Google Antigravity gives 10,000 researchers six months of free top-tier access for /teamwork
Google has opened applications for 10,000 academic researchers to receive six months of free access to the highest usage tier of Google Antigravity, its agentic development app, and has published a long report on what its multi-agent /teamwork feature produced when paired with Gemini 4 models. The announcement came on November 13, 2026, in a Google Research post signed by Vahab Mirrokni, Google Fellow and VP at Google Research, which states that external scientists and developers can use Gemini 4 models with /teamwork and /boost in Antigravity, and with Deep Review, starting the same day.
The free program targets the people most likely to run long, compute-heavy agent sessions: academics working on proofs, simulations, algorithm design and large code rewrites. Google describes the offer as six months on "our highest usage tier", which on Antigravity's own science page is tied to Gemini 4 Argon, the model Google calls its most advanced. Applications go through a Google form linked from the post and from the Antigravity science page, and Google has not published selection criteria, a deadline or the review timeline.
The report itself reads as a catalog of results that Google attributes to /teamwork and its formal-verification pipeline. Among them are a full proof of the Courtade-Kumar conjecture formalized in roughly 50 million lines of Lean 4, an autonomous 1.8-million-line Lean formalization of the Weil conjectures produced across 28,800 subagent invocations, and a roughly 800,000-line rewrite of the remaining C++ in Google's Zircon microkernel into safe Rust in about two weeks. Every one of these figures is self-reported by Google and its academic partners, and most come without an independent baseline.
··········
GOOGLE ANTIGRAVITY /TEAMWORK AND /BOOST: WHAT EXTERNAL USERS CAN RUN NOW.
/teamwork is Antigravity's slash command for multi-day, milestone-based projects. According to the Antigravity documentation, a session starts with a scoping interview in which the main agent asks about objectives, requirements, an objective verification method for each requirement, acceptance criteria and a working directory, which defaults to a folder under ~/teamwork_projects. The interview ends with a prompt artifact that the user approves with a single confirmation, after which the run proceeds autonomously.
Once approved, a role called the Sentinel records the request, routes work and posts progress updates. A Project Orchestrator breaks the brief into milestones and hands off to a fresh successor agent between milestones, so that no single conversation has to carry days of accumulated context. Read-only Explorers research the repository, Workers implement changes in non-overlapping tracks with one owning Worker per file, a Critic reviews correctness and interface conformance, a Challenger writes adversarial tests and edge cases, and an Auditor checks that reported test results match real command output. A Success Auditor runs a final end-to-end verification before the Sentinel presents the project.
The Google Research post describes the same design from the research side. /teamwork routes each prompt to a workflow "pattern" such as distributed coding, iterative problem-solving, long proofs, document review or rewrites, gives each specialized agent only the files and tools its step needs, and relies on automatic "falsifiers" to test candidates. The falsifiers named in the post include the Lean 4 proof checker, unmodified test suites, execution simulators, web search and hardware counters. Critics and auditors work from clean contexts and do not see the reasoning of the agent that produced the work they are checking.
/boost is the narrower tool. The documentation describes it as an on-demand multi-agent pipeline for hard software problems such as race conditions, intermittent deadlocks, algorithm work and tightly coupled refactors. An orchestrator routes the task to an implementation path or a root-cause analysis path, a first worker makes a candidate change and runs tests, follow-up workers try to break that output and fix what they find, and the orchestrator verifies the result against each requirement before delivering a summary. Google's post explains that /boost packages the iterative coding pattern of /teamwork after an analysis of past runs showed that tightly coupled code responds better to sequential implementation and revision than to parallel splits. Runs last anywhere from seconds to hours.
Users also choose an integrity mode during the /teamwork interview. The default development mode only flags fabricated outputs and facade implementations, demo mode forbids copying core logic from open source or reading tests to reverse-engineer expected behavior, and benchmark mode requires a from-scratch implementation using only the standard library. For researchers who want to publish results produced with the tool, the choice of mode determines how much of the output can be presented as original work.
··········
ANTIGRAVITY PLANS, FREE RESEARCHER ACCESS AND WHERE /TEAMWORK FITS.
