Gemini 4 Carbon sells itself on security and stamina, while a rival proves narrow beats broad

What Carbon claims

Google has released Gemini 4 Carbon, its strongest model for frontier performance in cybersecurity and long-horizon agentic workflows, and built it for sustained agentic workflows. Developers can reach it through the paid tier of the API in Google AI Studio, and Google AI Ultra subscribers can use it in Google Antigravity. The price is $2 per million input tokens and $10 per million output tokens. m:…1648 m:…1647

Google DeepMind says Carbon ranks first on CursorBench 4.0, Perplexity's WANDR, and Zapier's AutomationBench, and second on the Multimodal Knowledge Agent benchmark. The launch pitch is aimed squarely at developers building agents that run for a long time and touch real systems, rather than at chat. Carbon was released to Fairwind Program participants two weeks ago, which means some teams have already been testing it in production-shaped settings. m:…0798

The security pitch, in Google's own numbers

The security claims are the sharpest part of the launch, and they are all Google's own measurements. In Project Fortify, Carbon achieved 100% recall on scanning and 95% precision on triage. In simulated cyberattack testing environments it blocked 99% of attack trajectories, and it achieved a 0% prompt injection attack success rate on the Gray Swan IPI benchmark. Its per-attempt breach rate is 0.8%, compared with 3.1% for Argon, and it passed 92% of tasks on the agentic honesty benchmark. The company says Wiz used Carbon to uncover 0-day vulnerabilities in under two hours. m:…1648 m:…0798

The science claims point the same way. Carbon increased the number of homologs retrieved by 40-fold on average for protein structure prediction, and researchers used it to generate millions of lines of code in Lean to prove security properties of the Signal protocol. These are capability demonstrations, not independent evaluations, and they should be read that way. A lab choosing its own benchmark and reporting its own score is normal for a launch, and it is also the reason the scores deserve a second test before they shape a buying decision. m:…1648

The counterpoint: a specialist outscores the frontier

The most useful contrast this week did not come from Google. Baseten and Genspark post-trained an open-weight Qwen3.5-397B model into a specialized Deep Research model, splitting the work so the research phase was trained with reinforcement learning and the writing and synthesis phase was trained primarily with supervised fine-tuning. The post-trained model scored 0.759 on Genspark's quality grader, and the companies say it outperforms Opus 5.5, Sonnet 5.5, GPT-5.6 Sol, Kimi K3, and DeepSeek V4 Pro on that grader. It also crawls 5 times more sources per user query than the base model. m:…0579

The production detail matters as much as the score. The model runs on 2x B200 GPUs per replica using NVFP4 weights, and the companies say the quantized model matched full-precision performance within eval noise. Genspark retains ownership of the model weights, evaluations, and training code. The implication for a builder is plain: for a narrow, well-defined job, post-training a large open-weight model can beat renting a frontier model, on a grader the team itself controls. That grader is the weak point too, since the comparison is judged by the same team that built the model. m:…0579

Same week, another surface

Google's model family also showed up in a different kind of product this week. Recraft Studio integrated Gemini Omni Flash 1.2, a video generation model Google released on November 5, 2026. Recraft says the model reduces drift and morphing compared with version 1.1, improves camera move adherence to prompts, and provides clearer lip-syncing for dialogue, with native audio and generations limited to 10 seconds in Recraft Studio. Google is previewing a swarm feature in Google Antigravity that enables coordinated multi-agent workflows, which fits the same pattern as Carbon: the pitch is not one clever answer, but sustained work across many steps. m:…6747 m:…1648

Google is also unusually open about where the video model struggles, identifying fast sports action, rapid script changes, and complex physics as problems. That candour is worth noting because it is absent from the Carbon launch, where every reported number is a win. Developers deciding between Carbon and a specialist post-trained model should treat the Carbon numbers as a starting point for their own evaluation, not as a verdict, and should ask what their own agents actually do all day before paying for frontier generality. m:…6747