Gemini 4 Carbon

Gemini 4 Carbon is a proprietary frontier large language model from Google DeepMind, announced on October 28, 2026 as the second model released under the Gemini 4 name in the Gemini family.^[1] Google calls it "our strongest model with state-of-the-art cybersecurity and long horizon capabilities" and presents it as the successor to Gemini 4 Argon, which the announcement says "was launched two weeks ago."^[1] Google said Carbon ranked first on version 5.0 of the Artificial Analysis Intelligence Index with a score of 59, a placing Artificial Analysis confirmed on its own site.^[1][8][9] As with Argon, the model was not generally available at launch. Google began giving access to "a subset of cyber defenders" in the week of the announcement, said it would phase in more trusted partners through its Fairwind Program, and said it was taking part in the U.S. government's voluntary process for pre-release model testing.^[1] It announced an introductory API price of $2 per million input tokens and $10 per million output tokens, the same as Argon's, rising to $4 and $20 after an introductory period whose end date it did not give.^[1][22]

The launch post, written by Koray Kavukcuoglu, Google DeepMind SVP and Google's chief AI architect, puts the most weight on three areas: long-running agentic software engineering, cybersecurity defense, and alignment. Google reports state-of-the-art results on three external security benchmarks: 88.2% on SEC-bench Pro for finding vulnerabilities, 54.8% on ExploitGym for exploitation and 71.0% on Collinear's CWE-bench for patching. It says these were "evaluated without guardrails."^[1] Independent results published on launch day were broadly positive but not uniform. Artificial Analysis placed Carbon first on its Intelligence Index and second on its Coding Agent Index. Vals AI ranked it second on its Vals Index, behind Claude Sonnet 5.5, and Collinear's CWE-bench leaderboard ranked it first.^[8][12][13]

Overview

Background and development

Google announced Gemini 4 Argon on September 30, 2026, first giving access to cyber defenders in the Fairwind Program. It made Argon generally available on October 13, 2026 through the Gemini API, Google AI Studio, Google Antigravity and the Gemini app.^[25][26] Carbon was announced 15 days after that general release, a gap VentureBeat drew attention to when it separated Argon's announcement from its public launch.^[22]

The name surfaced before the launch. On October 9, 2026, Business Insider reported that Google employees had been trying "a new version of Gemini 4 internally named Carbon," which had been made available inside Google's internal coding platform, Jetski, "in recent days." The report described it as "possibly a new update to Argon, also known as a 'checkpoint'."^[20] According to Business Insider, an internal document showed that Google had tested a series of Gemini 4 models internally named Argon, Barium and Carbon. One document said the internal "Barium-B" model had been chosen to become the public Argon, and the outlet could not determine whether Carbon would ship under the Argon name or as a different Gemini 4 model.^[20] One employee told the publication that Carbon "feels like Opus 5.5" for coding, with the caveat that they would need to test it more. Google declined to comment.^[20] On October 23, TestingCatalog reported posts on X claiming that an early access program for Carbon had started, with partners receiving access and the model appearing on LMArena. Testers quoted there described it as better than Argon at 3D work but slower, and still weak at frontend and UI work.^[21]

Announcement

Google published "Gemini 4 Carbon: our next frontier model" on its Keyword blog on October 28, 2026.^[1] Sundar Pichai introduced it on X with the line "I did say we'd be releasing our models periodically...introducing Gemini 4 Carbon, our next frontier model," and wrote that Carbon "shows the pace at which we're pushing the frontier."^[4] Demis Hassabis wrote that Carbon "is SOTA across key benchmarks and combines frontier capabilities with deep safety & security research," and Google DeepMind's account called it "state-of-the-art across cybersecurity and long-horizon capabilities."^[5][6] The launch post attributed the release pace to Google's "researchers and models ... working hand in hand to advance capabilities faster, while strengthening our safeguards in tandem."^[1]

On the same day Google DeepMind's Gemini models page added Carbon to its benchmark table and listed the Carbon post first among its Gemini announcements. Google also published a separate evaluation methodology document for the model.^[2][3]