Both /teamwork and /boost require a paid plan, according to Antigravity's slash-command catalog, and both run in the Antigravity App and in the Antigravity CLI. The free individual plan keeps basic weekly rate limits and the everyday agent commands, while some orchestration features sit only on the Google AI Pro and Ultra subscriptions.
........
Antigravity feature | What it does | Plan required |
/teamwork | Multi-day agent teams with milestones and audits | Paid plans |
/boost | Sequential solve-and-verify loop for hard bugs | Paid plans |
/swarm | Named workers and a verifier on a shared board | Google AI Pro and Ultra |
/run-workflow | Python script orchestrating many agents | Google AI Pro and Ultra |
/goal, /plan, subagents | Autonomous goals, plans, background tasks | All plans, including free |
Researcher program | Highest usage tier free for six months | 10,000 academics, by application |
........
Antigravity's pricing page lists a free individual plan with basic weekly rate limits, Google AI Pro and Google AI Ultra with higher rate limits and a flexible AI credit pool, and a business plan through Google Cloud starting at $30 per seat per month for Standard and Plus seats, alongside a pay-as-you-go option with no seat fee. In its May 19, 2026 plan update, Google set Google AI Pro at $20 a month and split Google AI Ultra into a $100 tier with five times the Pro token allowance and a $200 tier with twenty times the Pro allowance. The "highest usage tier" in the researcher offer therefore corresponds, on the published structure, to the plan Google sells at $200 a month, which puts the nominal value of six free months at about $1,200 per accepted researcher, or up to $12 million across all 10,000 places if every place is filled at that rate. Google has not confirmed that the program uses exactly that plan, so the calculation rests on the published price list.
The arithmetic is a useful guide for who should apply. A researcher who already pays for Google AI Pro to try /teamwork on small projects would move from the Pro allowance to twenty times that allowance for half a year, which is the difference between occasional runs and the hours-to-days sessions Google describes in its report. For readers comparing the free plan with paid access more generally, our guide to whether Google Antigravity is free, its pricing and its limits covers the baseline tiers, which have since been reorganized around the Google AI plans shown above.
Deep Review, which Google bundles into the same announcement, is a separate product on the Google Research portal. Google introduced it on October 14, 2026 as an agentic reviewer for technical documents, with two free document reviews per week for every user and higher limits for users who connect a Gemini API key. At STOC 2027, 578 submissions, more than a third of the total, opted in to Deep Review feedback. In a survey of 70 authors, 90% rated the feedback mostly or almost entirely factual and 90% said they would use it again, while over 30% reported that it caught substantive proof bugs or significant typos.
··········
/TEAMWORK FROM AUGUST PREVIEW TO GEMINI 4 ARGON.
The feature itself is three months old. Google first offered it on August 27, 2026 as /teamwork-preview on all paid Antigravity plans, at a time when the headline model for the work was Gemini 3.7 Flash, used together with Gemini 3.1 Pro for long proofs. The November report moves the feature to the Gemini 4 generation and drops the preview label from the command name in the documentation.
........
Google Antigravity /teamwork | August 2026 preview | November 2026 report |
Command | /teamwork-preview | /teamwork, plus /boost |
Main models named | Gemini 3.7 Flash, Gemini 3.1 Pro | Gemini 4 Argon, with Carbon and Hydrogen |
Availability | All paid plans | Paid plans; free top tier for 10,000 academics |
Headline math results | Seven open problems, 71% on TCSBench | Courtade-Kumar proof, 1.8M-line Weil formalization |
Headline systems result | RISC-V simulator booting xv6 | 800,000-line Rust rewrite of Zircon |
........
Google's post names Gemini 4 Carbon and Gemini 4 Hydrogen in its introduction, but most of the reported results are attributed to Gemini 4 Argon. For paying users the model choice affects cost, because Antigravity's shared Gemini quota counts usage by API price, so a long /teamwork run on Argon draws down the allowance faster than the same run on a cheaper model. Developers who want to compare the models before committing a multi-day job can read our coverage of Gemini 4 Hydrogen and its API pricing and of Gemini 4 Carbon for developer agents.
