Google's Gemini 4 Argon Tops Benchmarks in Limited Release

Google unveiled Gemini 4 Argon, leading 12 of 18 disclosed benchmarks over OpenAI and Anthropic rivals, but broad access and its release date remain unscheduled.

@sharedot

Google's Gemini 4 Argon Tops Benchmarks in Limited Release

By @sharedot · · 8 pages

  • AI
  • Gemini 4 Argon
  • Google
  • Frontier Models
  • Benchmarks

Google unveiled Gemini 4 Argon, leading 12 of 18 disclosed benchmarks over OpenAI and Anthropic rivals, but broad access and its release date remain unscheduled.

What happened: Google re-enters the frontier race

Google announced Gemini 4 Argon on September 30, 2026, its first frontier model in more than seven months following Gemini 3.1 Pro, after a drawn-out development period in which the already-announced Gemini 3.5 was skipped entirely, as The Decoder reports.

Google says Argon leads or ties rival systems on enterprise-relevant benchmarks spanning long-horizon software engineering, cybersecurity vulnerability remediation, business automation and economic-impact knowledge work.

The model supports up to one million output tokens, an industry first up from the prior 64,000-token limit, enabled by a new Long Decode Continuation API feature that pauses and resumes long reasoning chains.

Access is deliberately narrow: rollout begins with trusted cyber defenders in Google's Fairwind Program and US government agencies under the voluntary pre-release process, with wider availability to paid API customers and AI Ultra subscribers promised only 'as soon as possible.'

Why it's surprising: breadth without a clean sweep

The surprise is that Google can claim frontier leadership by benchmark count without winning everything. Across the 18 benchmarks in Google's disclosed comparison, Argon leads outright on 12 and ties on one, while GPT-6 Astra leads outright on three and Claude Opus 5.5 on two, VentureBeat reports.

Argon's biggest wins are in enterprise workflows: 19.6% on Harvey's Legal Agent Benchmark versus 5.4% for GPT-6 Astra and 3.8% for Claude Opus 5.5, 51.3% on Zapier's AutomationBench, and 84.2% on the long-context GraphWalks test.

Yet the race stays close where it counts — GPT-6 Astra beats Argon by 10.5 points on both FrontierSWE v2 (65.5% to 55.0%) and Terminal-Bench Science 0.1 (68.1% to 57.6%), and Claude Opus 5.5 leads Terminal-bench 4.0 by 9 points, 66.4% to 57.4%.

The Decoder notes Anthropic likely still holds the overall lead.

Independent evidence tempers the marketing

Independent evaluations suggest a tighter picture than Google's own table.

According to The Decoder, Artificial Analysis scores Argon at 'High' reasoning 53 on its Intelligence Index, tying GPT-6 Astra (max) and Claude Fable 5.1 but sitting five points behind Claude Opus 5.5 at 58 — a 23-point jump over Gemini 3.1 Pro Preview.

Argon wins AutomationBench-AA at 77.5% and posts a strikingly low 15% hallucination rate on AA-Omniscience, versus 51% for GPT-6 Astra (max), but its 50% accuracy trails Astra's 63%.

Efficiency is the catch: Argon burns an average 62,000 output tokens per task against Astra's 27,000, so its cost edge comes from cheaper rates, not thrift.

On arena.ai's human preference Text Arena, Argon (High) takes first place with 1,525 points, 20 ahead of Claude Opus 4.6 (High), and Arena rates it the most cost-efficient model in its ranking.

Internal wins and defensive cyber claims

Google says thousands of employees already use Argon internally for specialized coding, research and writing, and cites concrete results, VentureBeat reports: Argon agents optimized quantum computing spacetime resources, reportedly beating a published baseline by 40% in minutes; identified memory optimizations expected to free 300 TiB across Google data centers; and migrated C/C++ codebases to Rust, including the Fuchsia OS Zircon kernel.

The most concrete example is the libgav1 video decoder, where Argon replaced 32,000 lines of SIMD code and produced a memory-safe Rust decoder running 2.7 times faster than the prior Rust port with identical output.

On security, Google says Argon can autonomously find, validate and patch critical vulnerabilities, and that Wiz's Scan for Good initiative used it to uncover a critical vulnerability in hospital software.

Trusted defenders get the model without cyber guardrails, while Google cites a 0.7% attack success rate on Gray Swan's indirect prompt-injection benchmark versus 8.5% for GPT-6 Astra.

The stakes: aggressive pricing, Wall Street watching

Google is pairing the launch with introductory pricing of $2 per million input tokens and $10 per million output, with cached input at a 95% discount — about $0.10 per million — before rising to $4/$20, which matches Claude Opus 5.5's base pricing and undercuts GPT-6 Astra's listed $10/$50, according to VentureBeat's pricing comparison.

The Decoder's independent math complicates the bargain: at promo rates one Intelligence Index task costs $1.99, 60% of GPT-6 Astra's $3.26, but post-intro it rises to $3.98, about 20% above Astra.

Markets responded cautiously-positive: TradingView reports Alphabet shares rose as much as 1.5% pre-market, with JPMorgan maintaining its 'Overweight' rating while saying Google needs to 're-establish itself at the frontier and as an AI leader for both the industry and investors.'

Retail sentiment on Stocktwits shifted to bearish, and analysts flagged Google's lagging Gemini app compared with Claude Cowork and ChatGPT Work.

What comes next

The immediate question is timing. Google has given no date for broad availability, saying only 'as soon as possible' after feedback from early testers feeds into the model's safety mechanisms, The Decoder reports.

Meanwhile the regulatory backdrop is hardening: the FTC has opened a broad probe into AI safety practices at OpenAI and Anthropic, examining potential unfair or deceptive practices and consumer harm from 'rogue AI' systems, per ABC News and CGTN, even as the Trump administration pushes a voluntary industry self-regulation accord.

For enterprise buyers, VentureBeat's takeaway is that model choice remains workload-dependent — Argon leads on breadth of top scores in legal, finance, automation and long-context work, while OpenAI and Anthropic keep advantages in software, science-terminal and computer-use tasks.

Argon's competitive picture will be settled less by Google's benchmark table than by how it performs once developers beyond the Fairwind cohort can actually run it.

Sources

  1. venturebeat.com › Google unveils Gemini 4 Argon, retaking benchmark lead over OpenAI and Anthropic — but in limited release
  2. the-decoder.com › Google Gemini 4 Argon closes the gap with OpenAI and Anthropic but doesn't take a clear lead
  3. tradingview.com › GOOGL Stock Rises After Gemini 4 Argon Launch, But JPMorgan Says Google Must Reclaim AI Leadership
  4. news.cgtn.com › US regulator probes Anthropic and OpenAI over AI safety

More on AI