Tech • AI • Robotics • Game
Google's New ARGON Is Finally Here... And It's Not What We Expected
9.4/10TL;DR
Google has launched Gemini 4 Argon globally, positioning it as a cheaper, more autonomous flagship AI model with strong business and engineering performance, though independent rankings still place it behind Anthropic’s Claude 5.5.
KEY POINTS
Global launch and competitive position
Sundar Pichai said Gemini 4 Argon is rolling out worldwide as Google’s new top model after an earlier limited release focused on cybersecurity users. The launch comes amid intense competition, with recent releases from Anthropic, OpenAI, and Meta increasing pressure on Google after delays around Gemini 3.5 Pro.
Benchmark gains, but not the overall leader
On the independent Artificial Analysis scoreboard, Argon scored 53, tying GPT-6 Astra, beating Meta’s Muse Wonder by one point, and trailing Claude Opus 5.5 and Claude Sonnet 5.5 by about five points. That still marks a major jump from Google’s previous best model at roughly 30, a 23-point gain in one generation. Google’s own benchmark summary claimed wins in 14 of 18 tests, but the reported totals appeared internally inconsistent.
Strength in office and workflow automation
Argon performed especially well on business-oriented evaluations, ranking first on tests covering finance, legal and tax work, and outperforming rivals by about nine points on a Zapier benchmark for completing end-to-end business tasks. Google also expanded app integrations to more than 40 services, including Instacart, OpenTable, Adobe, and Shopify, while adding a task mode that lets the model keep working across apps in the background.
Mixed coding results
In software engineering, Argon set a record of about 78% on Google’s long-project coding benchmark, but lagged behind three rivals on a command-line test, where Claude led. External observers also noted mismatches between Google’s published rival scores and the figures reported by those rivals themselves, underscoring growing skepticism about benchmark comparisons.
Lower pricing as a major selling point
For developers, Argon is priced at about half the cost of Claude Opus 5.5 and about one-fifth the cost of GPT-6 Astra during its introductory period. Artificial Analysis estimated its tests cost about $2 per task on Argon versus about $6 on Opus. Google offers consumer access through Google AI Pro at $20 a month, an Ultra plan at $100, and a top-tier plan cut from $250 to $200.
Used inside Google’s own engineering
Google said teams using Argon are shipping more than 35% more code, while code changes that later had to be rolled back fell by nearly one-third. The Chrome team used it to help untangle around 6 million lines of legacy code, accelerating work that had taken months and enabling browser features such as moving pages between tabs, side panels and floating windows without reloading.
Autonomous overnight coding test
In one internal trial, Argon was given a short instruction to improve software speed without breaking anything, then worked alone for 11 hours through about 205,000 lines of code. It returned six ideas, of which engineers accepted two. Those two changes improved long-session setup times by about 40% and startup speed by 21%, suggesting useful autonomy but with continued human review.
Science and robotics demonstrations
Google said Argon helped rewrite AlphaQubit, a quantum error-correction system, making it run nearly twice as fast, and redesigned a simulation component that used more than 100 times fewer resources than the best published alternative. In robotics, it learned a cornhole-throwing motion from a human recording and enabled an Apollo humanoid robot to make seven shots in a row. In another simulation-heavy test, it lifted a robot’s success rate from 56% to 75% after more than 4,500 virtual practice rounds.
Medical, fusion and enzyme research
Google also highlighted research uses across medicine and physics. In one newborn genetics case, Argon identified a chromosomal rearrangement that earlier software had missed. Working with Matt Landreman of the University of Maryland, Google said it found a new family of solutions related to stellarator fusion design. In enzyme design, Argon produced five new candidates within hours, and all five worked in lab testing.
Cybersecurity focus and safety controls
Argon was first deployed in cybersecurity and is now being opened to hundreds more trusted defenders. Google said it helped identify a loophole that exposed a master key in a healthcare system and assisted a federal agency in closing a data exposure risk. The company added that the model refuses help with cyberattacks or biological weapons, that biology misuse attempts succeeded less than 1% of the time in outside testing, and that hidden-instruction hijacking worked only 0.7% of the time versus 1% for Claude Opus 5.5 and 8.5% for GPT-6 Astra.
Real-world tests show uneven performance
In one developer comparison on a live app feature, Claude Opus 5.5 scored 77 out of 100 while Argon scored 59 when both plans were graded by another model. In a separate set of 12 difficult coding and security tasks, Argon tied top rivals on 11, was the fastest at about 2.5 minutes per task, and cost roughly $3 for the run versus about $10 for Claude Sonnet 5.5. Its clearest weakness came on a long counting puzzle, where token limits caused blank or incorrect answers.
CONCLUSION
Gemini 4 Argon does not appear to be the smartest model overall, but it narrows the gap sharply while competing on price, speed and workflow automation. Its real significance may lie less in headline benchmarks than in how aggressively Google is deploying it across engineering, science, business software and security.
Ask a question