NEWS
Google Adds Gemini 4 Carbon to a 27,000-Application Queue
Google added Gemini 4 Carbon to its 10,000-scientist program after cyber defenders, with API prices doubling on December 1.
Google added Gemini 4 Carbon to a free six-month Antigravity seat for 10,000 scientists on November 10, after cyber defenders had the model from October 28. Koray Kavukcuoglu, Google DeepMind’s senior vice president and Google’s chief AI architect, says Carbon uses more tokens than Gemini 4 Argon “to explore thoroughly, reason deeply, and run extended verification loops.”
The public line is a lab giveaway. The first users were security teams, a Signal proof stack, and paying API customers whose meter doubles on December 1.
Scientists Join a Line Cyber Defenders Already Cleared
Google opened Carbon to scientists and developers the same day it put the model in the Gemini API, Gemini for business, Google AI Ultra, and Antigravity. Argon stays the default for most software work, including the Argon agents that beat Minecraft. Carbon is the slower, heavier pass for long debugging, open-ended research, and cyber review.
The scientist offer is not a new program. It started on October 13 with Argon: six months on Antigravity’s highest usage tier, up to 10,000 people, reviewed every week. Google’s interest form, updated November 10, says the company had received more than 27,000 applications and granted upgraded access to the first 3,000 scientists. Current and future participants now get Carbon as well, and Google says later frontier models will follow.
Consumer Carbon in the Gemini app and AI Mode in Search is a Google AI Ultra feature. The paid API has no free Carbon tier. Cyber-tuned “Carbon for cyber” still runs through a separate gate.
Our strongest frontier model, Gemini 4 Carbon, is now available.
Built for deeply complex, long-horizon workflows that push a frontier model’s intelligence, Carbon delivers frontier performance in agentic workflows across complex knowledge work, software engineering and web… pic.twitter.com/KSNrOQ0VVj
— Google AI (@GoogleAI) November 10, 2026
THE ACCESS CALENDAR
- October 13, 2026: Opens the 10,000-scientist Antigravity seat with Gemini 4 Argon.
- October 28, 2026: Announces Carbon and gives it first to a subset of cyber defenders.
- November 10, 2026: Puts Carbon in the paid API and Ultra products, adds it to the scientist seat, and publishes the Signal proof work.
- December 1, 2026: Raises standard API rates from $2 and $10 per million tokens to $4 and $20.
That order is the story the 10,000-seat headline hides. Labs are joining a rollout that already had a security customer, a formal-verification demo, and a dated price cliff.
Zero-Days in Under Two Hours
Carbon’s first public life was not a notebook. On October 28 Google said it would start with a subset of trusted partners in its Fairwind Program for cyber defenders, then widen access after more guardrail work. Fairwind still holds the cyber-tuned build, which can run alone or with CodeMender. The public API is a different door.
On the open cyber benches Google cites, Carbon scored 88.2% on SEC-bench Pro for finding flaws, 54.8% on ExploitGym for exploitation, and 71.0% on CWE-bench for patching, all measured without guardrails. Internally, Google says Carbon more than doubled Argon’s success rate at writing secure code. The Chrome team reported it produced high-quality patches more than 10 times better than another lab’s large model on the tests they ran.
Wiz, which Google calls its leading cloud security platform, used Carbon in its Scan for Good work and, Google says, uncovered 0-day vulnerabilities at two major software-as-a-service providers in under two hours. The team then worked with those firms on fixes. Google’s Cloud Vulnerability Research group used the model to show that two bugs in a popular open-source project were exploitable at a severity it compared to PHP2Shell; maintainers recast them as critical and shipped patches. Names of the vendors and the project stay behind responsible disclosure.
Inside Google’s Project Fortify, Carbon hit 100% recall on scanning and 95% precision on triage on an internal benchmark, with more than a 25% relative gain over Argon on open-source recall. John White, director of the Advanced Security Initiatives Group at Cisco, said deep reasoning models like Carbon for cyber let defenders “evaluate complex codebases and determine whether vulnerabilities are genuinely exploitable.”
That is production security work, not a campus perk. Scientists now get the same weights through Antigravity. They do not get the Fairwind cyber pack unless they apply on that track too.
