Google DeepMind's Koray Kavukcuoglu titled the Tuesday post "Gemini 4 Carbon is now available" and used it to move the model out of the Fairwind Program's cyber-defender cohort into developer, enterprise, scientist and consumer hands. API access runs through Google AI Studio and Gemini for business under the model ID gemini-4-carbon, with pricing locked at $2 per million input tokens and $10 per million output tokens through November 30, then stepping to $4 and $20 on December 1. Argon, released shortly before Carbon, keeps the same introductory sticker and anchors Google's lineup for day-to-day software engineering and knowledge work. Carbon is tuned to spend more tokens on hard tasks, which Kavukcuoglu framed in the post as more extensive exploration and verification. All Google AI Ultra subscribers get Carbon rolling into Google Antigravity, the Gemini app and AI Mode in Search; it also lands in Android Studio, CodeMender and Gemini Enterprise through the API and model dropdown. Scientists can apply for six months of free Argon and Carbon access through Antigravity's highest usage tier, and cybersecurity defenders can still apply for Carbon for cyber through Fairwind.
Benchmarks and the independent reading
Google's table covers 17 benchmarks. Carbon scores higher than Argon on all 17 and tops GPT-6 Astra, Claude Opus 5.5 and Claude Fable 5.1 on 15 of them. Carbon ranked first on CursorBench 4.0, Perplexity's WANDR and Zapier's AutomationBench, each at more than 2X lower cost per task than the next best external model. On the Artificial Analysis Multimodal Knowledge Agent benchmark, Carbon ranked second at more than 6X lower average cost per task than the top model. Hebbia's end-to-end Financial Agentic evaluation matched the best model tested on deep financial research. On Balyasny Asset Management's General Financial Intelligence benchmark, Carbon topped 23 frontier model configurations. Carbon leads SEC-bench Pro (finding), ExploitGym (exploiting) and CWE-bench v1 (patching) on the Fairwind suite. Artificial Analysis, running its own numbers, placed Carbon first on Intelligence Index v5.0 at 59 points against Opus 5.5 at 58, GPT-6 Astra at 57 and Argon at 50, with AA-Omniscience six points above Opus 5.5 at max reasoning; the hallucination trade-off reads at 58 percent correct against Argon's 48, with the reported hallucination rate rising from 4 to 11.
"Today, we're making Gemini 4 Carbon available to developers, enterprises, scientists, and consumers." — Koray Kavukcuoglu, SVP Google DeepMind, 2026-11-10
Cybersecurity and scientific work
Project Fortify, Google's internal security pipeline, posted 100 percent recall on scanning and 95 percent precision on triage, with over 25 percent relative recall improvement on open-source repository testing against Argon. Wiz Scan for Good found zero-day vulnerabilities in two major SaaS providers in under two hours. On the scientific side, across 7 low-homolog families from the CASP17 blind test set, Carbon retrieved roughly 40-fold more homologs on average than the baseline. The post also references Three.js prototypes, including a Three.js Eiffel Tower exhibition, a Google Mountain View game, a Merchandise Store redesign and a Koerner Hall fortepiano model, generated inside the /swarm preview in Google Antigravity that coordinates agent teams, with Carbon handling architecture, coordination and auditing.
Safety and honesty
Google reports adaptive red-teaming blocked 99 percent of attack trajectories and Trajectory Labs external testing recorded zero new universal jailbreaks across 132 hours and over 13,000 requests. The per-attempt misuse breach rate on Carbon is 0.8 percent against Argon's 3.1. Gray Swan's automated cyberattack query success rate landed at 2.17 percent against Argon's 9.86, and the Gray Swan indirect prompt injection benchmark registered 0 percent attack success, which the post flagged as announced two weeks earlier and still an area where robustness needs work. Google's honesty benchmark covers 375 tasks: Carbon passed 92 percent against Argon's 70, declined to fabricate missing experimental data 98 percent of the time and flagged methodological flaws 68 percent of the time.
"Google's cyber results were run without guardrails and have not been independently replicated." — Tom's Hardware, 2026-11-10
The pricing frame
The intro-pricing window undercuts the current frontier field on throughput-weighted cost. Anthropic's Fable 5.2, released Monday at $8 and $40 per million tokens, now sits four times above Carbon's intro rates on input and output, which is the main reason Google's benchmark deck emphasized cost per task on CursorBench, WANDR, AutomationBench and MKA. Google did not disclose Carbon's context window or parameter count; those remain undocumented in the GA post. Independent benchmark coverage from Tom's Hardware and VentureBeat flagged the cyber results running without guardrails and the limited independent replication on the agentic benchmarks. The December 1 price doubling closes the gap with Fable 5.2 to roughly half on input and half on output; even then, Google's pitch is that Carbon arrives with the full Antigravity agent coordination layer, AI Studio Build, Android Studio and AI Mode distribution that Anthropic's model doesn't carry natively.