The November post also lists work that external users cannot yet reproduce. Google says the rewrite pattern behind the Rust results is "soon to be available in Antigravity", that the formal proof pattern and other specialized workflows will arrive in upcoming updates, and that it will publish deep-dive tutorials on the Antigravity blog in the coming weeks. The post states that "many of these features are available in Antigravity" and promises further updates to bring all of them to external users.
··········
[[ADPLUS_300x250_1]]
GOOGLE'S REPORTED /TEAMWORK RESULTS IN MATHEMATICS AND FORMAL PROOFS.
The mathematics section is the longest part of the report. Google says the Courtade-Kumar conjecture, posed in 2013, is now proved in full generality. Researchers at CUHK had reduced it to a 14-variable optimization problem, and Gemini found counterexamples to a generalized version, extended the proof to unbalanced functions and helped develop the bounds. The resulting Lean 4 formalization runs to about 50 million lines, and the work is credited to Google, CUHK, USC and CMU.
........
Problem or task | Result reported by Google | Verification named |
Courtade-Kumar conjecture (2013) | Proved in full generality | Lean 4, about 50M lines |
Weil conjectures | Autonomous formalization, 28,800 subagent calls | Lean, about 1.8M lines |
Hilbert's tenth problem, rings of integers | First formalization of a recent proof | Lean, over 919,000 lines |
Deque Conjecture (Tarjan, 1985) | New upper bound on splay-tree overhead | Lean 4 |
Boolean matrix multiplication | Deterministic algorithm answering STOC 2024 question | Paper result |
Quantum query complexity | Functions needing more than n/2 queries | Paper result |
........
The formal-verification results come from a pipeline Google calls LEAP, paired with Gemini 4 Argon and built with Google Cloud AI colleagues. Beyond the Weil and Hilbert formalizations, the post claims an end-to-end formalization of a known 3D Navier-Stokes blowup construction with smooth forcing, first formalizations of the Birkhoff conjecture and the 3D Bochner-Riesz conjecture, and a gap found and fixed in a recent informal proof of the latter. It also reports C¹ regularity results for infinity-harmonic functions in three and higher dimensions, and machine verification of informal proofs from a UCLA group, including the 40-year-old Jamison Caterpillar Conjecture and an improved bound on Erdős Problem #1063.
The theoretical computer science results are framed as answers to long-standing questions. Google reports near-optimal sample bounds for adaptive data analysis, closing a gap open since 2014, a deterministic Boolean matrix multiplication and triangle detection algorithm that reaches the target speedup over cubic time named as the main open question of STOC 2024, and the first n-bit Boolean functions that need strictly more than n/2 quantum queries at constant error. Several of these are listed as joint work with named academics, such as Samson Zhou at Texas A&M on adaptive data analysis and Sorrachai Yingchareonthawornchai at ETH Zurich on the Deque Conjecture.
The distinction between machine-checked and unchecked results is the most useful filter for readers. A Lean 4 formalization that compiles is a verified statement about the theorem as formalized, while results described only as algorithms or bounds still depend on human review of the corresponding papers, and the post does not say which of them have been peer reviewed or accepted at conferences.
··········
ANTIGRAVITY /TEAMWORK IN SYSTEMS ENGINEERING, RUST REWRITES AND GOOGLE INFRASTRUCTURE.
The engineering results are closer to what most Antigravity subscribers will try. Google says /teamwork rewrote the remaining C++ of the Zircon microkernel into roughly 800,000 lines of safe Rust in about two weeks, building on an earlier port that had covered 25% of the code, and that the result passes kernel unit tests and boots on arm64 hardware. A second rewrite turned the libgav1 AV1 decoder into about 170,000 lines of Rust with 100% bit-exact conformance. The Rust version started about four times slower than the original, and after profiling Google reports it ran 20-30% faster than the C++ and hand-optimized assembly on x86_64 and ARM.
........
