Gemini 4 Argon: Pricing, Benchmarks and What the Google Model Can Do

Google released Gemini 4 Argon on October 13, 2026, a frontier AI model for coding, science, cybersecurity and long agentic work, priced at 2 dollars per million input tokens at launch. The announcement, written by Google DeepMind’s Koray Kavukcuoglu, is long on case studies and short on benchmark tables, and it is worth reading closely for what it says and for what it leaves in a footnote. You can read the full Gemini 4 Argon announcement on the Google blog. Two short Google DeepMind videos released alongside it, one on drug discovery research at Isomorphic Labs and one on a humanoid robot learning cornhole, are embedded below with notes on what they add.

TLDR

Gemini 4 Argon is available now through the Gemini API, Google Antigravity, the Gemini app and AI Mode in Search, with an introductory price of $2 per million input tokens and $10 per million output tokens that doubles after November 30, 2026. Google makes its case with work the model has already done: a refactor inside Chromium, speedups to open-source projects, security findings with Wiz, and research results in quantum computing, fusion, rare disease and enzyme design. It claims the top spot on the Vals Index, an 87% score on the ERQA embodied reasoning benchmark, and a 0.7% attack success rate on the Gray Swan prompt injection benchmark. The safety section describes a model trained to stop and ask before irreversible steps, plus a monitor that reads the agent’s reasoning in real time.

Thoughts

The safety section holds the most consequential engineering in the whole announcement. Google says it trained Argon to respect system boundaries, to report negative findings honestly, and to pause for a human before doing anything irreversible, and that attempts to take a shortcut around a roadblock fell to 2% on coding benchmarks. That matters more than any leaderboard position, because every impressive example elsewhere in the post depends on it. An agent that works on a production codebase overnight with a one-paragraph goal is only useful if it does not quietly delete the failing test to make the build pass. Google also says it built a monitor that reads the agent’s step-by-step reasoning to block workarounds as they happen, and that it deliberately avoided penalizing the model’s thoughts during training so that the reasoning stays readable. That second choice is the careful one. Train a model against its own visible reasoning and you risk teaching it to hide the reasoning, which leaves you with a monitor that watches nothing.

The price in the headline is not the price. The $2 and $10 figures are introductory, and footnote one says that after November 30, 2026 the model costs $4 per million input tokens and $20 per million output tokens. Anyone building a product on Argon has about seven weeks at half price, so the budget should be written against the second number. The more telling detail is the 95% discount on cached input tokens, paired with a 1 million token limit shared between input and output. That combination is priced for agents that reread the same large context hundreds of times across a long job, and it says a lot about the workload Google expects: fewer chat turns, more hours-long runs over a whole repository or a whole document set.

What sets this launch post apart from most is that part of the evidence can be checked. The speedups to the Agent Development Kit, to the xTB quantum chemistry package and to Gemma inference under vLLM are described as merged or submitted changes in public repositories, with specific numbers attached (a 40% cut in session prep time, up to 2.3x faster structure prediction, a 2.85x inference speedup). A reader can go and look at the diffs. The internal figures deserve more caution. A 35% rise in agentic code submissions is a measure of volume, and volume is not the same as value. The companion number, rollbacks down by nearly a third, is the one that suggests quality held, and neither can be verified from outside Google. The Isomorphic Labs video shows the right habit for everyone else: the researcher first gave the model a small set of patents he already knew well, and only trusted the output after comparing it with what his team had on file.

The robotics demonstration is easy to misread, and the companion video corrects the blog post in a useful way. The written version reads like a single pass: watch a video of a person playing cornhole, write the simulation, transfer it to Apptronik’s Apollo, sink seven shots. The video describes a loop. The model practiced in the MuJoCo simulator, reviewed footage of each failed attempt, changed the stance and the swing, and then on the real robot watched each bag land and adjusted the release by a few centimeters. So the model is doing two jobs, the robotics engineer who builds the simulation and the coach who reviews the tape. What it is not doing is running the fast, moment-to-moment motor control. If this holds beyond one internal test, the scarce input in robot training shifts from hand-built simulations to ordinary video, which exists in nearly unlimited supply. One test with one skill is thin evidence, so treat it as a direction and not a result.

The cybersecurity section shows how access to frontier models is splitting into tiers. The public Argon has cyber guardrails. Vetted defenders in the Fairwind Program get a version without them, and Google is extending that to hundreds more partners. The reasoning is sound: the capability that chains several vulnerabilities into an administrator key at a healthcare provider is the same capability an attacker would want. The consequence for ordinary security teams is that the model they can buy today is deliberately less capable at their job than the one a competitor with program access is using, and the application queue becomes part of the product.