Availability

Google described a "deliberate approach to rollout." In the post's words: "Starting this week, we will make Carbon available to a subset of cyber defenders and will gradually phase access to additional trusted partners through our Fairwind Program. We're continuing to gather feedback and iterate on guardrails before making Carbon available to developers, businesses, and consumers via our product surfaces as soon as possible."^[1] Google AI's account described the first cohort as "a subset of trusted cyber defenders in our Fairwind program," and Pichai told users to "hang tight" ahead of "wider availability soon."^[4][7] Google gave no date for a public release.^[1]

Google describes the Fairwind Program as giving "high-priority defenders (like governments, healthcare providers, and telecommunications services) early access to advanced models." As of October 29, 2026, its program page still described exclusive access to Gemini 4 Argon and did not mention Carbon.^[27]

When it launched, Carbon had no public API. Artificial Analysis wrote that "as with launch of Gemini 4 Argon, Gemini 4 Carbon is being rolled out to selected users and is not yet publicly available," and its charts mark the model as having "No publicly available API."^[8][9] VentureBeat noted that Google's announcement gave no API specification for Carbon, so it was not clear whether Argon's documented limits and features would carry over. VentureBeat also reported a post by Google Antigravity head Varun Mohan saying that access would expand through Fairwind ahead of wider availability in Antigravity.^[22]

Pricing

Google said Carbon "will launch at the same introductory price as Argon," and a footnote to the post says "After the introductory period expires, the price of $4 per 1M input tokens and $20 per 1M output tokens will apply."^[1] Unlike Argon, whose introductory period Google set to end on November 30, 2026, the Carbon post gives no end date. VentureBeat and Artificial Analysis both pointed this out, and Artificial Analysis described the introductory rate as a "50% launch discount, with no confirmed end date."^[8][22][25] Independent evaluators priced the model at the standard rate. Artificial Analysis said all its charts use standard $4 and $20 pricing, which makes cached input $0.20 per million tokens. Vals AI also lists $4 and $20.^[8][12]

Capabilities claimed by Google

Google says Carbon is "designed to take on much harder, long-horizon engineering problems" and "built for sustained agentic workflows." According to the post, "Internally, our teams have seen Carbon run for days to persist through dead ends to tackle complex, multi-step engineering tasks like root-causing and patching subtle bugs across large codebases and low-level systems."^[1] Google says it worked on coding, agentic and scientific tasks, full-stack web development and multimodal understanding in response to user feedback. It also says that "with improved quality of output and artifacts in web development, Carbon jumps +89 elo points to land at #4 (1768) on WebDev Arena."^[1]

Google's materials do not give a parameter count, architecture details or a training data cutoff for Carbon. VentureBeat wrote that Google had not released a technical report.^[1][22]

Benchmarks

Google's launch comparison

Google's benchmark table compares Carbon with Gemini 4 Argon, OpenAI's GPT-6 Astra and Anthropic's Claude Opus 5.5 and Claude Fable 5.1.^[2][3] Carbon has the highest figure in 15 of the 17 rows. GPT-6 Astra is higher on FrontierSWE 2 and Claude Opus 5.5 on the OSWorld-2.1 offline subset.^[2][22] According to the methodology document, Gemini scores are pass@1 with the highest thinking settings through the Gemini API, and scores for other companies' models are "sourced from providers' self-reported numbers unless otherwise mentioned."^[3]

Source: Google DeepMind's Gemini models page and launch graphic, October 28, 2026. Bold marks the highest figure in each row.^[2][4]

The methodology document says which results Google ran itself and which it took from outside leaderboards:^[3]

• Taken from outside sources: the AA Index and the two Terminal-Bench rows (Artificial Analysis), the Vals Index (Vals AI), FrontierSWE (Proximal's public leaderboard), AutomationBench (Zapier's private set and public leaderboard), Vending-Bench 2 and Blueprint-Bench 2 (Andon Labs), Chartography and RiemannBench (Surge AI), and CWE-bench (Collinear AI).
• Self-computed by Google: SWE-sweep for Carbon (using a mini-swe-agent harness), Agents' Last Exam for the Gemini models (122 tasks, the default ALE-Claw harness, a five-hour window), OSWorld-2.1 for the Gemini models (82-task offline subset, best of three runs), GraphWalks for all models, and ExploitGym and SEC-bench Pro for the Gemini models (Antigravity harness).

