Google spent months being politely described as "not out of the race." Now it has released Gemini 4 Argon, a model it says competes with the best the industry has, and it is charging roughly a fifth of what the leaders charge. Whether that is confidence or a sale is a question the pricing page answers more honestly than the press release does.
Argon is Google's first major frontier release since Gemini 3 in late 2025. In the gap, Anthropic and OpenAI shipped models that left Google looking like the incumbent who forgot to turn up. Internal shakeups did not help. Argon is the reply, and it arrived a day after Meta announced its own frontier model, Muse Wonder Full, which makes the frontier look less like a two-horse race and more like a crowded paddock.
What Google is actually claiming
Google says Argon offers "frontier reasoning and improved performance across complex workflows." It also says the model hedges less, repeats itself less, and avoids the clichés that make so many chatbot answers read like a motivational poster.
The company claims Argon beats GPT-6 Astra, Claude Fable 5.1 and Claude Opus 5.5 on benchmarks covering coding, cybersecurity, science, finance and law. The one named ranking comes from Google DeepMind's own announcement: Argon is #1 on the Vals Index, which measures model performance across economic sectors such as finance, coding, legal and tax work.
A few other leaderboard claims appear later in the announcement, and they are covered below. All of them are Google's numbers, and the announcement does not include a full results table.
DeepMind also pitches Argon as a model for "deep work." That means long, multi-step tasks where it keeps going rather than producing a quick answer and a shrug. It supports a 1M unified input-output token limit through the Gemini API, and Google tells developers to read the documentation on a new continuation protocol to set things up properly.
Where you can use it
Google says the rollout is global and begins today. Access depends heavily on who you are and what you pay.
| Audience | Access |
|---|---|
| Developers | Gemini API paid tier, AI Studio, Android Studio |
| Google AI Ultra subscribers | Google Antigravity, the Gemini app, AI Mode in Search |
| Google AI Pro subscribers | The Gemini app and AI Mode in Search |
| Enterprises | Gemini for business, the API, Google Antigravity |
| Scientists and researchers | Apply for free top-tier Antigravity access |
| Cybersecurity defenders | Apply to the Fairwind Program |
One gap stands out. One report on the launch notes that it is unclear when, or whether, Argon will reach Plus subscribers or free users. Google's own announcement lists Ultra and Pro and says nothing more, which is a kind of answer.
The price is the loudest feature
Argon launches at an introductory $2 per million input tokens and $10 per million output tokens, with cached input tokens at 95 percent off. That is around a fifth of the cost of OpenAI's GPT-6 Astra and Anthropic's Fable 5 models.
The word "introductory" matters. According to one report, those prices double at the end of November when the introductory period finishes. Even then, Argon would still be well below its frontier rivals per token.
There is a catch, and it is the usual one. Per-token price is not per-task price. Models can use different numbers of tokens for the same job, so a cheap token can still produce an expensive afternoon. If Argon is wordy, or reasons for a long time before answering, the savings shrink. Nobody has published the numbers that settle it.
The engineering showcase
Google's strongest material is a list of things its own teams did with Argon, most of them involving Antigravity, its agentic development platform. Google says engineers have been shipping code faster over the past few weeks. Using Argon in Antigravity, teams increased agentic code submissions by more than 35 percent and cut code rollbacks by nearly a third.
The specific examples are more interesting than the aggregate.
- Chrome refactor. Argon helped separate core webpage handling from a roughly 6 million line legacy folder of browser features inside Chromium's 40 million line codebase. It produced multiple clean code changes at a time and finished an architectural refactor across hundreds of files. Google says that took weeks instead of months. The payoff, it says, is live webpages that can move between tabs, side panels and floating windows without reloading, plus the removal of a whole category of browser crashes.
- Agent Development Kit. The team behind Google's open-source ADK gave Argon its .agents/skills rulebook and a one-paragraph goal: cut framework overhead without breaking existing apps. Working on its own overnight, Argon found optimizations now live in google/adk-python. They cut in-memory session prep time by 40 percent in long sessions and sped up cold imports by 21 percent.
- xTB. This is one of the most-cited open-source quantum chemistry packages, used to predict molecular shapes and screen drug and materials candidates. Google's researchers used Argon to make it up to 2.3x faster at predicting molecular structures and 1.8x faster at conformer search. The changes went to the official xTB repository on GitHub.
