Google introduced Gemini 4 Carbon on October 28, 2026, two weeks after Gemini 4 Argon, as its strongest model for long-horizon engineering and cybersecurity defense. Independent testing puts Carbon at No. 1 on the Artificial Analysis Intelligence Index v5.0 with a score of 59 and No. 2 on the Vals Index at 70.36%. It matches Argon on list price while posting a nine-point index gain, which is why this launch matters for anyone buying frontier capability by the task rather than by the token.
What Just Happened
Carbon is Google DeepMind’s next frontier model after Argon, focused on sustained agentic workflows, coding, and the full cyber defense lifecycle from finding flaws to shipping fixes. It takes the No. 1 spot from Claude Opus 5.5, which led the newly updated Index v5.0 at launch with 58 before Carbon’s 59 displaced it a day later. Google says internal teams have run Carbon for days on complex multi-step engineering work, including root-cause subtle bugs across large codebases and low-level systems.
Rollout is deliberately narrow. Starting this week, Carbon goes to a subset of cyber defenders, then to additional trusted partners through the Fairwind Program. Google says it is also taking part in the U.S. government’s voluntary pre-release model testing process. There is no date yet for developers, businesses, or consumers, and Google describes general availability as coming as soon as possible after feedback and guardrail work.
Key facts at launch:
- Announced October 28, 2026, by Google DeepMind
- 1M token context window, with text, image, video, and speech input and text output
- Private weights, reasoning model with a high-effort setting used for benchmark runs
- Introductory API price of $2 per million input tokens and $10 per million output tokens, with cached input at 95% off, moving to $4 and $20 after the introductory period
How Much Better Than Argon
The short version is broader gains rather than one breakthrough, with the largest jumps in terminal science work, long context retrieval, and exploit validation.
| Measure | Gemini 4 Carbon | Gemini 4 Argon |
|---|---|---|
| Artificial Analysis Intelligence Index v5.0 | 59 | 50 |
| Vals Index v2.2 | 70.36% (#2 of 29) | 69.24% |
| AA output speed | 99.8 tokens per second | 82.0 tokens per second |
| Cost per Intelligence Index task, standard pricing | $7.76 | $7.68 |
| Output tokens per task | 123,000 | Not disclosed in AA Carbon writeup, described as in line on cost |
| Terminal-Bench 4.0, AA harness in Google table | 65.2% | 57.1% |
| Terminal-Bench Science 0.1, AA harness in Google table | 64.8% | 52.9% |
| SEC-bench Pro, vulnerability discovery | 88.2% | 83.4% |
| ExploitGym, exploit validation | 54.8% | 42.0% |
| CWE-bench, root cause patching | 71.0% | 68.0% |
| GraphWalks 256K to 1M, long context | 96.2% | 84.2% |
Output speed is close at about 100 tokens per second for Carbon against 82 for Argon on Artificial Analysis provider runs. That similarity has prompted speculation that Carbon could share lineage with Argon through further tuning or post training. Google has not confirmed that, so treat it as unconfirmed context rather than fact, and expect early speed figures to shift once access widens.
For how Argon itself compared to OpenAI’s flagship on price and task cost, see our Gemini 4 Argon vs GPT-6 Astra comparison.
Price and What Tasks Really Cost
Per token pricing tells only half the story because Carbon is a verbose reasoning model. Artificial Analysis calculates cost per completed Index task from input, cached, reasoning, and answer tokens, which is the more useful budget number.
| Model | Input per 1M | Output per 1M | Cached input per 1M |
|---|---|---|---|
| Gemini 4 Carbon, list, 50% intro is $2 and $10 | $4.00 | $20.00 | $0.20 |
| Gemini 4 Argon, list, $2 and $10 through Nov. 30 | $4.00 | $20.00 | $0.20 |
| Claude Opus 5.5 | $4.00 | $20.00 | $0.20 |
| Muse Wonder | $4.00 | $20.00 | $0.20 |
| GPT-6 Astra | $10.00 | $50.00 | $1.00 |
| Claude Fable 5.1 | $10.00 | $50.00 | $0.25 |
During the introductory period, Carbon halves to $2 and $10, which Artificial Analysis prices at $3.88 per Index task.
Carbon costs about 23% less per task than Opus 5.5 at max effort while using about 29% fewer output tokens per task, 123,000 against 173,000. It costs slightly more per task than Astra despite much lower token prices because it generates roughly twice the tokens Astra does per task. In other words, Google’s low token rate absorbs Carbon’s verbosity but does not erase it.
Two latency details matter for planning. Artificial Analysis records about 200 million output tokens across its Carbon Index run, about three times the 64 million median, and a median of 97.5 seconds to first answer token against a 3.6 second median. Streaming is fast once it starts, near 100 tokens per second, but reasoning-heavy tasks wait longer before the first token.