Engineering task | Reported outcome | Comparison point |
Zircon microkernel to Rust | About 800,000 lines in about two weeks | Earlier 25% port |
libgav1 AV1 decoder to Rust | 100% bit-exact, 20-30% faster | Original C++ and assembly |
FarArray connected components | 60x-90x throughput, 6x lower RAM | TCP key-value stores |
JAXBench TPU kernels | 4.09x geometric-mean speedup, 17 kernels | Reference kernels |
vLLM TPU Pallas kernels | 2.99x speedup, 7 kernels | Hand-optimized kernels |
Borg TPU scheduling | 24% less workload preemption | Some hardware types |
........
On distributed systems, Google reports that FarArray, built on its SNAP framework, ran lock-free connected components over graphs with one trillion edges at 60 to 90 times the throughput of TCP key-value stores with six times less RAM, and that a Personalized PageRank job called InstantEmbedding used up to 14 times less peak DRAM while running 11 to 27 times faster than accelerated MapReduce. For TPU kernel work, the DuoDiscover workflow replaced 7-to-32-minute hardware evaluations with 23-second proxy measurements validated by rank correlation, then reported a 4.09x geometric-mean speedup across 17 JAXBench kernels and 2.99x across seven vLLM TPU-Inference Pallas kernels.
Two results come from Google's own production fleet. Reusable workflows found memory savings that exceeded an initial estimate of one pebibyte, although the post gives no final figure, and two shipped changes to the Borg scheduler cut TPU workload preemption by 24% on some hardware types while keeping service-level objectives and utilization intact. The report adds a 30-agent chemistry loop that reached 95% top-1 exact match and 100% top-3 accuracy in identifying complex benchmark molecules from NMR data, roughly double single-agent performance, and a 9x9 Go engine that gained about 2,600 Elo in under eight days to reach 1,885, a rating Google places between KataGo at 16 and 32 visits on the strength of a single showcase game.
··········
HOW TO APPLY AND START A /TEAMWORK PROJECT IN ANTIGRAVITY.
Academic researchers apply for the free tier through the form linked from the Google Research post and from the "Apply here" button on Antigravity's science page. The post limits the offer to 10,000 academic researchers and six months, and says nothing about country restrictions, required institutional email addresses or what happens to an account when the six months end, so applicants should expect the account to fall back to whatever plan they hold afterwards unless Google says otherwise.
Users who already hold a paid plan can start immediately. In the Antigravity App, a new conversation that begins with /teamwork followed by the goal launches the scoping interview, and the same command works at the Antigravity CLI prompt. Progress appears in the subagent panel in the App and in the status bar in the CLI, where Alt+J jumps to subagents waiting for approval. Hard single problems are better suited to /boost, which starts immediately without an interview, while broad tasks that split into independent pieces fit /swarm on Google AI Pro and Ultra.
Subagents inherit the file-access and command policies of the active workspace, and protected terminal commands or edits outside trusted scopes trigger an approval prompt, which means a multi-day run can stall overnight if it reaches a step that needs a human click. The documentation gives no hard runtime limit for /teamwork and does not state how much quota a typical run consumes, so the practical way to size a project is to start with a bounded goal in development mode and read the progress artifact before approving larger milestones.
··········
GOOGLE ANTIGRAVITY /TEAMWORK: CONFIRMED FACTS AND OPEN QUESTIONS.
Several points are documented on Google's own pages. /teamwork and /boost run on paid Antigravity plans in both the App and the CLI, /swarm and /run-workflow require Google AI Pro or Ultra, Deep Review offers two free reviews per week, and Google is accepting applications from 10,000 academic researchers for six months of its highest usage tier. The roles, interview flow and integrity modes of /teamwork are described in the official documentation.
The research results are Google's claims. The Lean formalizations carry their own form of verification, but the algorithmic and systems results lack independent replication, several lack baselines, and the Go rating rests on one game. Google has also not published the quota a /teamwork run consumes, the selection process and deadline for the researcher program, or a date for the rewrite and formal proof patterns that produced some of the most striking numbers.
For academics, the free tier lowers the cost of testing these claims on their own problems for six months. For paying developers, the November report confirms that /teamwork is now a standard paid feature built around Gemini 4 Argon, and the unpublished quota cost of long Argon runs remains the main unknown before planning multi-day projects.
··········
FOLLOW US FOR MORE.
·····
DATA STUDIOS
·····
datastudios.org
