Google DeepMind and Gemini 4 Argon: direction, spending and pricing

Google DeepMind is pushing Gemini toward agents, coding, science and cyber defense, and Alphabet is paying for it with sharply higher capital spending. Alphabet's second quarter 2026 filing shows $44.9 billion of capex in one quarter, up from $22.4 billion a year earlier. Gemini 4 Argon, released on 13 October 2026, carries a stated introductory price of $2 per million input tokens.

This post walks through what Google has said in primary sources, what is a vendor claim, and what is still unknown. It is not investment advice and it does not predict what Google will do next.

Key numbers at a glance

All figures come from Alphabet's own filing or Google's own pages unless noted. Benchmark scores are reported by Google and have not been checked by an independent party in this post.

• Alphabet revenue, quarter ended 30 June 2026: $119.796 billion, up 24 percent. Source: Alphabet earnings release, 22 July 2026.
• Google Cloud revenue, same quarter: $24.768 billion, up 82 percent from $13.624 billion. Source: same release.
• Google Cloud operating income, same quarter: $8.814 billion, against $2.826 billion a year earlier. Source: same release.
• Capital expenditure, same quarter: $44.924 billion, against $22.446 billion a year earlier. Source: same release.
• Capital expenditure, trailing twelve months: $132.402 billion. Source: same release.
• Gemini app monthly active users: 950 million, per the chief executive's statement in the release.
• Gemini API volume: 22 billion tokens per minute, per the same statement.
• Gemini Enterprise adoption: "nearly 90%" of the Fortune 100, per the same statement.
• Employees: 198,933, up from 187,103 a year earlier.
• Gemini 4 Argon introductory price: $2 per million input tokens and $10 per million output tokens until 30 November 2026. Source: Google blog, 13 October 2026.
• Gemini 4 Argon standard price after that: $4 per million input tokens and $20 per million output tokens.
• Cached input discount: 95 percent off the input price.
• First announcement of Argon: 30 September 2026, to trusted cyber defenders first.
• Broad release: 13 October 2026, through the API, Google AI Studio, the Gemini app and Search AI Mode.

What Google actually released

Google released Gemini 4 Argon in two steps. The first post, dated 30 September 2026 and written by Koray Kavukcuoglu (Senior Vice President at Google DeepMind and Google's Chief AI Architect), said Argon would roll out first to trusted cyber defenders through a group called the Fairwind Program. Google said it would widen access gradually. The post also said Google was taking part in the United States government's voluntary pre-release model access process, where a model is made available to government evaluators before it ships widely.

The second post, dated 13 October 2026, announced broad availability. Developers and enterprises can reach the model through the Gemini API in Google AI Studio. Subscribers to Google AI Ultra can use it in Google Antigravity, which is Google's agent based coding tool. Subscribers to Google AI Ultra and Pro can use it in the Gemini app and in AI Mode in Search.

The Google DeepMind models page lists the wider family as of 19 October 2026. Gemini 4 Argon sits at the top, described as "our next era of frontier intelligence." Beside it are Gemini 3.8 Flash for agentic tasks at scale, Gemini Omni and Omni 1.1 Flash for creating media starting with video, Gemini Audio for real time speech, Gemini Robotics 2, Gemini Embedding, and the image model branded Nano Banana. Separately, the open Gemma family continues, and the page lists EmbeddingGemma 2 as an October 2026 release.

That list tells you something about direction even before you read the numbers. Google is not shipping one model. It is shipping a ladder: a top model for hard problems, a cheaper fast model for volume work, and specialist models for audio, video, images, embeddings and robots.

The direction in plain terms

Reading the two launch posts and the models page together, four themes show up.

Agents and coding. Most of the engineering results Google quotes are about software work. The 13 October post says agentic code submissions rose by more than 35 percent in Antigravity and that code rollbacks fell by nearly a third. It describes a refactor of roughly 6 million lines of a legacy folder inside a codebase of about 40 million lines in Chrome that took weeks instead of months. The 30 September post says Argon set a new best score on DeepSWE v1.1, a software engineering benchmark, at 77.9 percent, and described work such as replacing 32,000 lines of hand tuned code in a video decoder library with a Rust version that Google says runs 2.7 times faster than an existing Rust port. All of these are Google statements about Google's own work.

