NOTE: This article is a little more benchmark and metric heavy than my normal articles - I’m experimenting with different levels of detail to see what people like.
Back in July, I audited a year of my own AI usage. Google’s Gemini had fallen to about two percent of my daily work while Claude and ChatGPT split the rest.
That number bothered me because Google owns the best collection of AI assets in the industry across Search, Workspace, DeepMind, and Cloud. Yet for most of this year, Google kept shipping fast, cheap Flash updates while its top reasoning tier stood still. Gemini 3.1 Pro entered preview in February and stayed there. Gemini 3.5 Pro never showed up. Then in late August, Google Cloud launched Gemini Enterprise for Legal (to a limited testing group), and its default “auto” router in my Enterprise Subscription failed my simplest procurement test. I asked for a formula-linked financial workbook. It promised to build one, failed three times, and handed me an apology instead of a file.
Today, October 13, Google opened Gemini 4 Argon to everyone after a two-week security test through its Fairwind cyber defense program. Argon is the first model in the Gemini 4 family, and it is the release I have waited for all year. I spent today running heavy builds with it in Google Antigravity and AI Studio and wiring it into my local review setup next to Claude and Codex. If you help run a law firm or legal department, put Google back in your testing rotation.
The output bottleneck is gone
For the past year, Google has pitched a one-million-token input window, plenty of room to drop in a closing binder or ten deposition transcripts at once. The problem sat on the other end.
Output stopped at 65,536 tokens, and that cap had to cover both internal reasoning and the finished document. Ask Gemini to reconcile defined terms across a 400-page credit agreement or check a multi-year damages schedule, and it could burn through its token budget thinking and cut off before finishing the memo.
Argon fixes that. It shares one 1,048,576-token window across input, thinking, and output. Ordinary synchronous API calls still stop at 65,536 output tokens for everyday chat and summaries. When a job demands deep reasoning and a long deliverable, developers using Google’s Interactions API can turn on background execution, which Google calls long decode. The server writes up to 64,000 tokens per hop under one interaction ID until it finishes the job, up to the full million-token limit. You also pick a low, medium, or high thinking budget based on what a mistake would cost.
Inside Google, the Chrome team pointed Argon at a six-million-line legacy folder inside Chromium’s 40-million-line codebase. Argon piped architectural changes across hundreds of files and cut months of refactoring down to weeks.
I saw that same stamina on my Mac today. I used Argon inside Antigravity to build and render a 60-second, 1,800-frame 4K video combining headless Blender 5.2, Three.js, an exploding 30-part Porsche 911 CAD model, 3D Excel and PowerPoint scenes, and a multi-track audio master. Every scene compiled and rendered to 4K without me having to hand-fix broken camera math or timing scripts. Here’s that video:
Where Argon wins on legal and financial work
On the Vals Index, an independent benchmark from Vals AI that weights multi-step work across law, finance, tax, and code by their share of U.S. GDP, Gemini 4 Argon takes first place out of 48 models at 68.90 percent. That beats Claude Sonnet 5.5 at 67.04 percent, Claude Opus 5.5 at 66.97 percent, and Claude Fable 5.1 at 65.83 percent. Argon also finished the index at $15.68 per test. The Anthropic models behind it cost between $21.34 and $32.14.
Look closer at the legal and financial tests and the jump is even sharper. Back in June, Gemini 3.1 Pro scored zero percent on Harvey’s Legal Agent Benchmark. Argon climbed to sixth out of 79 models at 19.58 percent. Across the Vals suite, Argon ranks third out of 150 models on LegalBench at 88.30 percent, fourth on Legal Research Bench at 54.81 percent, fourth on Tax Agent Bench at 44.56 percent, and seventh on Public Benefits Bench at 69.76 percent. On Vals Finance Agent v2, it ranks first out of 78 models at 65.40 percent.
Why does a finance benchmark matter to lawyers? Because the numbers that decide a dispute or hold up a closing rarely sit in clean text files. They hide in scanned PDFs and footnotes buried inside financial exhibits.
