본문으로 건너뛰기

Gemini 4 Argon: The Field Guide for Claude Users (Sep 30, 2026)

중급

Google opened the Gemini 4 generation on September 30, 2026 with a single model, Argon, and a number nobody else has: a 1 million token output limit, up from the previous 64K. Not the context window — that stayed at 1M. The output ceiling went up 16x.

It landed at the top of Hacker News and stayed there. It also landed in a strange place commercially: Google's own post says Argon "leads or ties for the highest score in more categories than GPT-6 Astra or Claude Opus 5.5," while the independent Artificial Analysis Intelligence Index published the same week puts Argon at 53 against Claude Opus 5.5's 58. Both statements are accurate. They measure different things, and the gap between them is the most useful thing on this page.

This lesson is for a Claude user deciding whether Argon changes anything about their routing. It covers what the 1M output limit actually buys, how to read two contradictory benchmark stories, the per-task economics that the "40% cheaper" headlines invert, and why your access request will probably be declined.

What you'll learn
  • Separate the 1M output limit from the 1M context window — they are different ceilings and only one of them moved
  • Reconcile Google's '13 of 18 categories' claim with the independent index that places Argon 5 points below Claude Opus 5.5
  • Price a task properly: why 62K output tokens against Astra's 27K means Argon's advantage is its rate card, not its efficiency
  • Read the 15% hallucination rate for what it is — an abstention behaviour, not a knowledge win
  • Understand what the Fairwind Program gates, what it requires of your org, and why Argon is effectively uncallable for most readers

The one-sentence version​

Gemini 4 Argon is Google's new frontier model, 1M-token input context and a 1M-token output limit (up from 64K), priced at an introductory $2 in / $10 out per 1M tokens with cached input at 95% off, rising to $4 / $20 after the introductory period. It is not generally available: it is rolling out to vetted cyber defenders through the Fairwind Program, then to paid API customers and Google AI Ultra subscribers. On Google's own 18-benchmark set it leads or ties in 13 categories; on the independent Artificial Analysis Intelligence Index it scores 53, tying GPT-6 Astra and Claude Fable 5.1 and sitting five points below Claude Opus 5.5.

The number that is not what it looks like​

Almost every summary of this launch says "1 million tokens" and leaves the reader to assume a context window. The context window did not change. Gemini has been at 1M input tokens since the 3.x line — Gemini 3.8 Flash has the same 1M window and a 64K output cap.

What moved is the output ceiling, from 64K to 1M. These are separate limits and they fail in separate ways:

LimitWhat it governsGemini 3.8 FlashGemini 4 Argon
Input contextHow much you can send — files, transcripts, repo1M1M
Max outputHow much the model can emit in one response, including reasoning tokens64K1M

The practical consequence is about where a long task breaks. With a 64K output cap, a genuinely long-horizon job — refactor this subsystem, draft this filing, triage this codebase — has to be chunked by you, and every chunk boundary is a place where the model loses its own working state. Google's pitch is that Argon can hold a single uninterrupted chain of reasoning and execution across the whole job instead.

Three things to keep in mind before you treat that as free:

  • Reasoning tokens come out of the same budget. On Gemini, thinking tokens are output tokens. A 1M output allowance is not 1M tokens of deliverable; it is 1M tokens of thinking plus deliverable.
  • Output is the expensive side. At introductory pricing, output costs 5x input ($10 vs $2). A response that actually used the full million output tokens would cost $10 on its own, and $20 once standard pricing lands.
  • Nothing obliges the model to use it. A raised ceiling changes what is possible, not what is typical. Argon averages 62K output tokens per Intelligence Index task — around 6% of its ceiling.

Two benchmark stories, both true​

This is the part worth slowing down for, because the same pattern will repeat at every launch for the rest of the year.