Over Six Million Lines of Lean on libsignal
The load-bearing research claim landed on November 10 beside the scientist update. Pushmeet Kohli and Swarat Chaudhuri at Google DeepMind described AlphaProof Code, an agent setup that writes machine-checked proofs about real software. Humans and mixed models drafted the specifications. Agents on Gemini Carbon then wrote the proofs.
The target was libsignal, the reference code behind Signal’s one-to-one encryption, which also sits under Google Messages. DeepMind says this is the first time the implementation of a widely deployed protocol has been formally verified at this depth, with or without AI. The agents produced over six million lines of Lean, a language whose kernel rejects a bad proof. They flagged 1 percent of the codebase for a second human look. The campaign took four weeks, work that expert teams have historically needed years to finish. When the team put two previously fixed bugs back into an isolated copy, the system proved those builds insecure.
While prior formal verification focused on smaller portions of the code, Google DeepMind’s approach enables significant expansion. Using AI agents to translate live code into valid Lean theorems is a major breakthrough. As AI-driven security threats grow, rigorously verifying the Signal protocol implementation gives us even greater confidence in our code’s privacy guarantees.
Ehren Kret, Chief Technology Officer, Signal
The result is a proof of specified properties of compiled code, not a stamp on every Signal feature. DeepMind open-sourced the artifacts. The scientist program does not automatically hand out that harness; it hands out Carbon, the model that wrote the proofs, on Antigravity’s top usage tier.
What the 40-Times Protein Result Measured
Google’s science write-up is specific, and it is narrower than a folding miracle. In a multi-agent test, researchers used Carbon to search sequence archives and structure databases for distant relatives of hard proteins. Across seven protein families from the CASP17 blind test set, the company says the model raised the number of related sequences retrieved by an average of 40 times. Those extra sequences feed structure tools such as AlphaFold 3. The figure is retrieval, and Google says broader benchmarking is still running.
THE RESEARCH CLAIMS GOOGLE PUBLISHED
| Job | What Google says Carbon did | What that number does not say |
|---|---|---|
| Protein homolog search | 40 times more related sequences, seven CASP17 families | Not a 40-times gain in fold accuracy |
| Lab notebooks and raw data | Flagged incomplete mixing, better peak splits, sample carryover | Not a published methods paper |
| Axiom Bio toxicity work | Company is moving production analyses onto Carbon and Argon; tests found mechanisms it had not seen | No independent tally of new findings |
| AlphaProof Code on Signal | Key properties of compiled libsignal in four weeks | Not a proof of the whole messaging app |
Other internal runs used Carbon on laboratory problems from notebooks and raw instrument files. Google says it spotted anomalies tied to incomplete automated mixing, improved chromatography peak splits, and found traces of sample carryover. Axiom Bio, a drug toxicity firm, is moving production scientific analyses and agents onto Carbon and Argon; Google says Axiom’s tests identified toxicity mechanisms the company had not previously found.
Those are Google’s own research notes, released with the model. The 10,000-scientist seat is how outside labs get the same weights long enough to try to reproduce, or fail to reproduce, that notebook. Derya Unutmaz, a professor of immunology, wrote that Gemini 4 Carbon and OpenAI’s Astra Pro are the two strongest AI scientists he has used in biomedicine. That is one lab voice, not a trial.
The Same Token Price, Until December 1
Carbon bills at Argon’s launch rate even though it is built to spend more tokens. Through November 30, 2026, the Gemini Developer API lists $2 per million input tokens and $10 per million output tokens, with cached input at $0.10. From December 1 those figures become $4, $20, and $0.20. Thinking tokens count as output. Carbon is not on the free API tier. Google’s footnote says the higher rates apply to the Carbon model meant for public access.
GEMINI 4 CARBON API RATES
| Meter (standard, USD per 1M tokens) | Through November 30, 2026 | From December 1, 2026 |
|---|---|---|
| Uncached input | $2.00 | $4.00 |
| Output, including thinking | $10.00 | $20.00 |
| Cached input | $0.10 | $0.20 |
Identical sticker prices still hide a higher bill, because Carbon is the model that keeps searching. A long Antigravity run can chew through a high Ultra allowance in an afternoon once thinking tokens start stacking. Google already shipped a cheaper teammate on November 12: Gemini 4 Hydrogen, with an Orchestrator mode that pairs it with Carbon. The Antigravity team said that mix reached near-Carbon success on Terminal-Bench 4.1 while using 24% less quota and finishing 19% faster. Scientists on the free highest tier are the group who can burn Carbon without watching that meter.
