Gemini 4 Argon is Google's new frontier model for long, multi-step work in software engineering, finance, legal and security. Since 13 October 2026 it runs in the Google Antigravity app and CLI for Google AI Ultra subscribers and eligible Gemini Enterprise plans, and in the paid Gemini API at $2 input and $10 output per million tokens until 30 November. From 1 December the price doubles to $4 and $20.
What is Gemini 4 Argon?
Gemini 4 Argon is the first model in Google's Gemini 4 family. Google announced it on 30 September 2026 and opened it widely on 13 October. Google aims it at long, multi-step jobs in three areas: real-world software engineering, knowledge work such as legal and finance, and cybersecurity defence.
According to the Google DeepMind model card, Argon is a large sparse mixture-of-experts (MoE) transformer. It was trained as a new model, not as a fine-tune of Gemini 3. It reads text, images, video, audio and PDFs, and it writes text.
- API model ID: gemini-4-argon
- Context window: 1,048,576 tokens, shared between input and output
- Output: 65,536 tokens by default, or up to about 1 million with long decode
- Thinking levels: low, medium and high
- Tools: function calling, Google Search and Maps grounding, code execution, URL context, file search, structured outputs and computer use (preview)
- Delivery modes: standard, Batch, Flex and Priority
What is Google Antigravity?
Google Antigravity is Google's agent-first development platform. Rather than an editor with a chat box bolted on, it is built around AI agents that plan, write, run and check code across a whole repository. It comes as a desktop app, an IDE, a CLI and an SDK, and plugins, skills and MCP integrations let the agent reach your own tools and data.
Google says Antigravity was designed for models capable enough to work through a task on their own for long stretches, and Argon is the first model it treats as fully meeting that bar. Inside Google, the internal version of Antigravity running Argon now has more than 100,000 monthly users across Alphabet.
How do you get access to Gemini 4 Argon?
Access depends on who you are. The one catch: individual developers only get Argon in Antigravity on the Google AI Ultra plan. Google AI Pro is not enough there.
- Google Antigravity desktop app and CLI: Google AI Ultra subscribers, plus eligible Gemini Enterprise subscriptions. Update Antigravity to the latest version, then pick Gemini 4 Argon in the model selector.
- Gemini API and Google AI Studio: paid-tier API keys, using the model ID gemini-4-argon.
- Gemini Enterprise Agent Platform (Google Cloud): through Model Garden for enterprise workloads.
- Gemini app and AI Mode in Google Search: Google AI Pro and Ultra subscribers.
- Android Studio and Firebase AI Logic: for Android and app developers.
- Fairwind Program: vetted cyber defenders, who got access first, on 30 September.
How much does Gemini 4 Argon cost?
Argon launched at an introductory API price of $2 per million input tokens and $10 per million output tokens. Cached input costs 95% less, which is $0.10 per million tokens at the introductory rate. Batch and Flex requests cost half, and Priority costs more.
At the introductory price, Argon costs the same as Claude Sonnet 5.5 and GPT-6.1 Sol, half the price of Claude Opus 5.5 ($4/$20), and a fifth of GPT-6 Astra ($10/$50). From 1 December it moves to $4/$20, the same as Opus 5.5.
Introductory pricing ends on 30 November 2026. If you are budgeting a production workload, use $4 input and $20 output per million tokens. Long-context and agentic runs consume output tokens quickly, so test with caching and Batch mode before you commit.
Gemini 4 Argon vs Claude Opus 5.5 vs GPT-6 Astra: benchmarks
Most of the figures below come from Google. The Vals Index and the Artificial Analysis Intelligence Index are independent. Argon leads in agentic coding, automation, finance and legal tasks, while Claude Opus 5.5 still leads on terminal work and on overall intelligence rankings.
- DeepSWE v1.1 (agentic software engineering): Argon 77.9%, Claude Opus 5.5 74.2%, GPT-6 Astra 74.1%.
- Zapier AutomationBench: Argon 51.3%, Opus 5.5 42.5%, Astra 41.4%.