Story one — Google's evaluation report. Across 18 disclosed benchmarks, Argon leads or ties in 13 (12 outright leads, 1 tie):

BenchmarkArgonGPT-6 AstraClaude Opus 5.5
Harvey Legal Agent19.6%5.4%3.8%
GraphWalks84.2%71.8%66.8%
Vals Finance Agent v265.4%53.5%58.6%
AutomationBench51.3%41.4%42.5%
LVBench (long video)91.7%87.5%83.7%
DeepSWE v1.177.9%74.1%74.2%
CWE-bench v168%68% (tie)67%
Gray Swan IPI (attack success — lower is better)0.7%8.5%1.0%
FrontierSWE v255.0%65.5%—
Terminal-Bench Science 0.157.6%68.1%—
Terminal-bench 4.057.4%—66.4%
PostTrainBench45.3%—49.3%

Story two — Artificial Analysis' Intelligence Index, an independent aggregate run the same week:

ModelIntelligence Index
Claude Opus 5.558 (57.6)
Claude Sonnet 5.556
Gemini 4 Argon (High)53 (52.6)
GPT-6 Astra (max)53
Claude Fable 5.153
GPT-6.1 Sol (max)52

Argon ranks #8 of 224 models on that index — excellent, and five points behind Anthropic's flagship.

Neither source is lying. They diverge for reasons you can name:

  1. Benchmark selection is a choice. Google picked 18 benchmarks that include three domain-agent evals (legal, finance, automation) where Argon's margins are enormous. An aggregate index weights a different basket. Change the basket, change the winner.
  2. The losses cluster. Argon's four clear losses — FrontierSWE v2, Terminal-Bench Science, Terminal-bench 4.0, PostTrainBench — are all interactive, multi-turn, tool-driven agent work. Its wins cluster in single-shot domain reasoning and long-context retrieval. That is a shape, not noise, and it maps cleanly onto routing decisions.
  3. Absolute numbers matter more than ratios. Harvey Legal Agent is the launch's most quoted row: 19.6% against 5.4%, a 3.6x lead. It is also a score of 19.6% — Argon fails roughly four out of five tasks on that benchmark. "Best available" and "good enough to deploy unsupervised" are different claims.
  4. Vendor benchmarks are not production behaviour. Trending Topics cites Bloomberg reporting that, per Google employees, the model performs worse in practical coding use than its scores suggest. Treat that as a single unverified report — but it is the kind of gap the is the model nerfed? protocol exists to measure.
Pro tip
  • When a launch post says 'leads or ties in N of M categories,' the load-bearing question is who chose M. Go find an aggregate index you did not have to be invited into, and compare the two baskets before you re-route anything.

The economics are the opposite of the headline​

Several outlets led with Argon being around 40% cheaper than GPT-6 Astra. The rate card supports that. Per-task cost tells a more interesting story, because Argon is a heavy token spender.

Artificial Analysis measured average output per Intelligence Index task:

  • Gemini 4 Argon: ~62,000 output tokens
  • GPT-6 Astra: ~27,000 output tokens

Argon emits roughly 2.3x more tokens to reach the same index score. Its cheaper bill is a pricing artefact, not an efficiency win. The measured cost per task:

ModelCost per Intelligence Index taskIndex score
GPT-6.1 Sol (max)$0.7452
Gemini 4 Argon (intro pricing)$1.9953
GPT-6 Astra (max)$3.2653
Gemini 4 Argon (standard pricing)$3.9853
Claude Opus 5.5 (max)$5.9858

Read that table twice. Three separate conclusions fall out of it:

  • Argon's advantage over Astra is scheduled to expire. At $2/$10 it costs $1.99 per task against Astra's $3.26 — a 39% saving. When pricing doubles to $4/$20, it costs $3.98, which is more than Astra. The entire cost argument lives inside the introductory window, and Google has not said how long that window is.
  • GPT-6.1 Sol is the actual value pick on this table. One index point below Argon at 37% of Argon's intro cost and 19% of Opus 5.5's. If your work is not in Argon's strong domains, the cheap tier is where the ratio lives — the same lesson as every other month.
  • Opus 5.5 costs 3x Argon's intro price for 5 more points. Whether that is worth it is entirely workload-dependent, which is the honest answer and also the reason choosing a model is a framework rather than a leaderboard.

