On November 16, 2026, Google DeepMind announced that it would include its flagship reasoning model Gemini 4 Carbon in the Antigravity scientist program, providing up to 10,000 researchers with six months of free access; OpenAI simultaneously announced that it would expand its academic researcher program from about 20,000 people to up to 200,000, adding 50,000 free slots, with each approved participant also able to invite two additional collaborators to join.
Structural Differences Between the Two Programs
Google's Antigravity program takes a selective approach: 10,000 places are reviewed on a rolling weekly basis, prioritizing high-intensity, multi-step reasoning tasks in physics, mathematics, and the life sciences. The program had previously opened Gemini 4 Argon to participants; the newly added Carbon is DeepMind's higher-tier reasoning model, officially priced at $2 per million input tokens and $10 per million output tokens (this is an introductory price until November 30, after which it will adjust to $4/$20).
Koray Kavukcuoglu, Senior Vice President at Google DeepMind and Google's Chief AI Architect, described Carbon's design logic in the release notes: "Carbon uses more tokens than Argon to explore thoroughly, reason deeply, and run extended verification loops." Argon is the recommended option for the vast majority of software engineering and knowledge work, while Carbon is reserved for more complex research reasoning scenarios.
OpenAI's plan has broader coverage. The program launched a pilot in July with 20,000 researchers across 33 countries; this expansion directly raises the cap to 200,000 people, with each approved slot allowing two collaborators, meaning actual coverage could approach 600,000. Applications open on November 23, and OpenAI explicitly committed not to use program users' data for model training.
What Concrete Evidence Exists for Carbon's Real-World Capabilities?
Google provided a set of traceable, concrete experimental results: in protein prediction, researchers used Carbon to search for distant evolutionary relatives across seven protein families in the CASP17 blind test set, achieving an average 40-fold increase in the number of relevant sequences retrieved. This figure refers to the increase in the number of sequences retrieved, not an improvement in AlphaFold 3's structural prediction accuracy.
In laboratory data analysis scenarios, Carbon was used to process lab notebooks and raw experimental data, identifying anomalies in automated sample-mixing workflows, improving chromatographic peak separation methods, and detecting residual sample signals in research instruments. After drug toxicity research company Axiom Bio migrated production-grade analytical work to Carbon and Argon, its tests identified previously undiscovered toxicity mechanisms. In security, cybersecurity company Wiz used Carbon to discover previously unknown vulnerabilities in two large SaaS vendors within two hours.
In terms of safety benchmarks, Carbon achieved a 92% pass rate on agent honesty tests, higher than Argon's 70%. This gap means that in automated scientific research workflows, Carbon has a lower probability of producing misleading outputs.
The Real Meaning of the Pricing Competition
Gemini 4 Carbon's current introductory price of $2/$10 (per million tokens) is significantly lower than Claude Fable 5.2's $8/$40; even after it rises to the standard price of $4/$20 on November 30, it will still be on par with Claude Opus 5.5's price structure. For computationally intensive scientific reasoning tasks, the pricing gap will be directly amplified by task volume.
This means the free scientist programs are not merely a public relations gesture. For institutional research teams that may convert into paying users after the six-month free period, the price benchmark has already quietly taken shape during the free period. Whichever model first becomes the default tool in laboratory workflows will gain a negotiating advantage in subsequent procurement negotiations.
Gains and Losses for the Parties Involved
For scientists, the simultaneous existence of two free channels to apply for is a rare window of opportunity. OpenAI's program has broader coverage and a lower application threshold, making it suitable for individual researchers who want to quickly get started with frontier models to assist their daily research; Google's Antigravity program has fewer slots but already has thousands of scientists participating, and is prioritized for professional scenarios requiring intensive computation and multi-step verification. Neither program currently restricts researchers from applying to both at the same time.
For enterprises and developers, a byproduct of this competition is a set of more credible real-world validation data. Commercial organizations such as Axiom Bio and Wiz have appeared in Google's case list, and these are measurement results from real workflows. When such data accumulates to a certain scale, the decision weight in model selection will shift from paper leaderboards to actual laboratory results.
For competitors, the number 200,000 means OpenAI is building academic penetration through scale effects, while Google is building scientific credibility through result density. Both paths will compress the middle layer—model providers that cannot participate in similar programs will find it difficult to engage in the contest for scientific endorsement.
© 2026 Winzheng.com 赢政天下 | 转载请注明来源并附原文链接