AivexaNewsSearch
AI news for builders and product teamsChecked every hour

Google Deepmind releases Gemini 4 Carbon just two weeks after Argon with gains across the board

Collected Oct 28, 2026

Google DeepMind has released Gemini 4 Carbon, a new frontier model that arrives just two weeks after its predecessor, Gemini 4 Argon. Carbon takes the top spot on the recently updated Artificial Analysis Intelligence Index v5 with 59 points, nine points above Argon's 50. Because Google highlights advanced cybersecurity capabilities that span the full defense lifecycle, from detecting and validating vulnerabilities to patching them, it is restricting early access to selected security experts, starting this week.

Google says the model is aimed at what it describes as much harder, long-horizon engineering problems. Internal teams reportedly ran Carbon for multiple days at a stretch, during which it worked through dead ends autonomously and handled multi-step work such as root-cause analysis and subtle bug fixes in large codebases and low-level software. Beyond agentic tasks and coding, Google cites improvements in scientific reasoning, multimodal understanding and full-stack web development.

On the Artificial Analysis Intelligence Index, Carbon at its highest reasoning tier leads mainly on complex reasoning, knowledge retrieval and scientific tasks, taking first place in four of the index's nine individual evaluations. Google's own benchmark selection places Carbon first in 15 of 16 tests, with the widest margins in long-context and security tasks.

Coding remains a contested area. On the Artificial Analysis Coding Agent Index, Carbon rises from Argon's 53 to 59, but Claude Opus 5.5 still sits ahead at 62. In agentic coding on Antigravity, Carbon trails Claude Code with Opus 5.5 by three points. In agentic knowledge work on AA-Briefcase, Carbon (high) reaches 1,609 Elo, a 126-point gain over Argon (high, 1,483 Elo), helped by presentation quality up 293 Elo to 1,598 and analytical quality up 120 Elo to 1,671. It still trails Kimi K3.1 Pro (1,678 Elo), Muse Wonder at max settings (1,659 Elo) and Grok 4.7 at xhigh (1,638 Elo).

Accuracy and hallucination both moved. Answer accuracy rises from 48 to 58 percent against Argon, while the hallucination rate climbs from four to 11 percent, which Artificial Analysis still ranks second-lowest overall.

Vals AI places Carbon second at 70.36 percent accuracy in its Vals Index v2.2, between Claude Sonnet 5.5 at 70.85 percent and Claude Opus 5.5 at 70.18 percent, with Argon fourth at 69.24 percent. At $10.51 per test and just under 48 minutes of runtime, Carbon runs cheaper and faster than both Anthropic models. It also leads Vals AI's Finance Index and Legal Research benchmark; in legal research the platform recorded roughly 45 percent more cited sources, more frequent tool calls and longer responses than Argon. The trade-off is latency and verbosity: Finance and Tax Agent tasks take more than twice as long as with Argon, legal research averages over 50 minutes per task, and the model generates 2.3 times as many reasoning tokens and 64 percent more output tokens than Argon, though still under half the output tokens of Sonnet 5.5 and Opus 5.5.

On Arena.ai, Carbon (High) takes the Text Arena lead at 1,528 points, edging Argon's 1,520. In the Agent Arena it places second across more than 21,900 real sessions, ahead of GPT-6 Astra and Fable-5.1, with a net improvement of 16.2 percent and the highest steerability score, up 18.6 percent. Its median task cost of $1.80 is 72 percent below the first-place Claude Opus 5.5 (Max). In the Code Arena WebDev category Carbon scores 1,768, 89 points ahead of eleventh-place Argon, for fourth overall and two behind Claude Sonnet 5.5 at xHigh (1,770), putting it on the Pareto curve across all three disciplines.

Security is where the release gets unusual. Several tests ran without safety guardrails. On SEC-bench Pro, which measures autonomous vulnerability discovery in real-world software, Carbon scores 88.2 percent and leads all tested models. On ExploitGym, which tests turning discovered flaws into working exploits, it reaches 54.8 percent at ten hours of runtime. It also holds a first-place lead in writing fixes on CWE-Bench at 71.0 percent, ahead of GPT-6 Astra and Argon. On Gray Swan's indirect prompt injection benchmark, the attack success rate fell from 0.7 percent with Argon to 0.0 percent, blocking every tested attempt according to Google; the test appears saturated rather than the problem solved. Google also cites internal results that cannot be independently verified, including secure code produced more than twice as often as Argon and, in Chrome team tests, more than ten times better at writing security patches than the frontier model of an unnamed competitor. This suggests the staged rollout reflects genuine dual-use risk: a model strong at finding and exploiting vulnerabilities can serve attackers as readily as defenders.

Access follows that logic. Carbon is going through the U.S. government's voluntary pre-testing process, with selected cybersecurity experts first, then expansion to more trusted partners through the Fairwind Program. Google plans to open it to developers, businesses and consumers through its product interfaces after gathering more feedback and adjusting safeguards. Pricing matches Argon: $2 per million input tokens and $10 per million output tokens, with cached input at a 95 percent discount. Artificial Analysis puts the cost at $7.76 per task versus $7.68 for Argon, about 23 percent below Claude Opus 5.5 (max, $10.04) and roughly 1.2 times GPT-6 Astra (max, $6.25), averaging 123,000 output tokens per task. Those calculations use list prices of $4 and $20 per million tokens, ignoring a 50 percent introductory discount running through the end of 2026.

Why it matters: Teams evaluating frontier models now see a leader on general intelligence and security benchmarks that is not yet generally available, with coding and knowledge-work leaderboards still led elsewhere on some measures. Developers should expect higher token consumption and longer runtimes when adopting Carbon for agentic work, which affects cost planning even at unchanged per-token prices. The cautious, security-gated rollout also signals that access to top-tier models may increasingly depend on vetting rather than a public API key.

Read at The Decoder

Based on reporting from the original publisher. Visit the source for full context and later updates.

Publisher excerpt

Just two weeks after Gemini 4 Argon, Google DeepMind is releasing Gemini 4 Carbon, which tops the Artificial Analysis Intelligence Index v5 with 59 points. Google says the model is built for extended agent workflows and the full cyber defense lifecycle. Because of those advanced security capabilities, access is limited to selected security experts at first. The article Google Deepmind releases Gemini 4 Carbon just two weeks after Argon with gains across the board appeared first on The Decoder .