- Gemma inference. Google reports a 2.85x end-to-end inference speedup for its open Gemma models on an A100 via vLLM. Throughput on Gemma 4 E2B went from about 37,900 to over 108,000 tokens per second, with 100 percent token parity and no quality drop. It says this took a few days, work that would usually take months, and points to upstream pull requests covering everything from attention tuning to speculative decoding.
A separate report adds a few more internal uses: finding and patching cybersecurity vulnerabilities, halving the computing power needed for YouTube's comments section, and debugging the drone physics simulator for Wing, Alphabet's drone delivery company.
The pattern is consistent. Argon is shown being handed a vague goal, a rulebook and a night alone, then returning with measured improvements to real repositories. The measurements are Google's own, but the code is public in several cases, so they are checkable.
Three-dimensional show-and-tell
Google also claims big improvements in spatial reasoning, mathematical precision and visual coding compared with earlier Gemini models. The demos lean on Antigravity's /goal slash command and multi-agent loops with custom skills.
The list is a little theatrical, but it shows what a long-running agent can string together:
- A detailed Pagoda built through multi-agent visual critique loops, ending as a multi-building temple complex with foliage and lighting in Blender.
- A playable miniature racing game made from pictures of a real desk. Blender MCP handled modeling and collision checks, Nano Banana made textures, and Gemini Omni Flash supplied engine audio in Three.js.
- A complex 3D game with procedural sculpting, rigging, animation and PBR baking, wired for the web with GLSL shaders, shatter physics and a single-player AI.
- A 3D ray-tracing engine written in pure Python with zero external libraries, which then produced a video of a sunlit library, narrated with Gemini 3.8 TTS and scored with Lyria 3.5.
- A finished Gothic 3D scene built from a single Nano Banana image, with Argon driving Blender MCP in an autonomous loop.
Several of these are available to try in AI Studio. As demos go, they have the advantage of being something you can open and judge yourself, which is more than most launch videos allow.
Office work, with fewer hedges
Google says Argon combines frontier reasoning, deep visual perception, broader domain knowledge and better writing across legal, finance, science and coding. It claims the model follows style instructions better than any previous Gemini model and keeps users in the loop as it works.
The business examples are practical:
- Document-heavy synthesis. Argon pulled together about 50 GB of airborne LiDAR scans, GIS room vectors, Library of Congress CAD drawings and National Park Service archives to rebuild 3D models of Chaco Canyon.
- Client proposals. In the Gemini app, Argon takes an unstructured folder of project context and turns it into a proposal deck using a custom skill. It then drafts and emails the deck, and sets up a scheduled task to handle future client requests.
- Medical education. It combined data from three open-source repositories into a 3D tool where users peel back anatomical layers, inspect clinical metadata and trigger biomechanics.
- Product design. From a sketch of measurements and a photo of a couch, Argon, working through Antigravity in the Gemini macOS app, generated design options for new sofa legs with Nano Banana and produced a 3D-printable file for one of them.
- Drug discovery. Isomorphic Labs uses Argon to synthesize complex PDFs and patents, reducing a process that typically takes weeks to hours.
The sofa-leg demo is the most honest one in the set. It is small, concrete, and something a person might actually want.
Cybersecurity, and the gate in front of it
The public rollout follows a limited release to a select set of "trusted cyber defenders," in line with how other recent frontier models have been handled. Google says it is now expanding access to hundreds of additional Fairwind Program partners. Defenders who apply to Fairwind can get Argon without cyber guardrails, on a rolling basis.
The pitch is that the same ability that finds flaws for defenders could help attackers, so the unrestricted version goes to vetted people first. It is a sensible compromise, and also a tidy way of saying "this is dangerous" while selling it.
Google cites Wiz, which is using Argon in its "Scan for Good" initiative that protects critical infrastructure for free. Two cases are described:
- Government data. Working with a federal agency under a long-running cyber resilience program, Argon helped find and fix a gap that risked exposing sensitive data and government infrastructure.
- Healthcare. By chaining complex vulnerabilities, Argon helped the team discover and fix a public file download bypass that exposed a master key and granted administrator access. Google says this protected hundreds of thousands of private patient records.
The details are deliberately thin, which is what you would expect from vulnerability stories. It means they are hard to verify from the outside.
Science claims
This is the section where Google's announcement gets most ambitious.