Measure tokens-per-task yourself before trusting any per-task cost claim

For the same 20 representative tasks from my own workload, I want to compare two models on cost, not on benchmark score.

For each model, log per task:
- input tokens, output tokens (split reasoning vs final answer if the API reports it)
- wall-clock latency
- a 0-5 quality score I assign blind, without knowing which model produced it

Then compute, per model:
- mean output tokens per task
- mean cost per task at CURRENT list price
- mean cost per task at the post-introductory price, if one is announced
- cost per quality point

Flag any model whose cost advantage disappears at post-introductory pricing.
Do not quote vendor benchmark scores back to me — I only want my own 20 tasks.

The 15% hallucination rate, read correctly​

On AA-Omniscience, the component of the index that measures whether a model asserts false things, Argon posts a 15% hallucination rate against 51% for GPT-6 Astra and 54% for GPT-6.1 Sol. That is not an incremental improvement; it is a different regime, and it is the most under-reported number in the launch.

The mechanism matters. Artificial Analysis' characterisation is that Argon is far more likely to admit it does not know than to produce a confident wrong answer. So the win is in calibration and abstention, not in knowing more. Practical consequences:

  • For retrieval, research and compliance work this is the headline feature, well ahead of any coding number. A model that says "not in the provided documents" is worth more than one that scores two points higher and invents a citation.
  • It will read as less helpful in casual use. A model tuned to abstain refuses to guess, and users experience refusal-to-guess as hedging. Expect "it keeps saying it doesn't know" complaints from exactly the same capability that makes it trustworthy.
  • It does not transfer to your prompts automatically. Abstention is a trained disposition interacting with your instructions. Tell a well-calibrated model "always give a definitive answer" and you have spent the advantage.

You probably cannot call it​

This is where most readers' interest ends in practice. Argon is not a public endpoint. It is gated behind the Fairwind Program, the same mechanism Google introduced for Gemini 3.8 Flash Cyber — but this time it gates the flagship, not a cyber-specific SKU.

Fairwind prioritises governments and national cyber authorities, critical-infrastructure operators (healthcare, telecoms, energy, finance) and core technology platforms, and vets applicants for "a proven track record of ethical operations and research." Accepted organisations must agree to:

Guided walkthrough1 of 6
  1. Not an org-level API key. Each human using Argon must be individually authenticated behind phishing-resistant MFA, plus applicable access controls. That rules out the usual pattern of one service credential shared by a team or embedded in a deployment.

For trusted defenders, Fairwind can pair Argon with CodeMender, Google's code-security agent that automates vulnerability fixes, and grants a build of Argon without the cyber guardrails that the eventual public model will carry. The cyber case is strong on its own numbers: real-world vulnerability discovery at 85.8% (against 71.0% for Gemini 3.8 Flash Cyber), the Wiz penetration-test benchmark at 70.9% (from 58.2%), CWE-bench v1 at 68%, and the best measured prompt-injection resistance of any frontier model — 0.7% attack success at k=15 on Gray Swan's indirect prompt-injection suite, against 1.0% for Claude Opus 5.5 and Fable, and 8.5% for GPT-6 Astra.

That last number deserves its own sentence. An 8.5% success rate means roughly one in twelve indirect injection attempts lands; 0.7% means about one in 143. If you build agents that read untrusted web content, that is a twelve-fold difference in exposure — and it is measured at k=15, meaning fifteen attempts per target, not one. Worth reading alongside prompt injection and coding agents under attack.

Watch out
  • Because Fairwind forbids sharing, redistribution and resale, any gateway, proxy or 'unlocked Argon' reseller claiming to offer it is either misrepresenting a different model or violating the program terms. Treat an Argon endpoint you did not get from Google as untrustworthy by default.

