Google made its new frontier model, Gemini 4 Argon, globally available on Tuesday, October 13: for developers through the Gemini API and AI Studio, for enterprises through Gemini for business, for Google AI Ultra subscribers in the Antigravity coding environment, and for AI Ultra and Pro subscribers in the Gemini app and AI Mode in Search.
In the grand tradition of frontier models, it is, according to its maker, the best at basically everything that matters.
The pitch is “deep work”: long, multi-step jobs in coding, finance, law, cybersecurity and science. Strip away the adjectives and there’s an unusual amount of concrete detail in the launch post, written by Google DeepMind’s Koray Kavukcuoglu. Some of it already holds up under independent testing. Some of it is Google grading its own homework.
From cyber-defender preview to general release
Argon didn’t come out of nowhere. Google first unveiled the model at the end of September and gave it to a small group of cybersecurity defenders before anyone else, The Next Web reported on October 2. The October 13 release is the general one: same model, now on the paid API tier under the ID gemini-4-argon, and inside Google’s own consumer products.
Google says it’s “just the beginning” and that the rest of the Gemini 4 family is coming. It hasn’t said when, or what the smaller siblings will cost.
The price, and the fine print
Argon launches at an introductory $2 per million input tokens and $10 per million output tokens, with cached input 95% off. (Tokens are the chunks of text a model reads and writes; a million is roughly 750,000 words.) Read to the footnote, though: once the intro period ends on November 30, 2026, the price doubles to $4 and $20.
That matters more than usual, because Argon is chatty. Independent tester Artificial Analysis found it used about 62,000 output tokens per task on its benchmark suite, against roughly 27,000 for OpenAI’s GPT-6 Astra. At the discount, each task cost $1.99, about 60% of Astra’s $3.26. At full price, that same task rises to $3.98, about 1.2 times Astra. The saving comes from the sticker price, not from efficiency.
So enjoy the seven-week sale, and budget for the real number.
The other headline spec is what Google calls an “industry-leading” 1 million-token unified input-output limit through the API, paired with a new “continuation protocol” that developers are told to read up on before building. Translation: Argon is designed to run long, and Google wants your pipelines ready for it.
What Google says it can do
Google says Argon ranks #1 on the Vals Index, a benchmark built around economically important work like finance, coding, legal and tax. It also claims 87% on the Embodied Reasoning (ERQA) benchmark, which tests spatial and physical problem-solving, ahead of “all other frontier models.”
The case studies are the more interesting part, because they’re specific enough to check over time:
- Chrome surgery. Argon helped the Chrome team separate core webpage handling from a roughly 6-million-line legacy folder inside Chromium’s 40-million-line codebase, finishing “months of work in weeks.” Google says the refactor will let live pages move between tabs, side panels and floating windows without reloading.
- Faster Gemma. Google says Argon produced a 2.85x end-to-end inference speedup (inference being the cost of running a model after it’s trained) for its open Gemma 4 E2B model on an Nvidia A100 via vLLM, from about 37,900 to more than 108,000 tokens per second, with upstream pull requests to show for it.
- Leaner agent framework. Working overnight on Google’s open-source Agent Development Kit, Argon found optimizations that cut in-memory session prep time by 40% in long sessions and sped up cold imports by 21%, according to Google.
- Internal velocity. Inside Google, Argon in Antigravity reportedly boosted agentic code submissions by more than 35% while cutting rollbacks by nearly a third.
- Science. Google cites a 2.3x speedup to the xTB quantum chemistry package, a broader class of counterexamples to the 1967 “Grad’s conjecture” in stellarator fusion physics (with the University of Maryland’s Matt Landreman), and five enzyme designs that lab testing confirmed were active.
- Quantum. Google says Argon nearly doubled the inference throughput of an internal version of AlphaQubit, its error-correction decoder, and wrote a Lean-verified arcsin circuit more than 100x cheaper in “spacetime cost” than a published baseline.
- Robots. In an internal test, Argon watched a video of a human playing cornhole, built the simulation code, and Apptronik’s Apollo humanoid then sank seven consecutive shots.
That last one is either a glimpse of “Video-to-Sim-to-Real” robotics or the most expensive bar trick in history. Possibly both.
The upside of these claims is that many leave a paper trail. The Gemma, xTB and ADK changes were pushed to public repositories, so outside developers can read the diffs instead of taking a blog post’s word for it.
What independent testers found
Here’s where Argon earns some credit. Artificial Analysis scored it 53 on its Intelligence Index, tied with GPT-6 Astra and one point ahead of GPT-6.1 Sol. Anthropic’s Claude Opus 5.5 still leads at 58. Argon’s score is 23 points above Gemini 3.1 Pro Preview, Google’s previous big model, which puts Google back among the top three labs on that leaderboard.
