Google published Gemini 4 Argon on September 30, and the row to read is the legal one. On Harvey’s Legal Agent Benchmark, Google’s table puts Argon at 19.6%. The rival models it lists score between 3.8% and 6.7%. That is nearly three times the best of them and more than five times the worst.
Then comes the catch. Argon is not generally available. For now it is rolling out to vetted cyber defenders through Google’s Fairwind Program, and the launch post says paid API customers and Google AI Ultra subscribers follow, with no date.
The table, row by row
Google lists the numbers as its own comparison, so treat them as published claims rather than audited results. On Harvey’s benchmark Argon scores 19.6%. VentureBeat’s read of the table has GPT-6 Astra at 5.4% and Claude Opus 5.5 at 3.8%, and the best rival at 6.7%.
AutomationBench, the business-process test, goes 51.3% against rivals between 31.4% and 42.5%. DeepSWE v1.1, long-horizon coding, goes 77.9% against 67.4% to 74.2%. On CWE-bench v1, which measures fixing security vulnerabilities, Argon scores 68% and ties for first. Google also claims state-of-the-art on the Vals Finance Agent v2; we could not open a figure for that one.
The number that matters is the gap, not the score. A leader at 19.6% on a 100-point scale is still a long way from done, and Google’s table does not say how a single task is scored, so read the figure as a ranking, not a share of a lawyer’s work. What changed on September 30 is the distance between first place and second.
Where Argon loses
The table is not a clean sweep. Per TechMyMoney’s summary, Argon scores 57.4% on Terminal-Bench 4.0 against 66.4% for Opus 5.5, and 55.0% on FrontierSWE v2 against 65.5% for Astra. Those are gaps of 9 and 10.5 points, in the other direction.
Third-party testing adds a wrinkle. Artificial Analysis gave Argon the lowest hallucination rate among leading models, 15%, Engadget reports. TechMyMoney’s summary adds that its accuracy was 50% against Astra’s 63%. Read together, the model says “I don’t know” more often than it guesses. In legal and finance work that is a feature if a person is checking the output and a cost if nobody is.
What it costs
Argon’s introductory price is $2 per million input tokens and $10 per million output, with cached input discounted 95%. After the introductory period it moves to $4 and $20, which is Opus 5.5’s price. VentureBeat lists Astra at $10 and $50. So for now the output rate is half of Opus and a fifth of Astra.
Hold that beside Sonnet 5.5, which Anthropic shipped on September 28 at the same $2 and $10 with a GDPval-AA score two points off Opus. Within two days, two labs have priced a model that claims a flagship-grade result at mid-tier rates. The introductory label on Google’s number is the part to watch. Nothing published says how long it lasts.
Against what we have already covered
On September 9, Harvey disclosed that 80% of the Am Law 100 already uses its product, and we flagged the Legal Agent Benchmark it launched the same day as the more important move: a score is what turns “how good was this work” into a line in a procurement file. Argon puts that benchmark in Google’s own comparison table, which makes the legal row part of a frontier launch rather than a vendor’s side project.
Two weeks ago Gemini 3.8 Live finished 35.1% of Sierra’s banking calls while topping a speech index. The pattern repeats: a vendor wins a general leaderboard, and the domain-specific row tells you what the job still needs. And when OpenAI built a banker’s workstation around the junior analyst on September 10, the product scope was the first-year workday. Argon’s pitch lists legal, finance and business-process work in one sentence.
Who is exposed
Four groups, in the order Google’s own table names them.
Legal research and first-draft work, the junior associate and paralegal layer, is where the Harvey score points. Business-process coordinators and operations back-office staff are what AutomationBench is built to stand in for. Financial analysts doing repeatable agent-style work sit behind the Vals Finance Agent claim. Security engineers triaging and patching known vulnerability classes are the CWE-bench row, and that is also the only group Google has actually handed the model to.
The timing is not immediate. A model that vetted defenders can use and paying customers cannot has no effect on a requisition list yet. But the price and the table are public now, and next quarter’s tooling decisions will be made against them.
What to watch
Three things. When the paid API opens, and whether $2 and $10 survives it. Whether Harvey publishes its own leaderboard entry, since a vendor’s chosen table and the benchmark owner’s table are different documents. And whether the 19.6% moves when a firm runs the model on its own matter files rather than a test set.