Knowledge work. Google says Argon ranks first on the Vals Index, a benchmark of economic work in finance, coding, legal and tax. The models page puts the score at 68.9 percent. The company also cites a partner, Box, reporting a 14 percent accuracy improvement over Gemini 3.1 Pro. Treat these as marketing claims backed by outside test providers. They are useful as a signal of where Google wants the model to be used, which is the office, not just the lab.

Science and robotics. The 13 October post lists work on a quantum error correction decoder, on enzyme design where five new designs showed confirmed activity in the lab, and on fusion mathematics. It also describes a demonstration in which Argon built a simulation from a video of a cornhole throw and an Apptronik humanoid robot then landed seven shots in a row. Google offers 10,000 scientists the highest Antigravity usage tier free for six months, subject to eligibility and capacity. A separate release, Gemini 4 Argon for Robotics, appears on the models page for October 2026; this post does not cover it.

Cyber defense. This is the theme that matters most to readers of this blog. Google says Argon can find, validate and patch vulnerabilities on its own. Notably, the 30 September post says the model will be released without cyber guardrails for trusted defenders and for Google's internal teams, while the general release keeps its protections. The post also reports a score of 68 percent on CWE-bench v1, a benchmark for fixing security flaws, tied for first by Google's account. On the models page the tie is with a model from another lab. Alphabet's own release highlights Gemini 3.5 Flash Cyber as a cost efficient model at the frontier.

A reader who builds software should note the pattern. Frontier labs are creating tiered access, where verified defenders get more capable and less restricted tools than the public. Whether that is a good model is a policy debate, and this post takes no side. It is simply a fact that Google has described it publicly.

Safety measures Google describes

Both launch posts say Argon was developed under Google's Frontier Safety Framework. The 13 October post lists several measures and results, all reported by Google:

• A multi layer defense for jailbreak resistance, monitoring and enforcement.
• External red teaming for biological risks that, per Google, could unlock less than 1 percent of the biology sub tasks tested.
• A drop of nearly half in violations on an adversarial test set compared with earlier Gemini models.
• Shortcut attempts on coding benchmarks falling to 2 percent. A shortcut attempt is when a model tries to game a test rather than solve the task.
• An automated monitor that reads the agent's reasoning and actions in real time and can stop execution.
• Attack success in the Gray Swan indirect prompt injection test falling to 0.7 percent across 15 adaptive attempts. Indirect prompt injection means hiding instructions inside content the model reads, such as a web page or document, so that it follows the attacker instead of the user.

The 13 October post says a model card is linked from it. The models page, however, links an evaluation methodology page and a cyber page, not a model card. If you need the model card for a risk review, check Google's site directly.

Pricing and what the numbers mean

Pricing on AI models is quoted per million tokens. A token is a small piece of text, roughly three quarters of a word in English. You pay for tokens you send (input) and tokens the model returns (output).

Google's stated prices for Argon are $2 per million input tokens and $10 per million output tokens as an introductory price, rising to $4 and $20 after 30 November 2026. Cached input, which is text you have sent before and that the service has stored for reuse, costs 95 percent less than normal input. Caching matters for agent work because an agent resends the same instructions and files again and again.

For context, other labs publish different price points in the same period. Mistral lists $1.36 per million input tokens and $4.18 per million output tokens for its new large model preview, and Moonshot lists $3.00 per million cache miss input tokens and $15.00 per million output tokens for Kimi K3. DeepSeek's page lists a much lower tier for its flash model. These models differ in size, capability and test results, so a price list alone does not say which is cheaper for your task. Compare on your own workload.

One point where Google's pages disagree with each other: the 30 September post says the output limit was raised to 1 million tokens from 64,000 and does not state an input limit. The 13 October post describes a "1M unified input-output token limit via the Gemini API." Those are not the same description, and the models page again speaks of a 1 million token output window. If your design depends on the context size, test it and read the API documentation.

The money behind it

Gemini is expensive to build and serve, and Alphabet's filing shows it. Capital expenditure, or capex, is money spent on long lived assets such as land, buildings, servers and chips. In the quarter ended 30 June 2026 Alphabet's purchases of property and equipment were $44.924 billion, more than double the $22.446 billion of a year earlier. The trailing twelve month total was $132.402 billion.

