Google launched its new flagship on October 13, 2026 — and it arrived carrying more baggage than most model releases. Gemini 4 Argon quietly replaces the Gemini 3.5 Pro that Google teased at I/O in May and never shipped, after months of delays and a shakeup at the top of DeepMind. Reuters reported global availability to developers and businesses the next day.
The headline isn’t a benchmark. It’s an engineering change: Argon can output up to 1 million tokens in a single response — roughly 16× the old 64K ceiling — which turns a code-completion model into one that can finish whole tasks. Add introductory pricing of $2/$10 per million tokens that doubles on December 1, and this is a release with a clock on it.
One honesty note up front: every benchmark figure below is Google’s own. Access is still staged, so independent verification hasn’t caught up yet.
Key Takeaways
-
- Launched October 13, 2026 — Google’s first flagship above the Flash tier in 7+ months, replacing the never-shipped Gemini 3.5 Pro.
-
- 1M-token output (default 65,536; up to 1,048,576 with long decode) on a unified 1,048,576-token context — the output ceiling, not the leaderboard, is the real story for agent workflows.
-
- Intro pricing: $2/1M input, $10/1M output — cached input 95% off, Batch API half price. Doubles to $4/$20 after November 30.
-
- DeepSWE v1.1: 77.9% per Google — ahead of Claude Opus 5.5 (74.2%) and just above DeepSeek V4.1 Flash’s 77.0 from days earlier. All figures are vendor-reported.
-
- Access is staged: paid Gemini API, AI Studio, Google AI Pro/Ultra, and the Antigravity coding tool (Ultra only).
The 1M-Output Story: Why the Ceiling Matters More Than the Leaderboard
Most model launches fight over benchmark points. Argon’s most consequential number is 1,048,576 — the maximum tokens it can generate in one response, up from 64K on the previous generation.
Why does that matter? Agent workflows don’t fail on reasoning nearly as often as they fail on truncation: a coding agent halfway through a multi-file refactor, a research agent cut off mid-report. A 1M-token output budget means the model can plan, write, and finish in one pass instead of being stitched together across calls. Google’s own example: a Chrome team refactored a ~6M-line legacy folder inside Chromium’s 40M-line
code-base
— the kind of job that dies on output limits.
Combined with the unified 1M-token context, Argon is positioned less as a chatbot upgrade and more as long-horizon infrastructure: coding, finance, law, cybersecurity, and science workflows where the task is measured in hours, not turns.
Gemini 4 Argon Pricing: The December 1 Clock
| Input / 1M tokens | Output / 1M tokens | |
|---|---|---|
| Intro (until Nov 30) | $2 | $10 |
| Standard (from Dec 1) | $4 | $20 |
Cached input is 95% off, and the Batch API halves both rates. For context: the intro rate matches GPT-6.1 Sol and Claude Sonnet 5.5 ($2/$10); from December 1 it matches Claude Opus 5.5 ($4/$20); and it undercuts GPT-6 Astra ($10/$50) throughout.
What that means in practice: a workload of 100M input + 20M output tokens per month costs roughly $400 at the intro rate — and $800 from December 1. If you’re evaluating Argon for production, the evaluation window is also the discount window.
Running the numbers on Argon? See how the other end of the market prices out in our Claude Haiku 5.5 pricing breakdown.
Benchmarks: What Google Claims (and Where It Loses)
Google’s launch table has Argon winning or tying 13 of 18–19 benchmarks. The highlights:
| Benchmark | Gemini 4 Argon | Notes |
|---|---|---|
| DeepSWE v1.1 | 77.9% | Opus 5.5: 74.2%; GPT-6 Astra: 74.1% |
| Vals Index | #1 | Finance, coding, legal, tax |
| Vals Finance Agent v2 | 65.4% | Widest gap in the set |
| AutomationBench | 51.3% | #1 (Zapier) |
| CWE-bench v1 | 68% | Tie for first |
| LVBench | 91.7% | Long-video understanding |
| ERQA | 87% | Embodied reasoning |
A
nd the honest part: Argon loses on FrontierSWE and Terminal-bench in Google’s own table. On safety, Google reports bio red-teaming vulnerabilities under 1% and a 0.7% attack success rate on Gray Swan’s indirect prompt-injection suite — again, self-reported.
