Google announced
Gemini 4 Carbon on October 28, 2026, fifteen days after Argon became broadly available. Carbon improves coding, long-context reasoning and cybersecurity results. It is initially being offered to selected cyber defenders; a public API release date has not been announced.
Google announcement,
Argon rolloutThe gains are substantial in some demanding tasks. Carbon's GraphWalks score rises from Argon's 84.2% to 96.2%, and its ExploitGym result rises from 42.0% to 54.8%. Independent evaluations also show a stronger model, although Opus 5.5 still leads AA's coding-agent index and Carbon does not lead every category of office work.
Gemini 4 Carbon vs Argon: the main changes
Google published results for both models across software engineering, computer use and long-context tasks. The changes below are measured in percentage points.
| Task and benchmark | Argon | Carbon | Change |
|---|
| Software engineering: FrontierSWE 2 | 55.0% | 61.2% | +6.2 points |
| Computer use: Agents' Last Exam v1.1 | 69.0% | 77.4% | +8.4 points |
| Desktop tasks: OSWorld-2.1 offline subset | 74.7% | 81.5% | +6.8 points |
| Long-context reasoning: GraphWalks | 84.2% | 96.2% | +12.0 points |
Source:
Google's model evaluation. These benchmarks use different tasks and scoring systems; their percentages are not interchangeable.
More demanding software engineering
FrontierSWE 2 measures software-engineering performance. Carbon improves on Argon by 6.2 points, but GPT-6 Astra scores 65.5% and Opus 5.5 scores 62.3% in the same published comparison, both ahead of Carbon's 61.2%.
Google describes internal use on bugs that require tracing a root cause through large codebases and low-level systems, with some assignments continuing for days. It also reports more than twice Argon's success rate on an internal secure-code generation test. That improvement concerns the security of generated code, rather than a doubling of every coding score.
Google's engineering resultsBetter results across long inputs and desktop tasks
GraphWalks tests graph traversal in inputs between 256,000 and one million tokens. Carbon's 96.2% result measures breadth-first-search F1 on that benchmark. It is evidence of stronger reasoning across long inputs, rather than a guarantee of equally accurate answers to any million-token document.
Computer-use results improve too. Carbon reaches 77.4% on Agents' Last Exam and 81.5% on OSWorld's offline subset. Opus 5.5 remains ahead on the latter at 87.2%. OSWorld here covers 82 tasks without internet access, so this result does not describe unrestricted web browsing.
Evaluation settings and resultsCybersecurity: finding, validating and fixing vulnerabilities
Cybersecurity is central to Carbon's launch. The three highlighted tests cover different parts of the work: discovering a flaw, showing that it can be exploited, and producing a repair.
| Benchmark | What it tests | Argon | Carbon |
|---|
| SEC-bench Pro | Vulnerability discovery and proof-of-concept generation | 83.4% | 88.2% |
| ExploitGym | Exploitation of real software vulnerabilities | 42.0% | 54.8% |
| CWE-bench | Auditing and patching vulnerable code | 68.0% | 71.0% |
These are the results published in
Google's launch announcement, which says the cyber evaluations were run without model guardrails.
SEC-bench Pro requires more than a plausible description of a bug: agents work on vulnerability-discovery and proof-of-concept tasks in executable target environments.
Benchmark description
ExploitGym includes vulnerabilities in ordinary userspace programs, Google's V8 JavaScript engine and the Linux kernel. Developing a working exploit tests a different capability from merely identifying suspicious code. Carbon's improvement over Argon is 12.8 percentage points in Google's reported run.
ExploitGym project
CWE-bench checks whether the repair blocks the vulnerability while preserving existing functionality. Its independent leaderboard also places Carbon first at 71% programmatic pass@1, compared with Argon's 68%. Each model gets four attempts per task; pass@1 and pass@4 are reported separately.
CWE-bench leaderboard and methodology
A public CWE-bench example involves a Rust image-placeholder decoder that can panic on malformed strings. A successful repair must return a recoverable error for invalid input while leaving valid images unchanged. Stopping the crash by breaking the decoder would fail the task.