Key Takeaways

Gemini 4 Argon launched globally on October 13, 2026 through the Gemini API in AI Studio and the Gemini platform, in Google Antigravity for AI Ultra subscribers, and in the Gemini app and AI Mode in Search for AI Ultra and Pro subscribers.
Introductory pricing is $2 per million input tokens and $10 per million output tokens, with cached input tokens at 95% off. After November 30, 2026 the price becomes $4 and $20.
The Gemini API offers a 1 million token limit that is unified across input and output, along with a new continuation protocol for long-running work.
Google says Argon ranks first on the Vals Index, which measures performance on finance, coding, legal and tax work.
Inside Google, Argon in Antigravity raised agentic code submissions by more than 35% and cut code rollbacks by nearly a third over a few weeks.
The Chrome team used it to separate core webpage handling from a legacy folder of about 6 million lines inside Chromium’s 40 million line codebase, finishing months of refactoring in weeks.
Open-source results include a 40% cut in session prep time and 21% faster cold imports in Google’s Agent Development Kit, xTB running up to 2.3x faster, and a 2.85x inference speedup for Gemma 4 E2B on vLLM.
At Isomorphic Labs, a researcher says Argon produced reports on 10 chemistry patents in an afternoon, work he puts at nearly a month by hand, and summarized academic papers in 10 to 15 minutes that would have taken him one or two days.
Wiz used Argon in its Scan for Good initiative to close a gap exposing government data and to fix a file download bypass that exposed a master key protecting hundreds of thousands of patient records.
In quantum computing, Argon nearly doubled the inference throughput of an internal version of the AlphaQubit decoder and wrote an arcsin implementation, verified in Lean, that cuts spacetime cost by more than 100x against a published baseline.
A Google team working with Professor Matt Landreman found a broader class of counterexamples to Grad’s conjecture, a 1967 hypothesis about stellarator magnetic fields.
Argon scored 87% on the Embodied Reasoning (ERQA) benchmark. Starting from a cornhole video, it built a MuJoCo simulation, practiced, and tuned the throw until Apptronik’s Apollo humanoid sank seven consecutive shots.
On safety, shortcut attempts on coding benchmarks fell to 2%, external bio red teaming unlocked less than 1% of tested bio sub-tasks, and prompt injection attack success across 15 adaptive attempts fell to 0.7% on the Gray Swan benchmark.
Google is offering 10,000 scientists its highest Antigravity usage tier free for six months, subject to eligibility, capacity and selection.
Argon is the first model in the Gemini 4 family, and Google says more are coming.

Detailed Summary

Availability and Gemini 4 Argon Pricing

Argon is available to developers and enterprises through the API, to AI Ultra subscribers in the Antigravity coding environment, and to Ultra and Pro subscribers in the consumer Gemini app and in AI Mode in Search. Google positions it for deep, multi-step work in coding, finance, legal, cybersecurity and science. The launch price of $2 input and $10 output per million tokens is introductory and doubles after November 30, 2026. Cached input is discounted 95%.

Engineering Work Inside Google

Google leads with its own use of the model. The Chrome refactor pulled core webpage handling out of a legacy folder across hundreds of files, which Google says will let live pages move between tabs, side panels and floating windows without reloading. The Agent Development Kit team gave Argon its skills rulebook and a one-paragraph goal, then let it work overnight. Its optimizations are now live in the public repository.

Open-Source Speedups and 3D Generation

Researchers used Argon to speed up xTB, a widely cited quantum chemistry package used to predict molecular shapes and screen drug candidates. A separate effort took Gemma 4 E2B from about 37,900 to more than 108,000 tokens per second on an A100 with identical output, in days and not months. Google also shows a run of visual demos built with multi-agent critique loops in Antigravity, including a temple complex sculpted in Blender, a racing game built from photos of a desk, and a ray tracer written in pure Python.

Business and Professional Workflows

For professional work Google stresses reasoning, visual perception, domain knowledge and writing. It says Argon follows style instructions better than earlier Gemini models, hedges less, repeats itself less and avoids clichés. Examples include reconstructing Chaco Canyon in 3D from about 50 GB of LiDAR scans and archival drawings, and turning a folder of project notes into a client proposal deck. Isomorphic Labs reports that extracting information from complex PDFs and patents went from weeks to hours.