Treat all of this as directional until independent testers get sustained access. Google’s early-customer list (Deloitte, LexisNexis, Ambience Healthcare, Isomorphic Labs) and its internal figures — agentic code submissions up 35%, rollbacks down a third — suggest real usage, but they’re still Google’s numbers.
The DeepSWE Face-Off: Argon 77.9 vs DeepSeek Flash 77.0
Days before Argon’s launch, DeepSeek shipped V4.1-Flash-1017 with a DeepSWE v1.1 score of 77.0. Argon’s 77.9 edges it out — a near-tie between Google’s flagship and an open-weights budget model that costs $0.15/1M input off-peak.
That gap (0.9 points) versus that price gap (more than 10×) is the real comparison shoppers should sit with. If you want the full picture on the cheaper side of this face-off, see our breakdown of DeepSeek V4.1 Flash pricing and benchmarks (coming soon). If your workloads are smaller and cost-sensitive, the Haiku 5.5 pricing changes are worth a look too — and Google’s image side is covered in our Nano Banana 2.1 explainer (coming soon).
How to Actually Get It
Access is staged, not universal:
-
- Developers: paid Gemini API tiers and AI Studio
-
- Consumers and pros: Google AI Pro and Ultra plans (Gemini app and AI Mode in Search)
-
- Coding: the Antigravity tool, top tier (Ultra) only
-
- Special programs: the Fairwind Program gives hundreds of vetted cyber defenders a guardrail-free version; 10,000 scientists receive 6 months of free top-tier Antigravity
If you can’t see Argon in your account yet, that’s the rollout working as designed — not a bug.
Verdict: Should You Switch?
Argon’s case rests on three things: the 1M-token output ceiling (a genuine workflow change, not a spec-sheet trophy), flagship-tier claimed benchmarks, and an intro price that matches mid-tier rivals until November 30.
The caveats are equally clear: the benchmarks are unverified outside Google, access is still rolling out, and the price doubles in about six weeks. The sensible play: if long-horizon agent work is your bottleneck, evaluate now while the intro rate lasts — and benchmark it on your own tasks before the December 1 price lands.
FAQ
Is Gemini 4 Argon’s $2/$10 pricing permanent?
No. $2 per million input tokens and $10 per million output tokens is introductory pricing. It doubles to $4/$20 after November 30, 2026 (from December 1).
When was Gemini 4 Argon released?
Google DeepMind announced it on October 13, 2026, with Reuters reporting global availability to developers and businesses on October 14. It replaces the Gemini 3.5 Pro that was teased at I/O in May but never shipped.
How does Gemini 4 Argon compare to Claude Opus 5.5?
On Google’s own DeepSWE v1.1 numbers, Argon scores 77.9% vs Opus 5.5’s 74.2%. Pricing ties from December 1 at $4/$20; Argon’s intro rate ($2/$10) undercuts Opus 5.5 until then. Independent head-to-head testing isn’t available yet.
Can I use Gemini 4 Argon today?
Possibly — access is staged. It’s live on paid Gemini API tiers, AI Studio, Google AI Pro/Ultra plans, and Antigravity (Ultra only). If it’s not in your account yet, the rollout hasn’t reached you.
What does “1M-token output” actually mean?
Argon can generate up to 1,048,576 tokens in a single response (the default is 65,536; the maximum requires long-decode settings). That’s roughly 16× the previous 64K ceiling, enabling agents to complete large multi-step tasks in one pass.
Never Miss a Model Launch
AI pricing moves fast — Argon’s $2/$10 intro rate expires December 1, and DeepSeek’s free Search API window closes November 1. Bookmark aiblitzo.com for weekly AI model breakdowns with verified pricing, honest benchmarks, and zero hype.
