After more than half a year of silence, Google is back at the frontier table. Gemini 4 Argon is the first flagship of the Gemini 4 generation, positioned for complex, long-horizon workflows — real-world software engineering, enterprise knowledge work in legal and finance, and cyber defense. Its single-run output limit jumps from 64K tokens to 1 million tokens: not context window, but how much a model can keep thinking and producing in one continuous task.
1. The scorecard: broadly leading, not universally
Across 19 official benchmarks, Argon takes 13 first places plus one tie. Third-party tests tell the same story:
- DeepSWE v1.1 (long-horizon software engineering): 77.9%, ahead of Claude Opus 5.5 (74.2%) and GPT-6 Astra (74.1%);
- Vals Index (finance / legal / coding / tax knowledge work): 68.9%, No. 1;
- Harvey legal agent benchmark: 19.6%, nearly 4× GPT-6 Astra (5.4%);
- AutomationBench (business workflow automation): 51.3%, ahead of Claude Opus 5.5 (42.5%);
- LVBench long-video understanding: 91.7%; CWE-bench v1 vulnerability fixing: 68%, tied first.
Weak spots remain: terminal operations (Terminal-Bench 4.0) and some coding tasks (FrontierSWE v2) trail GPT-6 Astra and Claude Opus 5.5, and Bloomberg reports internal employee skepticism about real-world coding experience. Bottom line: Argon excels at deep, long-running reasoning, not everything.
2. Google already uses it internally
Datacenter memory optimization
An agent team analyzed telemetry, autonomously identified and applied optimizations, freeing more than 300 TiB of memory — with total savings estimated at 500 TiB to 1 PiB.
C/C++ to Rust migration
Covers re2, libgav1 and the Fuchsia Zircon kernel (800K+ lines); in libgav1, Argon rewrote ~32,000 lines of SIMD code and delivered a 2.7× speedup.
Quantum algorithm optimization
Argon improved a published space-time resource benchmark by 40% in minutes.
3. Pricing: turning compute advantage into cost advantage
Argon launches at $2 per million input tokens and $10 output (promo; standard $4/$20), with cached input 95% cheaper. GPT-6 Astra sits at $10/$50 — Google is roughly halving the cost of frontier inference. For developers, top-tier model access is getting cheaper, but compute remains the hardest cost line.
4. Rollout: security teams first, then API and subscriptions
Argon is currently limited to trusted cyber-defense teams via the Fairwind program; paid API customers and Google AI Ultra subscribers come next. It's a staged release to harden safety before broad access.
What it means for you: smarter models make compute the real bottleneck
Workhorse models like Gemini 4 can burn hundreds of thousands of tokens in a single run — far beyond chat workloads. Building your own high-end rig means hardware depreciation, power, and maintenance. For individuals and small teams, cloud GPU on demand is the cost-effective path: elastic scaling, pay-per-use, zero ops.
Whether you're fine-tuning open models or calling frontier APIs, having cloud compute ready is what lets you ride the next wave without a capital splurge.
5. Local rig vs cloud GPU: the math
| Dimension | Self-hosted GPU | Cloud GPU / cloud LLM API |
|---|---|---|
| Upfront cost | Thousands of dollars; hardware churns yearly | Start for tens of dollars, scale as needed |
| Elasticity | Fixed capacity; idle or queued | Minutes to scale up/down, pay per use |
| Ops burden | Drivers, environment, cooling, uptime on you | Managed platform; swap instance on failure |
| Typical use | Steady, predictable long-running loads | Dev/test, inference services, peak bursts |
Put compute in the cloud and start building today
New-user promo pricing on GPU cloud servers — spin up your first AI app in minutes.
Sources
- Google official: Gemini 4 Argon — our next era of frontier intelligence
- 36Kr: Google releases Gemini 4 Argon with 1M-token output limit
- QbitAI via BAAI Hub: Google's Gemini 4 surprise launch
- Tencent Cloud Dev Community: Gemini 4 Argon ties GPT-6 Astra, but you can't use it yet
- Bloomberg: Google Grapples With Employee Skepticism About New Gemini Model