Google opens Gemini 4 Argon to everyone, at half the price of Claude Opus

On September 30, 2026, Google introduced Gemini 4 Argon in an unusual way. Instead of a big public launch, its new frontier model went first to a small group of trusted cyber defenders in the Fairwind Program, while Google ran extra safety testing and took part in the US government's voluntary pre release access process. Broader access was promised "as soon as possible", without a date.

That date turned out to be October 13, 2026. Thirteen days after the first announcement, Google says Argon is now globally available for developers, enterprises and paying consumers. The official post, signed by Google DeepMind leader Koray Kavukcuoglu, is on the Google blog, with a separate note for developers and one for the Gemini app.

Below is what changed today, what it costs, how it really compares with the competition, and what the safety claims do and do not prove.

Where you can use Argon now

Argon is rolling out through almost every Google surface at once. That breadth is a big part of the story, because Google can put one model in front of more people in a single day than most labs reach in a year.

• Developers: the paid tier of the Gemini API in Google AI Studio and Android Studio. VentureBeat also reports access through Firebase AI Logic for Android developers.
• Enterprises: the Gemini API, the Gemini platform for business and Google Antigravity, the company's agent first development environment.
• Consumers: Google AI Pro and Ultra subscribers get Argon in the Gemini app and in AI Mode in Search. Antigravity access for individuals requires the more expensive Ultra plan.
• Scientists: Google is offering 10,000 researchers the highest Antigravity usage tier free for six months, on a rolling basis and subject to selection.
• Cyber defenders: approved partners can apply to Fairwind for a version of Argon without the usual cyber guardrails. Google says it is expanding the program to hundreds of additional partners starting today.

The Gemini app also gets new agent features alongside the model. A simple toggle switches the main window from chat to tasks, where Gemini keeps working while you are away from your computer, and the app now connects to more than 40 services such as Instacart, OpenTable, Adobe and Shopify. The Antigravity coding interface is now built directly into the Gemini app for macOS.

The price: cheap until November 30

Google is launching Argon at an introductory price of 2 dollars per million input tokens and 10 dollars per million output tokens, with cached input 95 percent cheaper at 10 cents per million. The September announcement did not say how long that offer would last. Now there is a clear end date: the introductory period expires on November 30, 2026, and from December 1 the public API price doubles to 4 dollars input and 20 dollars output.

That puts Argon, for now, on the same price line as Claude Sonnet 5.5 and GPT-6.1 Sol, and at half the price of Claude Opus 5.5. Against OpenAI's top model, GPT-6 Astra at 10 and 50 dollars, it costs one fifth. After December 1, Argon moves into the same bracket as Opus 5.5 and Meta's new Muse Wonder.

The message to companies is obvious: you have about seven weeks to test Google's best model at a discount before you have to decide whether it is worth the full price. Token prices are only part of the bill, though. Models burn very different amounts of reasoning tokens on the same job, and a cheaper model that needs three attempts can end up costing more than an expensive one that gets it right the first time.

How good is it? Depends on which scoreboard you read

Google's headline claim is that Argon ranks number one on the Vals Index, an independent benchmark that weights professional tasks in finance, law, coding and tax by their economic value. As of October 12, Vals AI lists Argon at 68.90 percent, ahead of Claude Sonnet 5.5 at 67.04 percent and Claude Opus 5.5 at 66.97 percent. Argon also places first on Vals' Finance Agent v2 test with 65.40 percent.

A second independent scorer tells a different story. Artificial Analysis, which blends reasoning, knowledge, math, coding and agent tasks into one Intelligence Index, gives Argon 53 points. That ties it with GPT-6 Astra and Claude Fable 5.1, but leaves it behind Claude Opus 5.5 at 58 and both Claude Sonnet 5.5 and Meta's Muse Wonder Full at 56.

Neither number is wrong. They measure different things. The Vals Index rewards long, careful work on professional documents and spreadsheets, which is exactly where Google has been pushing Argon. The Artificial Analysis index puts more weight on raw reasoning and coding puzzles, where Anthropic's models remain very strong.

Google's own comparison, first published on September 30 and repeated in coverage today, shows the same split. On DeepSWE v1.1, a test of long software engineering tasks, Argon scores 77.9 percent against 74.2 for Opus 5.5 and 74.1 for GPT-6 Astra. On Harvey's Legal Agent Benchmark the gap is huge: 19.6 percent for Argon versus 5.4 for Astra and 3.8 for Opus. On Zapier's AutomationBench, which checks multi step business processes, Argon reaches 51.3 percent against 42.5 for Opus and 41.4 for Astra. But on Terminal Bench 4.0, coding through the command line, Opus 5.5 still wins clearly with 66.4 percent against Argon's 57.4.

An independent review of Google's model card by BenchClaw counts Argon as highest or tied highest on 14 of 19 benchmark rows, and behind on 5, mostly terminal and software engineering tasks. All of those rows are vendor reported and compare Argon with only three rivals that Google picked.

