Gemini 4 Argon: $2 per million input tokens, 95% cached discount

Google released Gemini 4 Argon on October 13, 2026, making its latest frontier model globally available across developer platforms and consumer applications. The model is accessible through multiple channels: the API and AI Studio for developers and enterprises, Google Antigravity for Google AI Ultra subscribers, and the Gemini app plus AI Mode in Search for Ultra and Pro subscribers.

The company positions Argon as a specialist for complex, real-world work spanning coding, finance, legal, cybersecurity, and science. According to the announcement, the model demonstrates depth and precision across multi-step tasks, outperforming other frontier models on benchmarks across professional domains. Argon ranks number one on the Vals Index, a benchmark measuring model performance across critical economic sectors including finance, coding, legal, and tax work.

What changed

Gemini 4 Argon launched on October 13, 2026, with broad availability across multiple platforms. Developers can access it via API with new introductory pricing. Google AI Ultra and Pro subscribers can use it in the Gemini app. Google AI Ultra subscribers have access through Google Antigravity. Google also began expanding Argon to hundreds of additional Fairwind Program partners focused on cybersecurity and defense applications.

The pricing structure aims to encourage adoption during the initial launch period. Introductory rates are two dollars per million input tokens and ten dollars per million output tokens. Cached input tokens are discounted 95% off, providing significant cost reduction when reused context. This pricing remains in effect through November.

Why it matters

Argon positions itself as a competitive option in the frontier model market, particularly for teams building workflows in finance, legal, and cybersecurity. The model’s specialization in professional work offers an alternative to general-purpose competitors.

Early results at Google demonstrate potential organizational value. Software engineers using Argon in Antigravity increased agentic code submissions by more than 35% while reducing code rollbacks by nearly a third. The Chrome team used Argon to complete an architectural refactor that would normally require months of work in a matter of weeks. The refactor separated core webpage handling from approximately six million lines of legacy code within Chromium’s 40 million-line codebase, enabling features that eliminate an entire category of browser crashes. This work will allow webpages to move fluidly between tabs, side panels, and floating windows without reloading.

Performance optimization work shows similar gains. Google’s open-source Agent Development Kit team used Argon to cut in-memory session prep time by 40% in long sessions and accelerate cold imports by 21%. Scientists using Argon optimized the xTB quantum chemistry package, making it 2.3 times faster at predicting molecular structures and 1.8 times faster at conformer search. The model achieved a 2.85 times end-to-end inference speedup on Gemma models via vLLM, accelerating throughput from approximately 37,900 to over 108,000 tokens per second.

For organizations evaluating frontier models, Argon’s focus on multimodal document comprehension, extended reasoning, and domain knowledge in legal, financial, and scientific work suggests differentiation from general-purpose alternatives. The model communicates more directly with less hedging, avoids unnecessary repetition, and skips clichés. The 95% discount on cached tokens particularly benefits workflows that reuse substantial context, such as analyzing repeated document sets or maintaining long-running agent loops. Enterprise customers and startups in early access programs have already achieved strong results with the model.

What to test

Before committing to Argon for production use, teams should validate the company’s specific performance claims for their workflows. The Vals Index benchmark focuses on finance, coding, legal, and tax work, but your use case may differ significantly.

Test the model’s performance on realistic tasks from your domain. Evaluate whether the introductory pricing through November reflects sustainable costs or whether pricing will increase substantially afterward. Benchmark token efficiency and cached token reuse patterns to understand actual per-task costs in your applications. Check how the model handles your specific compliance requirements, particularly for legal and financial work where accuracy matters. Validate claims about reduced hedging and more direct answers on questions critical to your workflow. Review multimodal capabilities including visual comprehension, document processing, and spatial reasoning needed for your use cases. Test integration with existing tools and workflows to understand implementation effort and time to productivity. Review the model’s performance on domain-specific benchmarks relevant to your industry.

Common questions

Is Gemini 4 Argon available now?

Yes. Gemini 4 Argon launched globally on October 13, 2026, and is available today for developers and enterprises via the API and AI Studio, for Google AI Ultra subscribers in Google Antigravity, and for Ultra and Pro subscribers in the Gemini app.

When did Gemini 4 Argon launch?

Gemini 4 Argon became available on October 13, 2026. The model is accessible through multiple platforms including AI Studio, the Gemini app, and the API.

What is the pricing for Gemini 4 Argon?

Introductory pricing is two dollars per million input tokens and ten dollars per million output tokens. Cached input tokens cost 95% less than standard input tokens. The introductory rates apply through November.

The conclusion

Gemini 4 Argon represents Google’s effort to compete in the frontier model market by specializing in professional use cases rather than pursuing general capability. The model’s ranking on the Vals Index and early results at Google demonstrate strength in coding, finance, and legal work. The 95% discount on cached tokens and introductory pricing through November reduce barriers to experimentation.

The key question is whether performance and pricing hold up as introductory terms end in December. Teams should test Argon on representative work now while pricing remains favorable, monitoring whether the model’s specific strengths in domain work justify adoption alongside existing models. Watch for updates on Fairwind Program expansion, which suggests growing emphasis on specialized cybersecurity and defense applications alongside general enterprise access.

Related reading

• Google DeepMind’s AlphaProof Nexus Solves Open Math Problems
• Meta Muse Wonder frontier model launches today
• Multi-Agent AI Teams Burn Up to 5x Tokens for Little Gain