Independent evaluations: where Carbon leads and where it does not
Artificial Analysis: stronger overall reasoning, second in coding agents
In AA's October 28 evaluation, Carbon scores 59 on Intelligence Index v5, up from Argon's 50. On AA-Briefcase, which evaluates agentic knowledge work, Carbon rises from 1483 to 1609 Elo. It still trails models including Kimi K3.1 Pro at 1678 in that evaluation.
AA's coding-agent index gives Carbon high with Antigravity CLI 59, versus Opus 5.5 max with Claude Code 62. Tooling and reasoning settings are part of those results: they describe the tested agent systems.
On AA-Omniscience v1.1, Carbon's accuracy improves to 58%, ten points above Argon, while the reported hallucination rate rises from
4% to 11%.
AA launch evaluationAA calculates that hallucination rate among responses that were not fully correct, including wrong answers, partial answers and abstentions. The higher rate means a larger share of those responses were wrong answers, rather than an admission of uncertainty or a partial answer. It is not the percentage of all questions answered incorrectly.
AA-Omniscience scoringVals AI: strong finance and research results, a lower coding rank
Vals tests professional tasks as well as coding. Its Carbon results include:
| Evaluation | Carbon result | Published rank |
|---|
| Vals Index v2.2 | 70.36% | 2 of 30 |
| Vals Finance Index | 66.35% | 1 of 34 |
| Legal Research Bench | 55.77% | 1 of 88 |
| Vibe Code Bench v1.1 | 89.06% | 8 of 122 |
Rankings checked on October 30, 2026; the number of models may change.
Sonnet 5.5 leads the overall Vals Index at 70.85%. The 0.49-point gap is small relative to Carbon's reported ±0.97-point uncertainty. Carbon ranks higher on finance and legal research than on the listed coding benchmark.
Vals resultsArena: better session outcomes than Argon, with longer tasks
The October 28 Agent Arena overall table puts Carbon second and Argon seventh. Users explicitly confirmed task completion in 18.05% of Carbon sessions and 12.84% of Argon sessions. Carbon is marked Preliminary.
Carbon's median session time is also longer:
11.4 minutes versus 7.6 minutes for Argon. These sessions contain different tasks, so the figures describe platform usage rather than a paired speed test.
Agent ArenaFrontierCS 2: long-running coding tasks reveal failure modes
In an independent
FrontierCS 2 evaluation published October 29, Carbon passed 25 of 48 tasks, compared with Argon's 26. Each model ran once per task in Antigravity, with up to five hours to improve code against published implementations. Among the 45 tasks where both submitted work, Carbon's mean score was 57.6 versus Argon's 52.9; that average and the pass count measure different things.
The team reviewed all 25 passing Carbon runs and identified three that used loopholes; one pass depended on the loophole. In the first evaluation, three Carbon runs ended without a submission because the CLI stopped waiting for a background command after 30 minutes; the team reran those tasks with a six-hour limit, and one of them then passed. These results describe the tested model and tool setup, not an inherent failure rate for every deployment. They show why inspecting completed work and execution traces matters alongside the headline score.
Context window, inputs and output
Artificial Analysis's model page lists a
one-million-token context window, text, image, audio and video inputs, and text output. Vals also lists a one-million-token context window.
Multimodal input means the model can work with material in those formats. It does not mean Carbon produces native images or videos. A 3D scene built from generated JavaScript, for example, is code rendered by a browser.
Google's public API catalog still lacks a Carbon entry with request parameters, quotas and supported controls. Argon's documented features therefore cannot yet be assumed to be Carbon's public API specification.
Google model catalogCan you use Gemini 4 Carbon today?
As of
October 30, 2026, Google’s public model catalog does not list Carbon, and EvoLink has not added access. Google’s initial rollout targets selected cyber defenders, with a gradual expansion to trusted partners through Fairwind; it has not announced a general release date.
Google announcement ·
Google model catalogFor current access, model ID, EvoLink pricing and availability alerts, see the
EvoLink Carbon model page.
Gemini 4 Argon remains an available alternative.
Frequently asked questions
Does Carbon replace Argon?
Google describes Carbon as building on Argon, but has not announced an Argon retirement date. Argon remains in the public API catalog. The two names should not be treated as interchangeable request IDs.
EvoLink TeamProduct Team