- Harvey Legal Agent Benchmark: Argon 19.6%, Astra 5.4%, Opus 5.5 3.8%.
- Vals Index (independent, 12 October): Argon 68.90% in first place, Claude Sonnet 5.5 67.04%, Opus 5.5 66.97%. Argon also leads Vals Finance Agent v2 at 65.40%.
- LVBench (long video understanding): Argon 91.7%.
- Where it trails: Terminal-Bench 4.0, Opus 5.5 66.4% against Argon 57.4%. On the Artificial Analysis Intelligence Index, Opus 5.5 scores 58 and Argon 53, level with GPT-6 Astra.
Benchmarks give a rough guide, not a verdict. Before switching models, run your own 20 to 50 real tasks through each one and compare accuracy, cost per task and how often a person has to step in.
What did Google's own teams do with Argon in Antigravity?
Google tested Argon on its own engineering teams before releasing it. Over four weeks, agent-written code rose 35% and post-submit rollbacks fell 30%, so the code was arriving faster and breaking less. The results Google reported:
- YouTube comments: Argon used production profiling data to find that redundant emoji rendering accounted for almost 45% of server overhead. Twelve fixes later, CPU use was half what it had been, server latency was down 30 ms and client latency down 20 ms.
- Bug backlog: 134 open bugs were checked against commit history and live telemetry. In under 3 hours, 97 stale or duplicate bugs were closed with proof, a 72.4% reduction, and the remaining 37 were sorted by owner.
- Legacy migration: a Fortran scientific codebase was rewritten in Rust, running 1.8x faster and gaining memory safety.
- Security audit: tracing how permissions flow at runtime, Argon found four critical bugs in access-control changes. Two earlier frontier models had passed the same code as clean.
- Wing drone simulator: Argon proposed 18 physics modelling fixes across aerodynamics, propulsion and structural dynamics.
- On-call support: tickets are grouped, root causes traced through code and logs, and replies drafted for a person to approve, saving 2 to 3 hours per shift.
Is Argon only for developers?
No. Usage outside engineering is growing fast inside Google: over half of Cloud sales staff, almost half of finance and over a third of business and operations staff now use Antigravity.
- Localisation: hundreds of documentation pages translated into 17 languages in hours, with people only reviewing.
- Operations: a technical programme manager with no dedicated engineers fixed a data centre dashboard outage that had been stuck for four months, landing a 3,200-line pipeline rewrite in a few days.
- Sales: an automated prospect check cut research time by 90%, saving an account executive more than 10 hours a week.
- Finance: 15.5 hours of human direction produced about 335 hours of checked reconciliation, budget and contract work across 80 workflows.
What does Gemini 4 Argon mean for businesses in Saudi Arabia and the GCC?
The examples Google picked are the same problems we see in projects across Riyadh, Jeddah and Dubai: legacy systems nobody wants to touch, bilingual content that takes weeks, and support queues that keep growing. For those jobs, an agent model that can work alone for hours changes the cost of the work. You still need people who know the business to scope it and check the output.
- Legacy modernisation: moving old ERP, PHP or .NET systems to a modern stack, with tests written first so behaviour is preserved.
- Arabic and English localisation: first drafts in minutes, with a native reviewer for tone, RTL layout and terminology.
- Compliance work: tracing ZATCA e-invoicing and VAT logic through existing code before a Phase 2 integration.
- Support and operations: triaging tickets and drafting replies for a person to approve, in Arabic and English.
Agent models raise the value of clean code, good tests and clear documentation. If your codebase has none of these, fix that first, because the agent will move faster in the wrong direction otherwise.
Should you switch to Gemini 4 Argon?
Try it if your work is long and agentic: multi-file refactors, migrations, bug triage, finance or legal document work, or anything involving long video. The introductory price makes testing cheap until 30 November.
Be careful about three things. First, most headline numbers come from Google, and independent rankings still put Claude Opus 5.5 ahead overall. Second, individuals need the Ultra plan to use it in Antigravity. Third, the price doubles on 1 December. In practice, many teams will run more than one model, routing each task to whichever does it best for the lowest cost.
