Google DeepMind released Gemini 4 Argon on October 13, 2026, a flagship model with a 2 million token input context, a 200,000 token output ceiling, and a Deep Think mode that can run autonomously for up to 24 hours on a single task. Google reported SWE-bench Verified at 82.4% and GPQA Diamond at 92.5%, positioning the model as a problem solver for research, engineering and large-scale code migration rather than a conversational assistant. It ships immediately through the Gemini app, AI Studio
At 16:01 on October 13, 2026, Google DeepMind published Gemini 4 Argon on its official blog, naming the model after the noble gas argon and framing it around solving the hardest problems. The release covers two tiers: Gemini 4 Argon and Gemini 4 Argon Flash. The published specifications are a 2 million token input context, a 200,000 token single-pass output limit, and a Deep Think mode that can execute a single task autonomously for up to 24 hours. The model natively ingests video, audio, PDFs and whole code repositories, and exposes browser and terminal control interfaces. Google reported SWE-bench Verified at 82.4%, GPQA Diamond at 92.5%, AIME 2025 at 99% and ARC-AGI-2 at 45%. Pricing is set at USD 3 per million input tokens and USD 15 per million output tokens for Argon, and USD 0.50 and USD 3 for Flash, with Deep Think billed separately by thinking compute. Distribution runs through
Since Gemini 3 shipped in late 2025, Google has kept to roughly a yearly cadence while OpenAI and Anthropic iterated densely in the same window, shifting the competitive focus from multimodal understanding to long-horizon agent execution. Enterprise feedback has been consistent: most automation failures come not from weak single-step capability but from broken consistency over long runs, where a workflow forgets its original constraints by step forty or drifts slowly across repeated judgments. That is the direct motive for pushing autonomous execution from a four-hour to a 24-hour horizon. Compute is the second precondition. Google's in-house TPU fleet and inference stack make 24-hour
Liu Gong's take: the decisive parameter in Gemini 4 Argon is not the benchmark table but the 24-hour autonomous execution window. Within six months, at least two other frontier labs will put continuous autonomous runtime at the centre of their launch keynotes, and static question-bank evaluations will be pushed to the margins of enterprise procurement within a year. From a technical standpoint, the combination of a 2 million token input context and a 200,000 token output ceiling is underrated. By published specifications, most same-generation frontier models cap single-pass output under 100,000 tokens; doubling that ceiling lets Argon emit a complete module or a full experimental report in one pass rather than stitching segments together, which removes state loss from the engineering chain. Liu Gong would add a caveat: a long context is not the same as an effective long context. In industrial quality inspection, whether a model holds the same decision boundary across an eight-hour shift matters far more than 99% accuracy on a single frame, because drift on a production line means batches of misjudgments and rework. Moving Deep Think from four hours to 24 hours is a bet on memory consistency rather than peak reasoning, which is the right direction and the hardest part to copy.
On the business side, pricing at USD 3 and USD 15 per million tokens sits clearly above the previous generation, and pairing it with thinking-compute billing effectively shifts the unit of account from tokens to tasks. Buyers will need to rewrite their ROI math: if a 24-hour autonomous run compresses a code migration from three person-weeks to two days, doubling the per-token price is not a barrier. Flash carries volume at a sixfold discount, creating a high-low portfolio. The risk sits in inference cost, since gross margin on 24 hours of continuous reasoning depends on amortising in-house TPUs, which is precisely Google's most structural advantage over rivals reliant on third-party compute.
For the industry, the first shock lands on software delivery and research support. Repository-scale migrations, long-cycle literature reviews and iterative experiment design, work currently handled by outsourcing teams and junior researchers, will be repriced within twelve months. For developer tooling this is both pressure and opportunity: products such as Cursor and Copilot must either build long-task orchestration into their own moat or become a front end for Argon. For Chinese vendors the signal is equally clear: context length and leaderboard scores are no longer sufficient; the next threshold is whether an agent can run reliably through a full working day. Liu Gong predicts that before the first quarter of 2027, long-horizon interruption rate will become a standard field on frontier model cards.
- Input context: 2 million tokens, double the 1 million of Gemini 3, enough to load a mid-sized code repository or several hundred pages of lab records in one pass.
- Output ceiling: 200,000 tokens in a single pass; by published specifications most same-generation frontier models cap output under 100,000 tokens, roughly half of Argon.
- Coding performance: SWE-bench Verified 82.4%, about 5 percentage points above the published Gemini 3 Pro figure; GPQA Diamond 92.5%, up roughly 4 points.
- Autonomy window: Deep Think supports up to 24 hours of continuous execution per task, six times the previous four-hour limit.
- Pricing: Argon at USD 3 input and USD 15 output per million tokens; Flash at
For the industry, the short three-month view is a wave of reworked agent benchmarks, with long-horizon interruption and task completion rates entering enterprise tender documents, alongside a repricing of software outsourcing and junior research support. Within a year, outcome-based contract structures are likely to appear, and inference cost plus gross margin will become the second battleground, favouring vendors with in-house silicon. For ordinary users, the near-term effect is priority access to long-task mode for Ultra subscribers and higher API unit costs that force developers to redo their budgets. Over twelve months, the range of automation an individual developer can afford will
Among competitors, neither OpenAI nor Anthropic issued a formal response to the specification comparison; researchers at both companies steered the discussion toward agent reliability, interpretability and safety boundaries rather than context length. Developer reaction split sharply. Top threads on Hacker News and Reddit focused on whether thinking-compute billing makes Deep Think spending unpredictable, and on who absorbs the cost when a 24-hour run fails midway, though several engineers said they would accept the premium if repository-scale migration works. Enterprise buyers were more positive, with consulting and software firms describing pilot programs that put legacy system migration
Questions worth exploring:
- Whether continuous autonomous runtime will displace static leaderboards as the primary enterprise criterion for choosing a frontier model.
- How the shift from token billing to thinking-compute and outcome-based pricing will rewrite business models and margin structures across the industry.
- Whether in-house silicon gives Google a durable cost moat in the agent era, and how rivals without it can respond.
Further reading: 2026 Frontier LLM Comparison: A Practical Guide — a durable, evergreen reference. Bookmark it.
This item comes from DeepMind Blog (published 2026-10-13 16:01:52). 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.