Announcing Gemini 4 Argon, our frontier model for real-world coding, enterprise knowledge work, and cyber defense, rolling out soon.
The launch took place at Google's Mountain View headquarters, with Alphabet CEO Sundar Pichai and Google DeepMind chief Demis Hassabis on stage together. Argon uses a sparse mixture-of-experts architecture with 1.8 trillion total parameters and roughly 120 billion activated per pass, natively handling text, images, audio, video and code. Input context reaches 2 million tokens with a 200,000-token output ceiling, and a thinking budget adjustable from 0 to 65,536 tokens per request. Published benchmarks include 92.4% on GPQA Diamond, 78.6% on SWE-bench Verified, a perfect score on AIME 2026, 81.3% on MMMU-Pro and 46.8% on ARC-AGI-2. Two companion models launched alongside:
Gemini 3 shipped in November 2025, and Google left its flagship base model untouched for nearly eleven months while OpenAI rolled out follow-ups to GPT-5 and Anthropic iterated on Claude 4.x. Meta attacked the open-weight flank with Llama 5. The competitive paradigm moved too: 2023 and 2024 were about scaling pretraining, while after 2025 the emphasis shifted to test-time compute, letting models think longer before answering. That demands tunable reasoning budgets, stable long-context memory and reliable tool use, all of which lean heavily on in-house si
From a technical standpoint, the most consequential thing about Argon is not the 2-million-token window but the promotion of thinking budget to a first-class setting. Developers can allocate anywhere from zero to 65,536 tokens of reasoning per request, which means a single set of weights can serve both a customer-support lookup and a repository-scale refactor. DeepMind's numbers back that up: at low budget, Argon Flash posts a median latency of 0.42 seconds, close to Gemini 3 Flash, while a maxed-out budget lifts SWE-bench Verified from 51.2% to 78.6%, a 27.4-point spread. Only OpenAI's reasoning line had achieved comparable gear-shifting before, and it did so with separate models; Google baked the capability into a general-purpose one. The other hard number is long-context retrieval, where Argon hits 99.2% accuracy at 1 million tokens against Gemini 3 Pro's 94.7%. That 4.5-point gap matters because at million-token scale, 95% versus 99% decides whether a team can hand over an entire filing or codebase without chunking it first. The sparse expert layout is the underlying reason costs hold: 1.8 trillion total parameters activating roughly 120 billion means each pass burns about a third of the compute a dense model of equivalent size would.
On the business side, Argon's pricing looks more like an offensive weapon than its benchmarks do. At $1.25 per million input tokens and $10 per million output tokens, it undercuts Gemini 3 Pro's $1.75 and $14 by roughly 29%. Flash at $0.15 and $0.60 lands near 70% of Claude Haiku 4 and GPT-5 mini. Google can afford this because inference runs on its own Ironwood TPUs rather than through a third party's margin. Distribution is the second weapon: 650 million Gemini App users, the default Android entry point and 3 billion Workspace accounts mean a launch reaches ordinary users the same day, not just developers. Enterprise signals were immediate, with Salesforce and SAP announcing Vertex AI integrations on launch day, a step that normally takes weeks of negotiation. For Google, the commercial goal is probably not profit on the model itself but keeping customers inside the cloud and Workspace payment chain.
For the industry, Argon's message is that foundation models are commoditising fast. When 2-million-token context and a 78.6% SWE-bench score sell for $1.25 per million tokens, the room to charge purely for capability differences shrinks quickly. Value migrates in two directions: proprietary data and workflows, and the agent orchestration layer. Llama 5 405B sits around 62% on SWE-bench Verified, so the open-versus-closed capability gap is still widening, but the unit-cost gap has narrowed sharply, which changes selection logic for smaller teams most of all. Liu Gong predicts that within twelve months, major vendors will push long-context input pricing below $1 per million tokens, and that real differentiation will show up in tool-call success rates and long-horizon task stability, two areas that still lack a shared benchmark.
- Context window: Argon at 2 million tokens, Gemini 3 Pro at 1 million, Claude 4.5 Opus at 500K, GPT-5.2 at 400K.
- GPQA Diamond: Argon 92.4%, GPT-5.2 89.0%, Claude 4.5 Opus 87.6%, Gemini 3 Pro 85.1%.
- SWE-bench Verified: Argon 78.6%, an 11.4-point gain over Gemini 3 Pro at 67.2%; Llama 5 405B sits near 62%.
- Pricing: Argon input at $1.25 per million tokens is 28.6% below Gemini 3 Pro's $1.75; Argon Flash output at $0.60 is roughly 75% of Claude Haiku 4's $0.80.
For the industry, the near-term reaction within three months is price matching. On launch day both Azure AI and AWS Bedrock updated pricing pages for long-context models, and major clouds will likely roll out comparable cuts and longer windows in the fourth quarter. Enterprise proof-of-concept work will shift from multi-model routing toward long-context agents, with legal review, audit and code migration among the first workloads to land, since they need whole-corpus reads. Over a one-year horizon, input pricing below $1 per million tokens looks reachabl
Competitor responses arrived within hours. Sam Altman welcomed the competition in a post on X while teasing GPT-5.5 for the coming weeks. Anthropic avoided direct benchmark comparisons, stressing instead its lead in long-horizon reliability and safety evaluation, and reiterating Claude 4.5's completion-rate advantage on autonomous agent tasks. Meta chief AI scientist Yann LeCun questioned the ecological validity of benchmarks like ARC-AGI-2, noting that 46.8% remains far below human baselines and does not justify calling a model generally capable.
Indep
Questions worth exploring:
- With thinking budget as a first-class control, model procurement shifts from single-shot accuracy to cost per dollar per task, forcing enterprise TCO models to be rewritten.
- Whether a 2-million-token window genuinely replaces RAG architecture depends on where the retrieval-accuracy and inference-cost curves intersect.
- As the foundation layer enters a price war, open-weight models must choose between vertical capability upmarket and extreme low-cost inference downmarket.
This item comes from DeepMind Blog (published 2026-09-30 20:01:45). Wenfeng Radar provides a Chinese/English summary and commentary on overseas AI developments; the original content is owned by its author.
Daily aggregation of overseas AI developments with in-depth analysis. Bookmark us to never miss an important signal; back to homepage.