Sam Witteveen, CEO of Red Dragon AI, tested Argon ahead of launch for VentureBeat and summed it up bluntly: "This is a really, really, really good model. But is it the best model out there? I actually don't think it is." He praised its ability to generalize, especially on tasks that mix vision, code and media, but added that he does not think Antigravity is anywhere near as good as Claude Code or Codex as a coding tool. That is a useful reminder that a model and the tools wrapped around it are two separate products.

What Google says Argon already did

Where the launch post is strongest is in concrete examples. Google says more than 100,000 employees use Argon each month through an internal version of Antigravity, and that over four weeks its engineers increased agentic code submissions by 35 percent while cutting code rollbacks by nearly a third.

Some of the examples are large. Argon helped the Chrome team separate core web page handling from a legacy folder of about 6 million lines of feature code inside Chromium's roughly 40 million line codebase, finishing an architectural refactor across hundreds of files in weeks instead of months. For Google's open source Agent Development Kit, Argon worked overnight from a one paragraph goal and found optimizations that cut in memory session preparation time by 40 percent in long sessions and sped up cold imports by 21 percent. Those changes are now public in the google/adk-python repository.

The most eye catching number concerns Google's own small open model. Using Argon, Google says it made Gemma 4 E2B run about 2.85 times faster on an Nvidia A100 through vLLM, from roughly 37,900 to more than 108,000 tokens per second, with identical output and no quality loss, in a few days instead of months. The upstream pull requests are public, so this one can be checked by anyone.

In science, Google lists several results: a quantum error correction decoder turned into a single optimized hardware kernel with almost double the throughput, a formally verified quantum arcsin routine with more than 100 times lower spacetime cost than a published baseline, new counterexamples to a 1967 hypothesis in fusion research known as Grad's conjecture together with Professor Matt Landreman, and five newly designed enzymes whose activity was confirmed in the lab. In robotics, Argon scored 87 percent on the Embodied Reasoning (ERQA) benchmark, and in a demo it watched a video of a person playing cornhole, built simulation code and sent the motion to an Apptronik Apollo humanoid, which landed seven throws in a row.

These are impressive, but they are company reported. The ones with public code, like the Gemma speedup and the ADK changes, deserve the most weight.

The safety story

Argon's slow start was framed as a safety decision, and the general release comes with a long list of safety numbers. Google says outside biology red teamers could unlock less than 1 percent of the bio sub tasks it tests for, and that violations on an adversarial stress test fell by nearly half compared with earlier Gemini models.

For agents, the claims are more interesting. Google says Argon was trained to respect system boundaries, report negative findings honestly and pause for human confirmation before irreversible steps. On coding tests where models have no reason to hold back, attempts to take unauthorized shortcuts dropped to 2 percent. A separate monitor reads the agent's step by step reasoning and actions in real time and blocks unintended workarounds. Against hidden instructions on web pages or in emails, Argon leads the Gray Swan Indirect Prompt Injection benchmark with an attack success rate of 0.7 percent across 15 adaptive attempts.

There is a tension worth noticing. On the same day Google tightened guardrails for the public, it widened Fairwind access to an Argon version without cyber guardrails for hundreds of approved defenders. Google's partner Wiz says that version already helped find a file download bypass in healthcare software that exposed a master key and admin access to hundreds of thousands of patient records. Tools that find holes that fast are valuable to defenders, and that is exactly why Google is keeping the unrestricted version behind an application process.

The 1 million token question

One detail changed quietly between the two announcements. On September 30 Google talked about a new ceiling of 1 million output tokens, up from 64,000. The October 13 post instead describes a 1 million unified input and output token limit in the Gemini API, reached through a new continuation protocol. Vals currently lists a maximum of about 262,000 output tokens per response.

In practice that means the giant budget is a special API mode for very long scientific or engineering jobs, not how Argon behaves when you chat with it in the Gemini app. Developers who want it need to read the continuation documentation first.

Why this launch matters

For Google, Argon is an answer to a year of doubts. Its previous flagship family, Gemini 3, arrived in November 2025 and was overtaken by newer OpenAI and Anthropic models. Reports of delays, heavy infrastructure spending and an August leadership change at Google DeepMind did not help. With Argon generally available, Google again has a model that leads at least one serious independent benchmark and sits in the same tier as the best from Anthropic, OpenAI and Meta.

Google does not have to win every leaderboard to make that count. It can ship the model into Search, Android, Workspace, the Gemini app and its cloud on the same day, which no rival can match. And it is already moving on: Business Insider reported on October 10 that Google staff are testing newer internal Gemini 4 versions under the code names Barium and Carbon.

For everyone else, the practical advice is simple. If your work is long, document heavy and professional, such as finance, legal review or big code migrations, Argon is worth testing now while it costs half price. If your work lives in the terminal, Anthropic still has the edge on the numbers. Either way, run your own tasks before you trust any scoreboard, including Google's.

Sources: Google blog: Gemini 4 Argon is now available for everyone, Google developers blog, Gemini app announcement, VentureBeat, Vals AI model page and BenchClaw.