Press "Enter" to skip to content

Google expands free AI access for scientists with Gemini 4 Carbon

Google DeepMind announced that its latest large‑language model, Gemini 4 Carbon, will join the Antigravity scientist access program. The move broadens the initiative that already provides Gemini 4 Argon to researchers, with the goal of granting up to 10,000 scientists six months of complimentary usage of advanced AI tools.

Program expansion and eligibility

The scientist program, which began with the Argon model, now welcomes applications for Carbon on a rolling weekly basis. Interested researchers can apply for the highest‑usage tier on Antigravity and, if selected, receive six months of free access. The program remains limited by eligibility criteria, overall capacity, and a selection process that continues to operate weekly.

Carbon’s public rollout to paying developers, enterprises, and Google AI Ultra subscribers is scheduled for 10 November. While the paid offering will be priced at $2 per million input tokens and $10 per million output tokens—identical to Argon’s introductory rates—the free scientist tier will remain cost‑free for qualifying participants.

Early research outcomes

Google highlighted several internal experiments that illustrate Carbon’s capabilities in scientific domains. In protein research, a multi‑agent workflow used Carbon to probe sequence archives and structural databases for distant evolutionary relatives of proteins. Across seven protein families drawn from the CASP17 blind‑test set, the model reportedly retrieved, on average, 40 times more related sequences than baseline searches. Those additional sequences can enrich structure‑prediction pipelines such as AlphaFold 3, although Google clarified that the metric reflects sequence retrieval rather than a direct boost in prediction accuracy.

Other internal studies examined laboratory notebook entries and raw experimental data. Carbon identified anomalies linked to incomplete automated mixing, suggested improved chromatography‑peak separation methods, and detected traces of sample carry‑over in research instruments. Axiom Bio, a drug‑toxicity research firm, is reportedly transitioning both Carbon and Argon into its production pipelines, where the models helped uncover toxicity mechanisms that had previously escaped detection.

In software verification, researchers combined Carbon with other models under a framework named AlphaProof Code. The effort translated security goals for Signal’s cryptographic protocol into formal theorems, generating millions of lines of Lean code to prove key security properties. Google said the team completed the verification in weeks, a timeline that traditionally stretches into years for expert verification teams. The claim pertains to specific properties rather than a blanket proof of the entire messaging service.

Cybersecurity applications

Carbon’s release follows an earlier, limited‑access window for cybersecurity defenders through Google’s Fairwind Program. The program continues to accept applications for Carbon’s defensive capabilities. Owen Larter, Senior Director and Head of Policy and Public Affairs at Google DeepMind, emphasized the model’s role in linking vulnerability discovery, verification, and automated patching.

External partner Wiz used Carbon to uncover previously unknown vulnerabilities in two major software‑as‑a‑service platforms within a two‑hour window. The affected providers collaborated with Wiz on remediation, though their identities were not disclosed. Google’s Cloud Vulnerability Research team also leveraged Carbon to demonstrate exploitable bugs in open‑source software; maintainers subsequently classified the bugs as critical and issued patches.

John White, Director of the Advanced Security Initiatives Group at Cisco, remarked that integrating deep‑reasoning models like Carbon enables defenders to assess complex codebases and confirm the exploitability of identified flaws.

Safety testing, pricing and multi‑agent tools

Google reported internal safety benchmarks that compare Carbon with Argon across 375 honesty‑related tasks. Carbon achieved a 92 % pass rate, whereas Argon recorded 70 %. The model also refused to fabricate missing experimental data 98 % of the time and flagged methodological flaws in 68 % of cases. On the Gray Swan indirect prompt‑injection benchmark, Carbon recorded a 0 % attack success rate, though Google noted that ongoing testing continues to reveal areas for improvement.

In parallel with the model launch, Google previewed a new “/swarm” feature within Antigravity. This capability allows a single user prompt to spawn a coordinated group of agents, each handling distinct subtasks such as architectural planning, coordination, or auditing. Faster models can be assigned to routine work while Carbon tackles more demanding steps. The system can operate across separate workspaces and includes agents that test proposed changes.

Google also highlighted enhancements in handling dense documents, charts, images, and tables. Outputs can now be formatted as spreadsheets, presentations, or reports. In a multimodal knowledge‑agent benchmark co‑developed with Artificial Analysis, Carbon placed second while costing less than one‑sixth of the average expense per task of the top‑performing model.

Access pathways for developers and enterprises

Paid developers can reach Carbon through the Gemini API available in Google AI Studio and Android Studio. Enterprise customers have multiple routes, including Gemini for Business, Antigravity, and CodeMender. Google AI Ultra subscribers receive Carbon via Antigravity, the Gemini app, and the AI Mode in Search.

Scientists seeking the free six‑month tier must apply separately on Antigravity’s highest‑usage tier. Selection remains contingent on eligibility, capacity, and the weekly rolling review process.

Overall, the addition of Gemini 4 Carbon to the Antigravity scientist program signals Google’s intent to accelerate AI‑driven discovery across biology, laboratory automation, software verification, and cybersecurity, while also providing a structured path for broader commercial adoption.

Be First to Comment

Leave a Reply

Your email address will not be published. Required fields are marked *