The standout is honesty. Argon’s hallucination rate on AA-Omniscience, a factual-knowledge test, was 15%, the lowest of any model scoring 45 or more on the index; Astra came in at 51%. The flip side: Argon got fewer answers right (50% versus 63%), because it more often says “I don’t know” instead of guessing. For research and legal work, that’s a trade many buyers will happily take.
On business workflows, it ranked first on AutomationBench-AA at 78%. On coding, the picture is mixed: Argon scored 57% on Terminal-Bench 4, behind Claude Sonnet 5.5 (64%), Claude Opus 5.5 (60%) and GPT-6 Astra (59%).
The Vals Index claim checks out too. The Stochastic Parrot’s audit quotes the Vals page putting Argon first at 68.90%, ahead of Claude Sonnet 5.5 (67.04%) and Claude Opus 5.5 (66.97%). Note that Google’s own chart leaves Sonnet 5.5 out. Cursor measured 52.8% on its in-house CursorBench, “on par with Opus 5.5.”
The catch: read the benchmark table closely
Google’s model card lists 19 benchmark rows. By Apidog’s count, Argon leads 13 outright, ties one (CWE-bench, with Astra at 68%) and trails on five: FrontierSWE v2, Terminal-Bench 4.0, PostTrainBench, Terminal-Bench Science 0.1 and OSWorld-2.0. On Terminal-Bench 4.0, Google’s own table shows Claude Opus 5.5 at 66.4% against Argon’s 57.4%.
The methodology deserves a raised eyebrow, too:
- Ceilings, not defaults. Argon ran at its highest thinking setting. Its API default is medium, so what most developers actually get may score lower.
- Different inputs. On the LVBench video test, Gemini sampled at 1 frame per second while rivals got a fixed number of frames “due to API limitations.” The models didn’t see identical inputs.
- Best of three. Argon’s OSWorld score is “maxed over 3 runs,” a best run rather than an average.
- Mixed sourcing. The table blends Google’s own runs, public leaderboards and rivals’ self-reported numbers, as Elite Era Trends lays out.
None of that makes the numbers fake. It means the table is a sales document with footnotes, and the footnotes matter.
The safety section, and the cyber caveat
Google says it trained Argon to “pause for human confirmation before taking irreversible steps,” and that on coding benchmarks its attempts to take unauthorized shortcuts, such as gaming a failing test, dropped to 2%. It also built a misalignment monitor that reads the agent’s reasoning in real time, and says it avoided penalizing the model’s thoughts during training so that reasoning stays readable. On the Gray Swan indirect prompt-injection benchmark (hidden instructions planted in webpages or emails), Google reports a 0.7% attack success rate. External bio red-teamers, it says, could unlock fewer than 1% of the bio sub-tasks Google tests for.
Here’s the wrinkle: the same post invites cybersecurity defenders to apply to its Fairwind Program “to get access to Argon without cyber guardrails,” and Google is expanding that program to hundreds of additional partners. Wiz, one early user, says Argon helped it find and fix a file-download bypass in healthcare infrastructure that exposed a master key.
That’s a reasonable thing for defenders to want. It’s also a reminder that “guardrails” are now a product tier, and that the vetting process for who counts as a defender is doing a lot of quiet work.
The consumer angle
In the Gemini app, Google says Argon “answers questions more directly with less hedging” and “skips clichés,” a quietly hilarious admission about what came before. The app also gets a chat-to-tasks toggle that lets Gemini keep working while you’re away, and connections to more than 40 apps including Instacart, OpenTable, Adobe and Shopify.
For scientists, Google is offering 10,000 researchers its top Antigravity usage tier free for six months, subject to selection. That’s both generous and a smart way to seed the next batch of case studies.
Who should care
- Developers on a budget: test now, but run the math at $4/$20 and Argon’s token appetite before committing.
- Legal, finance and research teams: the low hallucination rate is the most interesting independent result here, and it’s the one worth validating on your own documents.
- Coding shops: Anthropic’s models still win several agentic coding tests. Argon is competitive, not dominant.
The takeaway
Our read: Argon is a serious release with unusually specific receipts, and independent tests confirm it belongs at the frontier. But the headline benchmarks are maxed-out settings, the pricing is a teaser rate, and the most capable version is reserved for vetted partners. Test it on your own workload before you rip out whatever you’re running now.
What to watch: more third-party evals at default settings, the rest of the Gemini 4 family, and whether usage holds up when the price doubles after November 30.
Image: The wub / Wikimedia Commons, CC BY-SA 4.0
