Google has unveiled Gemini 4 Argon, its newest frontier model, and pushed it live the same day across the Gemini app, AI Mode in Search and the developer API. Announced on October 13, 2026, Argon is the first release in the Gemini 4 generation and now tops a lineup that also includes Flash and Pro. Google CEO Sundar Pichai confirmed the global rollout, while DeepMind SVP Koray Kavukcuoglu had first introduced the model on September 30 in a limited run restricted to participants in Google’s Fairwind Program.
- Simultaneous rollout: Argon ships to the Gemini app for AI Pro and Ultra subscribers, to AI Mode in Search, and to developers through AI Studio and the Gemini Enterprise Agent Platform.
- Pricing with an expiry date: Introductory rates of $2 per million input tokens and $10 per million output tokens hold until November 30, then double to $4 and $20 on December 1.
- Aimed at heavy lifting: Google is positioning Argon for multi-step reasoning, large-codebase refactoring and cross-document synthesis rather than casual chat.
What sets Argon apart from the rest of the Gemini 4 family
Google describes Argon as the tier where reasoning quality justifies the cost, rather than the fastest or cheapest option. Its headline capability is a long context window, which the launch narrative places at twice that of Pro, alongside a multimodal architecture that accepts text, images, audio and video on the input side. Workloads that previously broke down — such as reconciling a 200-page document set against a second source — are precisely the ones Google claims Argon now holds together.
There is a caveat on the specifications. Third-party gateway documentation published for Day 0 support lists a combined input-and-output window of 1,048,576 tokens, with up to 65,536 output tokens per request, while the launch framing points to a ceiling double that of Pro. The same integration notes confirm selectable thinking levels of low, medium and high, with thinking tokens billed as output and no support for a separate thinking budget.
On Google’s published evaluations, Argon reaches 77.9% on DeepSWE v1.1, its long-horizon software engineering test, and takes the top spot on the Vals Index, which weights finance, coding, legal and tax work by each sector’s contribution to US GDP. On AutomationBench, Zapier’s measure of end-to-end business tasks, it scored 51.3%, roughly nine points clear of the nearest model. Independent results are more mixed: on the Artificial Analysis Intelligence Index, Argon ties GPT-6 Astra at 53 with high reasoning enabled — one point ahead of GPT-6.1 Sol and Muse Wonder Max — but sits about five points behind Claude Opus 5.5 and Claude Sonnet 5.5, and trails GPT-6 Astra on FrontierSWE v2.
Inside Google, Argon is already doing the work
Google says it has been running Argon against its own hardest engineering problems for weeks. In Antigravity, its agentic development platform, the model helped teams lift agentic code submissions by more than 35% while cutting rollbacks by nearly a third. The Chrome team used it to untangle core webpage handling from a roughly 6-million-line legacy folder inside Chromium’s 40-million-line codebase, completing an architectural refactor across hundreds of files that had been scoped in months.
The open-source Agent Development Kit saw Argon reduce in-memory session preparation time by 40% on long sessions and speed up cold imports by 21%. Researchers reported a 2.85x end-to-end inference speedup for the Gemma 4 E2B model via vLLM on an A100, pushing throughput from roughly 37,900 to more than 108,000 tokens per second with full token parity. The quantum chemistry package xTB became up to 2.3x faster at predicting molecular structures and 1.8x faster at conformer searches.
“As software development becomes more agentic, teams need models that can support long-running, multi-step work across all stages of the development lifecycle.” — Stuart Moncada, VP, GitLab
Enterprise partners report similar gains in regulated domains. Harvey’s applied research lead framed the advance as a substantial improvement over Gemini 3.1 Pro in legal agent evaluations, while ServiceNow’s AI products chief cited roughly 20% better task success than Gemini 3.8 Flash, with standout results in HR and customer service.
Pricing, distribution and the enterprise battle
The commercial terms are aggressive. Cache reads cost $0.10 per million tokens during the introductory window, a 95% discount, with Batch and Flex options halving input and output charges. Even after the increase, Argon undercuts Anthropic’s Claude Opus 5.5 at $4/$20 and lands far below GPT-6 Astra at $10/$50. For heavy users, context caching remains the main lever on effective cost.
Beyond the premium tiers, the launch reshapes the consumer experience. Complex queries — comparative analysis, multi-source synthesis, step-by-step reasoning — now route to Argon inside AI Mode instead of Pro, at the price of a slightly longer wait of roughly 1.5 to 3 seconds. The Antigravity coding interface has been folded into the Gemini app, a new toggle switches Gemini from chat to tasks, and the assistant now connects to more than 40 external services including Instacart, OpenTable, Adobe and Shopify. Whether Plus subscribers or free users ever gain access remains an open question.
Argon’s arrival lands a day after Meta introduced its own frontier model and ends a stretch in which Google watched rivals set the pace. The context-window ambiguity in early documentation is worth watching, but the more consequential test is adoption: with a pricing cliff on December 1 and native hooks into BigQuery, Workspace and Vertex AI, Google is betting that enterprises already inside its cloud will treat that deadline as a reason to commit rather than an excuse to wait.
Sources & References