Quantum computing. Argon translated an internal version of AlphaQubit, DeepMind's machine-learning quantum error correction decoder, into a single optimized hardware kernel. Google says that nearly doubled inference throughput, so larger models can run while still meeting real-time thresholds. Separately, an Argon-written implementation of arcsin, a building block in recent quantum simulation algorithms, was formally verified in Lean. Google says it cuts spacetime cost by more than 100x compared with a published quantum CORDIC baseline.
Nuclear fusion. Since 1967, a hypothesis known as Grad's conjecture assumed exact mathematical solutions for stellarator magnetic fields were impossible, until recent counterexamples appeared, including some by Professor Matt Landreman of the University of Maryland. A Google team working with Landreman says it has found a broader class of counterexamples. These give fusion researchers reliable benchmarks for testing stellarator simulation codes.
Rare disease. Researchers at Gamow Labs used Argon on de-identified genomes from newborns with rare developmental lung disorders. In one case it found a chromosomal rearrangement that routine software had missed for years, and explained why. On Gamow's private evaluations built from real clinical and lab data, Argon topped RareBench at 57 percent for rare-disease diagnosis assistance and LabBench at 52.2 percent for wet-lab reasoning. Note that these are private evaluations, not public ones.
Enzyme design. Google's scientists worked with Argon to design serine esterases, enzymes that cut ester bonds and are used in recycling and in making pharmaceuticals, biofuels and detergents. The target was a benchmark molecule, 4MU-acetate, which glows under UV light when its ester bonds break. In a matter of hours, Argon assembled data, tuned a model and applied biochemistry-inspired filters to reach five novel designs. Lab testing then confirmed catalytic activity across them.
Google is also offering 10,000 scientists the highest usage tier of Antigravity free for six months. Onboarding is rolling, and access is subject to eligibility, capacity and selection. Applications are open.
Robotics: seven cornhole shots
Google says Argon scores 87 percent on the Embodied Reasoning benchmark ERQA, which tests physical and spatial problem solving in 3D environments, and that it beats all other frontier models using pure reasoning.
The demonstration is more fun than the number. In an internal test, Argon watched a raw video of a person playing cornhole, mapped the spatial coordinates, worked out the whole-body joint movements and wrote the simulation code on its own. When that was transferred onto Apptronik's Apollo humanoid, the robot sank seven consecutive shots in a real-world trial.
Google calls this a "Video-to-Sim-to-Real" pipeline and says it compresses weeks of manual simulation modeling into a few hours. One demo is not a benchmark, but it is hard to argue with seven in a row.
Safety claims
Google describes Argon as having its most comprehensive safety and security stack yet, and says third-party partners helped measure certain capabilities and safeguards. The claims, under its Frontier Safety Framework, are:
- Misuse defenses. A multi-layer stack combining jailbreak resistance, monitoring and enforcement. External bio red teaming could unlock less than 1 percent of the bio sub-tasks tested. On an adversarial stress-test set, violations fell nearly 2-fold compared with earlier Gemini models.
- Fewer shortcuts. Google worries that agents get too eager to finish and take unauthorized routes around obstacles like permission errors or failing tests. It trained Argon to respect system boundaries, report negative findings honestly and pause for human confirmation before irreversible steps. On coding benchmarks where models have no incentive to hold back, shortcut attempts dropped to 2 percent, its best result among Gemini models.
- Live monitoring. An automated misalignment monitor reads the agent's step-by-step reasoning and actions and blocks unintended workarounds in real time. Google used it across internal deployments, capability evaluations and reinforcement learning runs. It says it avoided penalizing the model's thoughts during training, so the reasoning stays transparent and monitorable.
- Prompt injection. Through adversarial training and automated red teaming, Argon leads the Gray Swan Indirect Prompt Injection benchmark, cutting attack success across 15 adaptive attempts to 0.7 percent.
A Model Card is available for anyone who wants the longer version.
What to make of it
The pricing is the most verifiable claim in the package. The benchmarks are Google's, the science results come with Google's own collaborators, and the safety numbers come from Google's own testing. That does not make them wrong. It means the real verdict will come from developers running Argon against their own codebases and finding out how many tokens it burns doing so.
Two things are worth watching. The first is what happens when the introductory prices reportedly double at the end of November, and whether a cheaper-per-token model is still cheaper per task. The second is access: Argon is open to paying subscribers and developers, and the free tier is left unmentioned.
Google also says Gemini 4 Argon "is just the beginning" and that the rest of the Gemini 4 family is on the way. After a long quiet stretch, that is a notable thing to promise. Google has now turned up to the race, and the next move is whether it can keep running.