Google unveiled Gemini 4 Argon on September 30, its strongest model to date, with a one million token output limit and a clear focus on cybersecurity work. Most users will have to wait, since the first seats go to vetted cyber defenders and the US government.
Key Takeaways
- Gemini 4 Argon targets software engineering, legal and financial work, and finding software vulnerabilities.
- Launch pricing sits at $2 per million input tokens and $10 per million output tokens, set to double later.
- Independent scores put it level with GPT-6 Astra and still behind Claude Opus 5.5 on the overall index.
Have an AI Sum Up This Article
ChatGPTGemini 4 Argon goes to cyber defenders first
Koray Kavukcuoglu, SVP of Google DeepMind and Google’s Chief AI Architect, signed the official post laying out the model’s capabilities and rollout plan. Google frames Gemini 4 Argon as the start of its next frontier era, after a summer spent shipping one 3.x release after another.
The rollout breaks with habit. Gemini 4 Argon is reaching trusted cyber defenders before anyone else, through Google’s Fairwind program, while US government agencies get pre-release access. Paid API customers and Google AI Ultra subscribers come next, “as soon as possible”, with no date attached.
That order matches the pitch. Google says the model can find, validate and patch critical software vulnerabilities on its own, and it scores 68% on CWE-bench v1, a vulnerability remediation benchmark. Security firm Wiz already runs it inside its Scan for Good initiative, where it caught a critical flaw in healthcare software that earlier frontier models had missed.
The spec sheet moves up a scale as well. Maximum output jumps from 64,000 tokens to one million, matching a one million token input window. Argon reads text, images, video and audio and answers in text only, while Google kept voice for a dedicated model we put through its paces in our Gemini 3.8 Live test. A new Long Decode Continuation feature pauses and resumes very long answers.
Google ran that stamina on its own codebases first. The company points to a migration of more than 800,000 lines of Fuchsia’s Zircon kernel, a video decoder made 2.7 times faster and memory savings across more than 300 TiB in its data centers. Those figures come from Google alone, the only source on its internal use.
On price, Gemini 4 Argon opens low. The introductory rate is $2 per million input tokens and $10 per million output tokens, with a 95% discount on cached input. Standard pricing will then climb to $4 and $20, and Google hasn’t said when.
How Gemini 4 Argon fares on independent scoreboards
Google’s own charts are full of first places. The model posts 77.9% on DeepSWE v1.1, a real-world software engineering test, tops AutomationBench at 51.3% and reaches 91.7% on LVBench for long video understanding. Google also claims the best resistance to prompt injection on Gray Swan’s benchmark.
Third-party numbers tell a more mixed story. The Decoder pulled the outside measurements together in its breakdown of where Argon lands against OpenAI and Anthropic. On the Artificial Analysis Intelligence Index, Gemini 4 Argon and GPT-6 Astra tie at 53 while Claude Opus 5.5 leads at 58.
Two leaderboards do lean Google’s way. Argon takes first place on the Vals Index at 68.9%, which covers finance, coding, law and tax, and it leads Arena Text with 1,525 points, a ranking built on blind user votes.
Errors are where the gap gets wide. On AA-Omniscience, Gemini 4 Argon hallucinates 15% of the time against 51% for GPT-6 Astra. For a team handing document research or legal analysis to a model, that matters more than a point on a composite index, and it follows the reset Google started when Gemini 3.5 Pro was pushed back so the company could rebuild it.
The real bill deserves a closer look. Argon averages 62,000 output tokens per task, against 27,000 for GPT-6 Astra, so a low unit price can still produce a heavier invoice. That makes it a different calculation from Gemini 3.8 Flash, which trailed Claude Opus 5 on coding at a fraction of the cost.
More articles on Horizon
- OpenAI Cancels GPT-6.1 Astra Over Deception
- Claude Sonnet 5.5 Nearly Matches Opus at Lower Cost
- Bernie Sanders Wants to Ban AI Superintelligence
OpenAI and Anthropic now face a Google back in the race
For developers, little changes while the API stays closed. Teams working on large codebases or huge document sets will still have a concrete reason to try Gemini 4 Argon the day it opens, with output nearly sixteen times longer than the previous generation allowed. Until then, the summer lineup remains the only option, from Gemini 3.8 Live and its aggressive voice pricing to the Flash models.
Legal and finance buyers are the first target. Google leans on its Vals Index lead in exactly those fields, a direct shot at OpenAI’s Astra for Law and at the Claude rollouts inside banks. The contest has shifted to how reliable a model is on professional work, and the low hallucination rate is now Google’s main sales argument.
Anthropic keeps the overall index lead with Claude Opus 5.5, five points clear. That edge rests on a single ranking, though, while Google takes the Vals Index and Arena Text, one built around finance and legal work and the other around user preference. The recent Claude Sonnet 5.5 release shows Anthropic is pushing on the mid-tier as well.
OpenAI is in the tightest spot. GPT-6 Astra is matched on the overall index and clearly behind on hallucinations, in the same week the FTC opened a probe into OpenAI, Anthropic and other labs. A price move or an Astra update in the coming weeks would be the obvious reply.
Google also holds a distribution lever its rivals lack. The Gemini app passed one billion monthly users in August, a milestone we saw coming when Gemini closed in on a billion users. Once Argon reaches the Ultra plan, that audience becomes the real test, a model measured at scale rather than in a lab.
Follow the story on Horizon.