In Box’s testing, Argon beat Gemini 3.1 Pro on accuracy by 14 percent by tying together facts across whole document sets, including numbers inside charts. Kepler tested Argon by having it build multi-year financial statements straight from SEC filings. Argon matched analyst-built reference models on more than 99 percent of figures. If your associates spend nights checking disclosure schedules against financial exhibits or tying bank statements to a general ledger, that accuracy changes how much cleanup lands on a partner’s desk.
LexisNexis has been testing Argon inside the multi-model engine behind Lexis+ with Protégé for finding relevant citations, contract review, complaint analysis, and research memos. At MinutaIA, CEO Caio Perona said Argon tied dates and financial math to legal precedent well enough that, on one test, it beat their human expert answer key by spotting a detail the lawyers had missed.
Argon also scored 91.7 percent on LVBench for long-video understanding. Litigators can drop hours of bodycam video or recorded depositions into the same window as the written exhibits and cross-check what a witness said on camera against the documents. Google also tuned Argon’s writing to cut hedging and stick closer to house style.
Better contract terms and a few real catches
If you let an AI agent read opposing counsel’s document production or inbound emails, you have to worry about indirect prompt injection, where someone hides instructions inside an untrusted PDF or email to hijack your agent. On the Gray Swan Indirect Prompt Injection benchmark, Argon held attack success across 15 adaptive attempts to 0.7 percent, the best score in the field.
Google also trained Argon to admit when a task fails instead of faking a clean answer. On coding tests where models often try to bypass a failing check, Argon’s shortcut rate dropped to two percent, backed by an automated monitor that checks its reasoning while it works. That matters in legal work, where a model that quietly invents a missing citation is worse than useless.
There is also good news on the contract issue I wrote about in August. In Google’s developer docs today, Google lists Argon 4 as a stable release, not a preview. Under Google Cloud’s pre-GA terms, preview models come with no service-level agreement, no IP indemnity, a $25,000 liability cap, and a ban on protected health information under a HIPAA business associate agreement. Making Argon stable in the API removes the biggest legal blocker to using it.
Still, watch the price tag before you roll it out firm-wide. Google’s launch rate of $2 per million input tokens and $10 per million output tokens ends on November 30, 2026. On December 1, those prices double to $4 and $20. Run your budget math on the December numbers, not the autumn sale.
Watch the speed, too. With thinking set to high on the Vals Index, Argon took an average of 46 minutes and 33 seconds per test. Forty-six minutes is an eternity if you are staring at a chat window waiting for a paragraph. It is a bargain if the model is grinding through six hours of associate diligence in the background while you are at dinner. On long agent jobs, per-test costs hit $57.82 on code migration and $193.78 on computer-use tasks, where Argon trails at 4.83 percent. Save high thinking for hard, high-value work where an unseen error is expensive.
What I’d do this week
You don’t need to rip out Claude or ChatGPT to get value from Argon. Here are the three moves I’d make now.
Use Argon as a second-opinion reviewer. Today I wired a local bridge script so Claude Code and Codex can pass a draft brief or contract redline to
Argon 4for an adversarial review. Having a third frontier lab attack your primary model’s draft and check its citations catches blind spots that one model alone will miss.Run the file test on one ugly matter packet. Pick a real assignment that ends in a finished file, like a Bates-linked chronology table or a multi-tab damages workbook. Run the same prompt and source packet through Argon three times with thinking set to high. Open the files and check the formulas against your current tools.
Get your routing and contract terms in writing. Before you put client work into any Google business plan, ask your account rep to confirm two things in writing. First, verify that your administrators can pin
Argon 4with high thinking turned on instead of trusting an automatic router. Second, confirm that your commercial indemnity and BAA terms cover Argon as a production service.
First, thank you for reading - I hope you enjoyed this article. I am genuinely excited to see what Argon 4 can do in additional testing. So far I am genuinely impressed with the reasoning and analytic capabilities and I’m just getting started, so stay tuned for more!
If you enjoyed this, please share with others!