What Google says it did with it internally​

The launch post's internal case studies are more informative than its benchmark table, because they describe long-horizon work with verifiable outcomes:

  • Quantum algorithm optimisation — beat the published baseline for spacetime resource cost by 40%.
  • Fleet memory profiling — found optimisations freeing over 300 TiB, with an estimated 500 TiB to 1 PiB total once applied.
  • C/C++ to Rust migration — including 800K+ lines of the Fuchsia Zircon kernel.
  • libgav1 video decoder — replaced 32K lines of hand-written SIMD, producing code 2.7x faster than the Rust port while emitting bit-identical output.

The libgav1 result is the most instructive one on the list. "Bit-identical output" is the detail that makes it credible: there was a mechanical oracle for correctness, so the model's work could be verified rather than reviewed. That is the shape of task where a 1M output budget genuinely pays — large, mechanical, and checkable by a machine. It is also the shape of task most teams do not have, which is worth being honest about before extrapolating from these numbers to your own backlog.

Should you change your routing?​

Short answer for most readers: no, because you cannot call it. When it opens up, here is the honest split based on where the benchmarks cluster.

WorkWhere it should goWhy
Interactive coding agents, terminal work, computer useClaude (Opus 5.5 / Sonnet 5.5)Argon's four clear losses are all in this family — Terminal-bench 4.0 57.4 vs 66.4, FrontierSWE v2 55.0 vs Astra's 65.5
Legal and financial document reasoningArgon, with reviewHarvey 19.6% vs 3.8–5.4%, Vals Finance 65.4% — but 19.6% absolute means supervision, not automation
Long-video and long-context retrievalArgonLVBench 91.7%, GraphWalks 84.2%
Research where a wrong citation is worse than no answerArgon15% hallucination rate against 51–54%
Defensive security, autonomous patchingArgon via Fairwind, if you qualify85.8% vuln discovery, CWE-bench 68%, lowest injection susceptibility
High-volume everyday workGPT-6.1 Sol or a Flash/Sonnet tier$0.74 per task for one index point less than Argon
Anything you are budgeting past the intro windowRe-price it first$2/$10 becomes $4/$20; cost per task goes $1.99 to $3.98

Check yourself

0/5
  1. Gemini 4 Argon's headline '1 million tokens' refers to which limit?
  2. Google says Argon 'leads or ties in more categories' than GPT-6 Astra or Claude Opus 5.5, yet Artificial Analysis scores it 53 against Opus 5.5's 58. What is the best explanation?
  3. Argon costs $1.99 per Intelligence Index task against GPT-6 Astra's $3.26. Why is it wrong to call Argon the more efficient model?
  4. Argon's 15% hallucination rate on AA-Omniscience, against 51% for GPT-6 Astra, is primarily the result of what?
  5. Your company wants Argon access through the Fairwind Program for its internal developer platform. What is the blocking problem?
Enter 또는 스페이스 키를 눌러 카드를 뒤집습니다. 좌우 화살표 키로 카드를 이동할 수 있습니다.용어가 표시되었습니다.
1 / 12
Key takeaways
  • The 1M is the OUTPUT limit (up from 64K), not the context window, which stayed at 1M. Reasoning tokens spend that budget, and output costs 5x input.
  • Google's '13 of 18 categories' and Artificial Analysis' 53-vs-58 are both true about different baskets. Argon's losses cluster in interactive multi-turn agent work; its wins cluster in single-shot domain reasoning and long-context retrieval.
  • Argon burns ~62K output tokens per task against Astra's ~27K. The cheaper bill is the rate card, not efficiency — and at standard pricing cost per task goes $1.99 to $3.98, above Astra's $3.26.
  • The 15% hallucination rate against 51–54% is the most under-reported number here, and it comes from abstention. It is the strongest argument for routing research and compliance work to Argon, and the first thing a 'always answer definitively' instruction destroys.
  • You almost certainly cannot call it: Fairwind gates it to internal cyber/IR/pentest teams with per-user phishing-resistant MFA and no redistribution. No API model ID has been published. Any third-party 'Argon' endpoint is suspect.

Sources & further reading​

Next​