Independent evaluations

Artificial Analysis released version 5 of its Intelligence Index on October 27, 2026, the evening before the Carbon launch (its announcement post on X is timestamped early on October 28 UTC). At that point Claude Opus 5.5 (max) led at 58 and Gemini 4 Argon (high) scored 50.^[11][31] Version 5 added FrontierCode 2.0 and Terminal-Bench-Science, replaced Humanity's Last Exam with HLE-Diamond, and, according to the firm, bases 60% of the index on evaluations with private test sets or held-out answers.^[11] Its methodology groups the nine component evaluations into four categories (knowledge work, coding, science and general) that each contribute 25%.^[9][29] Carbon's 59 put it at the top of that leaderboard on October 28. The firm said Carbon had the top score on 4 of the 9 component evaluations (HLE-Diamond, Terminal-Bench-Science, AA-Omniscience and AutomationBench-AA). It added that Carbon "makes substantial gains elsewhere but is not leading in coding and knowledge work."^[8] Artificial Analysis says scores from different versions of the index are not intended for long-run comparison.^[30]

Artificial Analysis reported a cost of $7.76 per Intelligence Index task at standard pricing, which it put at about 23% less than Claude Opus 5.5 (max, $10.04) and about 1.2 times GPT-6 Astra (max, $6.25). It reported an average of 123,000 output tokens per task, 29% fewer than Opus 5.5.^[8] At the same standard pricing, Carbon's cost per task was in line with Argon's $7.68.^[8] The firm's model page lists Carbon's measured output speed as 99.8 tokens per second and calls the model "very verbose," at 200 million output tokens across the index against a median of 64 million.^[10]

On factual reliability, the two sets of results point in different directions. Google says Carbon "cuts unsupported claims by approximately 2x over Argon" in internal testing.^[1] On Artificial Analysis's updated AA-Omniscience v1.1, Carbon's accuracy rose 10 points over Argon to 58%, but its hallucination rate rose from 4% to 11%. Artificial Analysis said this was still the second lowest of any model it tested, behind only Argon, compared with 50% for both Claude Opus 5.5 (max) and GPT-6 Astra (max).^[8] VentureBeat pointed out that the two measures come from different evaluations and do not establish a uniform improvement or regression.^[22]

The Vals result gives context for Google's Vals Index row. Google's table shows Carbon ahead of Claude Opus 5.5 by 70.4% to 70.2%, but the full Vals leaderboard places Claude Sonnet 5.5, which is not in Google's comparison, above both.^[2][12] On Vals AI's other benchmarks, Carbon ranked first of 33 models on the Vals Finance Index and first of 87 on Legal Research Bench.^[12]

Cybersecurity

Google says Carbon "achieves state-of-the-art performance across the cybersecurity defense lifecycle," which it divides into vulnerability detection, validation and remediation.^[1] It ties each of its three headline security benchmarks to one of those stages.^[1]

• SEC-bench Pro extends SEC-bench, a 2025 framework for benchmarking LLM agents on real-world security tasks.^[15][28] Its May 2026 paper describes 344 validated vulnerabilities and reports that, among the models it tested, Codex with GPT-5.5 solved 58%.^[15] Google says it ran Carbon on "the 183-task May 2026 JavaScript engine split," averaging three attempts per task. A task counts as solved if the proof of concept crashes the vulnerable build but not the patched one, or if it crashes both and an LLM judge reviewing the source code confirms the crash comes from the intended vulnerability.^[3]
• ExploitGym was introduced in a May 2026 paper by researchers including Zhun Wang, Nicholas Carlini and Dawn Song. The paper describes 898 instances.^[17] Google ran it with one attempt per task, without internet access and "with mitigations off," measuring success under both the default 6-hour timeout per task and an extended 10-hour timeout; its methodology does not say which setting produced the 54.8% figure. A task counted only if the exploit captured the target flag and an LLM judge verified that the intended vulnerability was used.^[3]
• CWE-bench is a private benchmark from Collinear AI, described in an October 2026 paper.^[14] Its public leaderboard lists Carbon, run in the Antigravity harness, alone in first place at 71% programmatic pass@1 and 81% programmatic pass@4. Carbon has no cost figure there yet.^[13]