What Independent Tests Actually Show
Artificial Analysis updated to Index v5.0 just before Carbon launched, so scores cannot be compared to older v4 releases. The new suite has nine evaluations covering briefcase knowledge work, automation, terminal coding, frontier coding, omniscience and hallucination, document reasoning, hard reasoning and knowledge, critical point physics, and terminal science. Carbon takes first place in four of them: HLE-Diamond, Terminal-Bench Science, AA-Omniscience, and AutomationBench-AA.
Carbon’s Index v5.0 component scores, via Artificial Analysis as republished by Opper AI’s model page which aggregates AA data:
| Index v5.0 component (Carbon High) | Score |
|---|---|
| AutomationBench-AA | 77.8% |
| HLE-Diamond | 69.1% |
| Terminal-Bench Science 0.1 | 64.8% |
| CritPt v2 hard subset | 63.0% |
| GDP.pdf all-pass | 28.4% |
| AA-Briefcase Elo | 1609 |
| AA-Omniscience Index | 54 (58% accuracy, 11% hallucination rate) |
GDPval-AA v2.1 was removed from the Index in the v5 methodology update, so no current Carbon figure applies to it. The AutomationBench-AA figure of 77.8% is the AA Index component; Google’s separate Zapier-variant table shows 53.9% on a different harness, and the two should not be mixed.
| Benchmark, Google table | Carbon | Opus 5.5 | Astra | Argon |
|---|---|---|---|---|
| AA Index v5.0 | 59 | 58 | 57 | 50 |
| Vals Index v2.2 | 70.4% | 70.2% | 65.0% | 69.2% |
| SWE-sweep | 9.6% | 3.6% | 9.5% | 4.3% |
| FrontierSWE 2 | 61.2% | 62.3% | 65.5% | 55.0% |
| Terminal-Bench 4.0 | 65.2% | 59.6% | 59.1% | 57.1% |
| Agents Last Exam v1.1 | 77.4% | 71.6% | 73.1% | 69.0% |
| OSWorld 2.1 offline | 81.5% | 87.2% | 80.9% | 74.7% |
| AutomationBench | 53.9% | 42.5% | 41.4% | 51.3% |
| Chartography | 72.5% | 66.3% | 71.0% | 71.6% |
| GraphWalks 256K to 1M | 96.2% | 66.8% | 71.8% | 84.2% |
| RiemannBench | 80.0% | 69.6% | 72.0% | 76.0% |
| Terminal-Bench Science 0.1 | 64.8% | 59.0% | 63.3% | 52.9% |
Source: Google’s Gemini 4 Carbon launch evaluation table. Vendor tables run in each lab’s native harness, so read the pattern rather than the decimals. Independent Artificial Analysis and Vals.ai runs confirm the headline placements above.
Vals.ai: strong on research, finance, and code
Vals.ai ranks Carbon No. 2 of 29 models on the Vals Index at 70.36%, behind Claude Sonnet 5.5 at 70.85% and ahead of Claude Opus 5.5 at 70.18%. Rankings here will move as more runs complete, so treat the gap as narrow rather than settled.
| Vals.ai benchmark | Carbon score | Rank |
|---|---|---|
| Vals Index overall | 70.36% | #2 of 29 |
| Vals Finance Index | 66.35% | #1 of 33 |
| Vals Legal Index | 75.83% | #3 of 30 |
| Legal Research Bench | 55.77% | #1 of 87 |
| EMB v1.1 | 76.47% | #2 of 35 |
| Finance Agent v2.1 | 67.16% | #5 of 33 |
| Tax Agent Bench | 48.86% | #3 of 78 |
| Terminal-Bench 4.0 | 63.13% | #4 of 57 |
| Vibe Code Bench v1.1 | 89.06% | #8 of 121 |
| HLAB-Vals-1.1 | 17.50% | Tied #6 of 30 |
The pattern is consistent across both suites, with Carbon strongest in research, finance, legal retrieval, science, and terminal-centered coding, and less dominant in pure office computer use and FrontierSWE-style engineering.
What Carbon Is Actually For
Benchmark tables are useful for placement, but capability is the better buying guide. Three themes run through Google, Artificial Analysis, and Vals.ai results.
Long horizon engineering without giving up early
Web development is one visible proxy, with Carbon gaining 89 Elo points to reach 1768 and No. 4 on WebDev Arena.
Cybersecurity across find, prove, and fix
Cyber is Carbon’s clearest flagship. Google reports state-of-the-art results across the defense lifecycle when evaluated without guardrails, with independent interest focused on whether defense teams can turn those scores into safer shipped code.