Revenue grew 24 percent to $119.796 billion, so spending grew far faster than sales. Google Cloud, the unit that rents computing and sells AI services to businesses, grew 82 percent to $24.768 billion and its operating income rose to $8.814 billion. In other words, the cloud unit is where the AI demand shows up in the numbers, and it is profitable on the reported basis.

Two things are not in the release. First, it does not disclose Google Cloud's backlog, the value of signed contracts not yet recognized as revenue. Second, the chief executive quotes do not mention Google's custom chips, called TPUs (tensor processing units). Neither launch post for Argon mentions TPUs either. The only hardware named in the 13 October post is an A100 GPU, used in a test of how fast the small open Gemma model can run. So these primary sources do not say what Argon was trained on, and Google has not stated it in the launch material. For chip context, see the guide to AI chip market share.

For company by company capex guidance, including Alphabet's full year range as it was reported after the July call, see the 2026 capex scoreboard. That post draws on call summaries, so treat the guidance numbers there as reported, not as figures from this release. For how the same kind of spending looks at AI labs, see how Anthropic's compute commitments work and what OpenAI has planned for compute.

What is firm and what is uncertain

Firm, because it comes from Alphabet's filing or Google's own pages:

• The quarterly revenue, cloud revenue, operating income and capex figures above.
• The release dates of 30 September and 13 October 2026 for Argon.
• The published price list and the 30 November change date.
• Availability through AI Studio, the Gemini app, AI Mode and Antigravity.

Uncertain or only vendor reported:

• Benchmark scores. Every score in this post is Google's own claim or a claim from a partner Google cites. Methodology pages exist, but no independent audit of them has been published.
• Internal results. The Chrome refactor, the memory savings and the code migrations are described without independent confirmation.
• Context size. The pages describe it differently, as shown above.
• Training hardware and training cost. Google has not stated either in the launch posts or the earnings release.
• Usage numbers. The 950 million monthly users and 22 billion tokens per minute are the chief executive's statements, and the release does not define "active".
• Whether the cyber variant stays limited. Google says it will be available without guardrails to trusted defenders. How that group is chosen and how it will change is not detailed.

What to watch next

• Alphabet's third quarter results. They are not out as of 19 October 2026. Check Alphabet's investor relations page for the date. Look at the capex line, the cloud growth rate and whether the company gives a figure for 2027.
• The end of the introductory price on 30 November 2026. The stated price doubles. Watch whether Google extends it, and whether other labs move.
• Independent testing. Third party evaluators will publish their own results for Argon. Compare them with Google's claims.
• The Fairwind Program. The post says it expanded to hundreds of additional partners. Watch for details on who is in it and on what terms.
• The open Gemma family. Google keeps shipping small open models next to its closed flagship. That is a different strategy from a lab that releases only closed models. For definitions, see open weights versus open source.

Quick check

If you want to know which open source components your own software pulls in, Safeguard offers SBOM and dependency scanning, and a free scan is available at https://gold.safeguard.sh/scan.

Sources

• Alphabet, second quarter 2026 earnings release (SEC Exhibit 99.1), filed 22 July 2026: https://www.sec.gov/Archives/edgar/data/0001652044/000165204426000066/googexhibit991q22026.htm
• Google, "Gemini 4 Argon: Built to solve your hardest problems," 13 October 2026: https://blog.google/innovation-and-ai/models-and-research/gemini-models/introducing-gemini-4-argon/
• Google, "Gemini 4 Argon: our next era of frontier intelligence," 30 September 2026: https://blog.google/innovation-and-ai/models-and-research/gemini-models/gemini-4-argon/
• Google DeepMind, Gemini models page, checked 19 October 2026: https://deepmind.google/models/
• Google DeepMind, Gemini 4 Argon page, checked 19 October 2026: https://deepmind.google/models/gemini/
• Mistral, "Introducing Mistral Large 4," 6 October 2026: https://mistral.ai/news/mistral-large-4/
• Moonshot AI, Kimi K3 post: https://www.kimi.ai/blog/kimi-k3
• DeepSeek API pricing page, checked 19 October 2026: https://api-docs.deepseek.com/quick_start/pricing

Last reviewed: 19 October 2026.