Google's other cybersecurity claims are internal or come from partners. Google says Carbon "achieves more than 2x uplift in success rate for generating secure code over Argon" on internal benchmarks. It says the Chrome team reported that Carbon "performed more than 10X better than another lab's large frontier model when generating high quality patches," without naming the other model.^[1] Google also says Wiz, which it describes as "our leading Cloud security platform," observed Carbon leading on web penetration testing benchmarks. Google calls that capability core to Wiz's Scan for Good initiative.^[1] Wiz describes Scan for Good as an effort that pairs AI with human security researchers to find public exposures at public services, critical infrastructure operators and nonprofits. As of a September 30, 2026 update, the initiative was using Gemini 4 Argon alongside Gemini 3.8 Flash Cyber.^[19]

Safety and alignment

Google says Carbon "builds on the same multi-layered safety and security stack of Argon," which combines model training, real-time misuse defenses, prompt injection resilience and chain-of-thought monitoring. It links the Gemini 4 Safety and Security report, in which it reported Argon's propensities to behave in an aligned way.^[1][26] It says that for Carbon it "trained the model to respond to complex or high-pressure tasks in the right way."^[1] All the figures below are Google's own and come from internal testing or Google-run evaluations.

The Terminal-Bench figures relate to reward hacking, in which an agent satisfies a test without doing the intended task. VentureBeat noted that the Gray Swan result applies to the attacks evaluated and does not establish immunity to every malicious instruction, and that the release does not establish how often an outside customer would see the same failures or reproduce the reductions.^[22] For external testing, Google's post says only that the company is "actively engaged in the U.S. government's voluntary process for pre-release model testing."^[1]

Reception

Coverage on launch day focused on the pace of releases and the restricted rollout. VentureBeat's headline said Google was "retaking the world's no. 1 spot on updated, harder benchmarks" while noting that "access is limited." It concluded that Google was "advertising improved benchmark performance at matching token rates," but had "not published enough operating data to establish lower spending on completed engineering or security work."^[22] Constellation Research wrote that "the takeaway is that Google is stepping up its frontier model release pace."^[24] 9to5Google reported the announcement as following "the launch of Argon two weeks ago."^[23] Artificial Analysis called Carbon's coding-agent result notable because "Google has historically underperformed in this area."^[8]