Google adds internal results that outside labs cannot yet verify. These include more than double the success rate for generating secure code versus Argon, more than 10 times better high-quality patches than another large frontier model in Chrome team testing, and a lead on web penetration testing benchmarks used in Wiz’s Scan for Good initiative. Treat vendor internal claims as directional until independent replications arrive.
More reliable behavior under pressure
Safety work focuses on fewer shortcuts and more honest collaboration. Google reports zero successful cheating on Terminal-Bench 4.0, with attempts down to 0.4%; almost 80% fewer high-severity misalignment alerts in live coding; about half the unsupported claims of Argon; 26% less overreach and 29% less underservice; and 0.0% attack success on Gray Swan indirect prompt injection testing against 0.7% for Argon. For background on how Google prices efficient reasoning models, see our Gemini 3.8 Flash benchmarks and pricing breakdown.
How to Get It
Most readers cannot use Carbon yet. There is no public API or consumer rollout date. Access starts with cyber defenders this week and expands through Fairwind to additional trusted partners as Google gathers feedback on guardrails.
What to watch is not only capability but also delivery history. Google announced Argon on September 30 and opened broad access on October 13, a staged gap of about two weeks. Carbon follows the same playbook so far, but staged does not mean fast, and Google’s own language emphasizes deliberate deployment plus voluntary pre-release testing with the U.S. government.
For the OpenAI side of the flagship race, see our GPT-6 Astra benchmarks against Claude and our broader Gemini vs ChatGPT assistant comparison.
What Could Take the Crown Back
A one-point lead on the Intelligence Index is thin. Artificial Analysis scores Carbon at 59 against 58 for Opus 5.5 and 57 for Astra in Google’s table, and Vals.ai has Carbon 0.49 points behind Sonnet 5.5. Either gap can flip with more runs or with the next releases from OpenAI and Anthropic.
Expectations center on GPT-6.1 Astra and Fable 5.5 updates, and GPT-6.1 Terra already appeared on Vals.ai on October 26. Those challengers cost more per token at current rate cards, with Astra and Fable 5.1 at $10 input and $50 output against Carbon’s $4 and $20 list. The economic question is whether extra reasoning tokens erase Carbon’s token price edge on specific workloads, as already happens against Astra on cost per task.
The broader business context is tighter than a simple profitable versus unprofitable split. Alphabet remains profitable overall, reporting $26.17 billion in Q3 net income, but it missed expectations, posted negative $9.07 billion free cash flow, and guided 2026 capital spending to $195 billion to $205 billion on data centers and AI chips. That makes efficient models that do more work per dollar strategically important, even for the best funded labs.
Frequently Asked Questions
When was Gemini 4 Carbon released?
Google DeepMind announced Carbon on October 28, 2026. Broad developer and consumer access has no confirmed date. Cyber defenders get first access through the Fairwind Program.
How much does Gemini 4 Carbon cost?
Introductory pricing is $2 per million input tokens and $10 per million output tokens, with cached input at 95% off. Standard pricing after the introductory period is $4 input and $20 output, the same rate card as Argon, Opus 5.5, and Muse Wonder, and below Astra and Fable 5.1 at $10 and $50.
How fast is Gemini 4 Carbon?
Artificial Analysis measures about 99.8 tokens per second output on Google’s API against 82.0 for Argon, which is faster than the peer median. Time to first answer is much slower at a median of 97.5 seconds because of long reasoning traces, so streaming feels fast only after a long initial wait.
Is Gemini 4 Carbon better than Claude Opus 5.5?
On overall intelligence it currently leads by one point on Index v5.0, 59 to 58 in Google’s table, and costs less per task at list prices. Opus 5.5 still leads coding agents overall and computer use tasks such as OSWorld, while Vals.ai ranks Sonnet 5.5 slightly ahead of Carbon overall. The better model depends on whether the work is reasoning, coding agents, or desktop computer use.
Can I use Gemini 4 Carbon today?
Not unless you are in the initial Cyber Defender cohort or a Fairwind trusted partner. Developers, businesses, and consumers are listed as coming soon with no date.
Is Carbon just a tuned Argon?
Google has not said that. Shared pricing, similar speed, and a two-week gap make further tuning plausible, but that remains speculation. Treat Carbon as a separate frontier release until Google describes the training relationship.
The Bottom Line
Carbon earns its headline as the new No. 1 on Intelligence Index v5.0, with genuine separation in science, long context, terminal coding, and cybersecurity defense. It does so at Argon pricing rather than Astra pricing, which resets what frontier capability should cost per task. The lead is narrow, access is narrow, and verbosity plus slow first answers are real tradeoffs. For teams choosing today, the practical move is to test Carbon where work is long, technical, and verifiable, while keeping Opus for computer use and Astra where terminal speed or FrontierSWE-style engineering matters most.