Isomorphic Labs: 500-Page Chemistry Patents in an Afternoon

The blog post gives the Isomorphic Labs work one sentence. This two and a half minute video, featuring researchers Brian Bender and Santosh Adhikari, shows what the job looks like. A chemistry patent can run to 500 pages, with a molecule’s structure sitting 50 pages away from its experimental readout, so pulling the data out by hand is slow and error-prone. One researcher tested Argon on a small set of patents he knew well. It extracted them into a spreadsheet, then went further and wrote a full report, which he found highly accurate when compared with his team’s own records. The other pointed it at academic papers on a protein target and got a report in 10 to 15 minutes that connected ideas across fields. They are clear about the division of labor: Isomorphic builds specialized models for specific drug discovery tasks, and a general model like Gemini handles the reading, summarizing and orchestration around them.

Cybersecurity and the Fairwind Program

Google is expanding its Fairwind Program to hundreds of additional partners, who can apply for access to Argon without cyber guardrails. Wiz used early access for its free Scan for Good work on critical infrastructure. In one case, with a federal agency, the model helped find and close a gap that risked exposing sensitive data. In another it chained vulnerabilities to uncover a bypass that exposed a master key at a healthcare organization.

Quantum Computing, Fusion, Rare Disease and Enzymes

The science section covers four fields. Argon rewrote an internal AlphaQubit error correction decoder as a single optimized hardware kernel and produced a formally verified quantum arcsin circuit. In fusion, a Google team and Professor Matt Landreman of the University of Maryland found new exact solutions that give stellarator simulation codes a reliable benchmark. Gamow Labs used the model to find a chromosomal rearrangement in a newborn’s genome that routine software had missed for years. Google scientists also used it to design five new serine esterase enzymes in hours, and lab tests confirmed they work.

Robotics: Teaching a Humanoid Robot to Play Cornhole

Argon scored 87% on the ERQA benchmark for physical and spatial reasoning, which Google says is ahead of every other frontier model. Google calls its robotics pipeline Video-to-Sim-to-Real and says it compresses weeks of simulation modeling into a few hours. The video walks through the internal test in four steps, narrated in the first person as the robot. First it watched a person throw a bean bag and mapped the arm swing and release onto the robot’s own joints and hands. Second it used the robot’s two eye cameras to measure the 3D position of the board and built a physics simulation of its body and the board in MuJoCo. Third it practiced in simulation, widening the stance and tuning the swing so the hand cleared the hip. Fourth, on the real Apollo humanoid from Apptronik, it watched each bag land and adjusted the release until seven went in. A researcher in the video says the model was shown only the human video and then left to iterate, make mistakes and review footage of what went wrong.

Safety, Alignment and Prompt Injection

Google describes this as its most comprehensive safety stack and says third parties helped measure certain capabilities and safeguards. Under its Frontier Safety Framework it deployed layered defenses against jailbreaks, and violations on an adversarial test set fell nearly 2-fold. For agents, it combined training (respect boundaries, report honestly, pause before irreversible steps) with a real-time misalignment monitor used in deployment, evaluation and training runs. Prompt injection attack success fell to 0.7% across 15 adaptive attempts.

Notable Quotes

“Argon in Antigravity has helped teams increase agentic code submissions by more than 35% while reducing code rollbacks by nearly a third.”
Koray Kavukcuoglu, on how Google engineers have used the model in recent weeks

“Argon answers questions more directly with less hedging, doesn’t repeat itself more than necessary, and skips clichés throughout the responses.”
Koray Kavukcuoglu, on the change in writing style from earlier Gemini models

“Patents are very rich with new chemical insights. They can be 500 pages long. The chemical structure can be 50 pages away from the experimental readout.”
Isomorphic Labs researcher, in the Google DeepMind video on chemistry patents

“In one case, Argon pinpointed a chromosomal rearrangement that routine software missed for years, and explained why it went undetected.”
Koray Kavukcuoglu, on the rare disease research at Gamow Labs

“And fourth, on the real robot, I watched each bag land and adjusted my release by a few cm until we sank seven shots straight in the hole.”
Narration of the model’s own account, in the Google DeepMind cornhole video

“First, we trained Argon to respect system boundaries, report negative findings honestly, and pause for human confirmation before taking irreversible steps.”
Koray Kavukcuoglu, on keeping agents from taking unauthorized shortcuts

The original post includes more video demos, customer statements and links to the model card and developer documentation. Read the full piece on the Google blog here.