On a new Multimodal Knowledge Agent test built with Artificial Analysis, Carbon ranked second at more than 6 times lower average cost per task than the top model. Google also puts it first on CursorBench 4.0, Perplexity’s WANDR, and Zapier’s AutomationBench at more than 2 times lower cost per task than the next outside model. The WebDev Arena score rose 89 Elo points to 1,768, in fourth place. Those are vendor-cited leaderboards, useful as a cost argument, not as a substitute for the protein caveat.
Who Can Apply for the Six-Month Seat
The form is still open. Selection is rolling, weekly, and not a guarantee. Access lands on the Google account used to apply, as an individual Antigravity upgrade, not as a campus Gemini Enterprise contract. Google says it weighs scientific credentials and likely research impact, and it has asked later applicants to wait because demand is high.
Antigravity is also /swarm preview inside Antigravity, which turns one prompt into a tracked team of agents. Carbon can take architecture, coordination, and audit roles while faster models build and test in isolated workspaces. That is the tool the free seat actually unlocks, not a raw API key.
WHO THE FORM ACCEPTS
- The work: Non-commercial scientific research only.
- The job: An active post at a university, degree-granting school, non-profit research institute, or national laboratory, including faculty, postdocs, staff scientists, research engineers, and graduate students.
- The ID: A public lab or group URL, plus a valid ORCID iD or Google Scholar profile.
- The account: Personal Gmail or institutional Workspace, so long as that login can already reach Antigravity; the upgrade follows that login.
Cyber defenders who want Carbon for cyber still apply to Fairwind, which Google describes as a limited channel for governments, critical infrastructure, and trusted security partners. The two queues share a model family and almost nothing else in paperwork.
Carbon Flags Missing Lab Data Instead of Filling It
Google’s safety notes are the other reason labs might want this build. On an internal set of 375 tasks about agent honesty, Carbon passed 92%, against Argon’s 70%. Google says Carbon refused to fill in missing experimental data 98% of the time and flagged methodological flaws in 68% of cases. On Gray Swan’s indirect prompt-injection test, Carbon’s attack success rate was 0.0%, after Argon’s 0.7%. Terminal-Bench 4.0 cheating success fell to 0.0%, with cheat attempts at 0.4%. Live internal coding alerts for high-severity misalignment dropped by almost 80%. Overreach fell 26% and under-service 29% on Google’s own tallies.
The same October 28 post says more advanced internal tests still find places where resistance to sophisticated attacks needs work. That clause sits next to the 0.0% Gray Swan figure, and both are Google’s. For a scientist, the useful behavior is the 98% refusal to invent a missing data point. For a security team, the useful behavior is the 0-day hunt that Fairwind already ran.
The 10,000-scientist seat is real, and 3,000 people are already on it. Carbon still arrived there after Wiz, after Cloud Vulnerability Research, and after a four-week Lean campaign on libsignal, and the public API price still doubles on December 1. Google open-sourced the Signal proof files. The Antigravity form remains open on a weekly review.
-
NEWS3 months agoMicrosoft 365 Auth Fault Took Down Exchange and Teams
-
GAMING3 months agoXbox Caps Game Pass Cloud Gaming at 15 Hours
-
NEWS3 months agoOpenClaw 2.0 Ships a Team Workplace With Host-Trust Defaults
-
NEWS3 months agoOpenAI Unveils GPT-6 Astra With a Critical Cyber Label
-
LIFESTYLE3 months agoSquishy Dumpling Toys Recalled After Hiding Illegal Water Beads
-
BUSINESS3 months agoDiesel Breaks Its Record as the White House Claims Credit
-
NEWS3 months agoIFA 2026’s Weird Gadgets Are Building a Sensor Home
-
BUSINESS3 months agoChargePoint Stock Rally Prices Wilmer’s Three-Year Cash Plan
