties such as Artificial Analysis, Vals AI, Zapier, Andon Labs and Surge. Others are self-computed by Google, including SWE-sweep, Agents’ Last Exam, OSWorld, ExploitGym and SEC-bench Pro for Carbon.
- Competitor scores are often self-reported. Unless noted, non-Gemini results come from the providers’ own numbers, using maximum reasoning settings or the best available results.
- Setups are not always identical. Carbon’s SWE-sweep score uses a mini-swe-agent harness, while rivals’ scores come from the public leaderboard. That makes a 0.1-point lead especially weak evidence.
- Some scores are best-case. Carbon and Argon ran on the highest thinking settings unless noted, and OSWorld scores are the maximum over three runs. Expect higher cost and latency from those settings.
- Cyber scores were run without guardrails. Google says the external cyber benchmarks were evaluated without guardrails, and ExploitGym with mitigations off. Since Google is still iterating on guardrails, the released product may not behave identically.
Two headline claims also cannot be checked from what is published. The Chrome team’s “more than 10X better” patch result is against “another lab’s large frontier model”, which Google does not name, and the 2x secure code uplift is from internal benchmarks.
What Google has not published yet #
The announcement and evaluation document leave open several items that a procurement or architecture review would normally require:
- A general availability date. Google says only “as soon as possible”.
- Carbon’s context window and output limits.
- When the introductory price ends.
- Latency, throughput and rate limits.
- The final guardrails, including what security work they may block for ordinary customers.
Should you adopt Gemini 4 Carbon? A decision guide #
| Your situation | Recommended decision now |
|---|---|
| You run security or vulnerability work | Ask your Google contact about the Fairwind Program. This is the group Google is prioritising. |
| You have a live project on Argon | Stay on Argon. Add an evaluation set so you can test Carbon on the day it opens. |
| You are starting a new AI project | Build on a model available today, and keep the model behind a thin interface so you can swap it. |
| You work across very large documents or codebases | Put Carbon first on your test list. GraphWalks is its largest published lead. |
| You rely on agentic coding or computer use | Run a head-to-head trial. Google’s own table shows other models leading FrontierSWE 2 and OSWorld-2.1. |
| You are in a regulated sector | Wait for final guardrails and contractual terms before committing. |
Whichever row fits, moving from a promising demo to dependable operation is usually harder than picking the model. Our guide on moving AI pilots into production covers that step.
Sources #
FAQ
Frequently asked questions #
5 questions readers ask about this topic
01What is Gemini 4 Carbon? #
It is the frontier model Google announced on 28 October 2026. Google describes it as its strongest model, with state-of-the-art cybersecurity and long-horizon capabilities.
02Can businesses use Gemini 4 Carbon today? #
Not generally. Google is starting with a subset of cyber defenders and will phase in more trusted partners through its Fairwind Program. It says wider availability will follow as soon as possible, without giving a date.
03How much does Gemini 4 Carbon cost? #
Google says it will launch at $2 per million input tokens and $10 per million output tokens, with cached input at 95% off the input price. After the introductory period, $4 and $20 apply.
04Is Gemini 4 Carbon better than GPT-6 Astra and Claude Opus 5.5? #
On Google’s published table it leads 15 of 17 benchmarks, but GPT-6 Astra leads FrontierSWE 2 and Claude Opus 5.5 leads OSWorld-2.1. Several leads are very small, and many scores are self-computed or self-reported.
05Should we wait for Carbon instead of using Argon? #
Usually not. If Argon meets your quality bar, build on it now, keep the model swappable, and test Carbon on your own tasks when it becomes available.
Still have a question?Ask me directly