Place in the Gemini 4 line

References

1. ^1 ^2 ^3 ^4 ^5 ^6 ^7 ^8 ^9 ^10 ^11 ^12 ^13 ^14 ^15 ^16 ^17 ^18 ^19 ^20 ^21 ^22 ^23 ^24 ^25 ^26 ^27 ^28 ^29 ^30 ^31 ^32 ^33 ^34 ^35 ^36 ^37 ^38 ^39Koray Kavukcuoglu. "Gemini 4 Carbon: our next frontier model." Google (The Keyword), October 28, 2026. blog.google/...introducing-gemini-4-carbon
2. ^1 ^2 ^3 ^4 ^5Google DeepMind. "Gemini" (models page, Gemini 4 benchmark table). Accessed October 29, 2026. deepmind.google/...gemini
3. ^1 ^2 ^3 ^4 ^5 ^6Google DeepMind. "Gemini 4 Carbon Model evaluation: Approach, methodology & results." October 2026. deepmind.google/...gemini-4-carbon
4. ^1 ^2 ^3Sundar Pichai (@sundarpichai). Post on X, October 28, 2026. x.com/...2115489219804352921
5. ^Demis Hassabis (@demishassabis). Post on X, October 28, 2026. x.com/...2115489550256746502
6. ^Google DeepMind (@GoogleDeepMind). Post on X, October 28, 2026. x.com/...2115489356437651654
7. ^Google AI (@GoogleAI). Post on X, October 28, 2026. x.com/...2115489593818816591
8. ^1 ^2 ^3 ^4 ^5 ^6 ^7 ^8 ^9 ^10 ^11 ^12 ^13 ^14 ^15Artificial Analysis (@ArtificialAnlys). "Google's Gemini 4 Carbon is the new #1 on the Artificial Analysis Intelligence Index v5..." Post on X, October 28, 2026. x.com/...2115494553780367384
9. ^1 ^2 ^3 ^4 ^5Artificial Analysis. "Artificial Analysis Intelligence Index v5.0." Accessed October 29, 2026. artificialanalysis.ai/...alysis-intelligence-index
10. ^1 ^2 ^3 ^4 ^5Artificial Analysis. "Gemini 4 Carbon (High) Intelligence, Performance & Price Analysis." Accessed October 29, 2026. artificialanalysis.ai/...gemini-4-carbon
11. ^1 ^2 ^3Artificial Analysis (@ArtificialAnlys). "Announcing Artificial Analysis Intelligence Index v5..." Post on X, October 28, 2026. x.com/...2115265133467635877
12. ^1 ^2 ^3 ^4 ^5 ^6 ^7 ^8Vals AI. "Gemini 4 Carbon Benchmarks, Cost and Capabilities." Updated October 28, 2026, accessed October 29, 2026. vals.ai/...google_gemini-4-carbon
13. ^1 ^2 ^3 ^4Collinear AI. "CWE-bench" leaderboard (v1). Accessed October 29, 2026. cwe-bench.com
14. ^1 ^2Gonzalo Gonzalez-Pumariega et al. "CWE-bench: Refocusing Agentic Security from Exploitation to Defense." arXiv:2610.17020, October 12, 2026. arxiv.org/...2610.17020
15. ^1 ^2 ^3Hwiwon Lee et al. "SEC-bench Pro: Can Language Models Solve Long-Horizon Software Security Tasks?" arXiv:2605.26548, May 26, 2026. arxiv.org/...2605.26548
16. ^SEC-bench Pro leaderboard. Accessed October 29, 2026. sec-bench.github.io
17. ^1 ^2Zhun Wang et al. "ExploitGym: Can AI Agents Turn Security Vulnerabilities into Real Attacks?" arXiv:2605.11086, May 11, 2026. arxiv.org/...2605.11086
18. ^sunblaze-ucb/exploitgym. GitHub repository. Accessed October 29, 2026. github.com/...exploitgym
19. ^Wiz. "Scan for Good: Finding Critical Exposures with AI." Wiz Blog (updated September 30, 2026). wiz.io/...scan-for-good-critical-ai-exposures
20. ^1 ^2 ^3Hugh Langley. "Google employees praise latest Gemini 4 internal model." Business Insider, October 9, 2026. businessinsider.com/...argon-barium-carbon-2026-10
21. ^"Gemini 4 Carbon early test results are in ahead of public release." TestingCatalog, October 23, 2026. testingcatalog.com/...-4-carbon-early-test-results
22. ^1 ^2 ^3 ^4 ^5 ^6 ^7 ^8 ^9 ^10Carl Franzen. "Google unveils Gemini 4 Carbon, retaking the world's no. 1 spot on updated, harder benchmarks - but access is limited." VentureBeat, October 28, 2026. venturebeat.com/...enchmarks-but-access-is-limited
23. ^Abner Li. "Google announces Gemini 4 Carbon with 'significant gains across coding and agentic tasks'." 9to5Google, October 28, 2026. 9to5google.com/...gemini-4-carbon-announcement
24. ^"Google rolls out Gemini 4 Carbon." Constellation Research, October 28, 2026. constellationr.com/...le-rolls-out-gemini-4-carbon
25. ^1 ^2 ^3Koray Kavukcuoglu. "Gemini 4 Argon: Built to solve your hardest problems." Google (The Keyword), October 13, 2026. blog.google/...introducing-gemini-4-argon
26. ^1 ^2 ^3Koray Kavukcuoglu. "Gemini 4 Argon: our next era of frontier intelligence." Google (The Keyword), September 30, 2026. blog.google/...gemini-4-argon; Google DeepMind. "Gemini 4 Safety and Security report." October 2026. storage.googleapis.com/...gemini_4_safety_report.pdf
27. ^Google DeepMind. "Fairwind Program." Accessed October 29, 2026. deepmind.google/fairwind-program
28. ^Hwiwon Lee et al. "SEC-bench: Automated Benchmarking of LLM Agents on Real-World Software Security Tasks." arXiv:2506.11791, June 2025. arxiv.org/...2506.11791
29. ^Artificial Analysis. "Intelligence Benchmarking Methodology" (Intelligence Index v5.0). Accessed October 29, 2026. artificialanalysis.ai/...intelligence-benchmarking
30. ^Artificial Analysis. "Progress Index Methodology." Accessed October 29, 2026. artificialanalysis.ai/...progress-index
31. ^Artificial Analysis. "Announcing Artificial Analysis Intelligence Index v5." October 27, 2026. artificialanalysis.ai/...sis-intelligence-index-v5