Gemini 4 Argon Goes GA With 1M-Token Output and $2 Intro Pricing

Google returns to the frontier

After most of 2026 without a new flagship model, Google has shipped one. Gemini 4 Argon is the first model in the Gemini 4 family. Google pitches it at real-world software engineering, enterprise knowledge work in fields like law and finance, and defensive cybersecurity.1611 Google DeepMind announced it on September 30, 2026, and opened it to developers and businesses on October 13.20 Google's Cloud lifecycle documentation also lists October 13 as the release date for the stable gemini-4-argon endpoint, with retirement set for October 13, 2027 or later.2

The timing matters as much as the model. In late September, DeepMind chief Koray Kavukcuoglu said Gemini 4 had entered post-training and that Google wanted to release it "as soon as possible." He made those remarks shortly after Anthropic and OpenAI had both refreshed their flagship models.5 The Information described the coming release as a long-awaited step for a company that had fallen behind Anthropic and OpenAI in the model race.7 Earlier in the summer, Sundar Pichai said on an earnings call that Google was "a bit behind" in coding and agentic coding. Those are exactly the areas where Argon now claims its strongest results.9

What Argon actually is

The model card calls Argon a large sparse mixture-of-experts transformer. Google says it is a new model, not a fine-tune of an earlier Gemini release.12 It accepts text, images, video, audio and PDFs, and it outputs text only.13

The main technical feature is output length. Google's launch post says Argon's output limit grows to 1 million tokens, up from 64K in earlier models. The argument is that a model with room to produce hundreds of thousands of tokens in one run can work through hard problems in a single pass.11 The developer documentation is more specific. Argon has one 1,048,576-token window shared by input, thinking and output. Synchronous calls default to 65,536 output tokens, and a "long decode" mode can use whatever space is left in the window.131520 Google Cloud's version of the model lists a full 1,048,576-token maximum output on all of its API surfaces.14

Some coverage describes this as a 1M-token output ceiling, well above the 64K–256K limits common on rival models.17 That's accurate as far as it goes, but readers should keep the shared window in mind. Every token of input you send reduces how much the model can write back. In our reading, long decode is a real change, but it's an option for long-running background jobs. Ordinary API calls still default to 64K of output.

Developers moving from Gemini 3.1 Pro will hit some breaking changes. Explicit context caching isn't supported. The old thinking_budget parameter now returns an error, and the "minimal" thinking level has been removed. Low, medium and high remain, with medium as the default.16

Pricing and access

Argon launched at an introductory price of $2 per million input tokens and $10 per million output tokens, with cached input 95% cheaper.11 One tracker says the introductory price runs through November 30, 2026.20 Ars Technica also reports that the discount ends with November.4 OpenRouter labels the current rate as 50% off.19 For Gemini 4 Carbon, Google's footnote gives post-introductory rates of $4 input and $20 output per million tokens. That suggests Argon's list price is likely double what developers pay today, though this is our inference.1 Teams budgeting for next year should plan for the higher rate.

Access requires payment. Argon is available through the paid Gemini API, Google AI Studio, Gemini for business, and the Antigravity coding environment for eligible AI Ultra and Enterprise users. It isn't offered on the free tier or the Google AI Pro plan.2018

The benchmark picture: strong, not dominant

Google's model card compares Argon with OpenAI's GPT-6 Astra and Anthropic's Claude Fable 5.1 and Claude Opus 5.5, with results dated October 2026.12 Argon leads on knowledge-work tests. It scores 68.9% on the Vals Index against 67.0% for Opus 5.5. On Vals Finance Agent v2 it scores 65.4%, and on Harvey's Legal Agent Benchmark it scores 19.6%, while every competitor stays in single digits.12 It also sets a new high of 77.9% on DeepSWE v1.1, a long-horizon software engineering benchmark.11 Its long-context retrieval is a clear strength too: it scores 84.2% on GraphWalks at 256K–1M tokens, against 71.8% for GPT-6 Astra.12

It doesn't win everywhere, even on Google's own chart. GPT-6 Astra leads on FrontierSWE v2 (65.5% to 55.0%), Terminal-Bench Science (68.1% to 57.6%) and the OSWorld-2.0 computer-use subset. Claude Opus 5.5 leads Terminal-bench 4.0 by about nine points and also wins PostTrainBench.12 The coding results split depending on the benchmark. Argon is strongest on long-horizon repository work, while the Anthropic and OpenAI models still lead on terminal-heavy agent tasks.

Reports from hands-on testers point the same way. They describe Argon as more ambitious in scope and in how many features it builds. It is also rougher on visual polish, especially 3D and game graphics, than Opus 5.5 or GPT-6 Astra.17 Google's best real-world example is internal. Argon agents rewrote 32K lines of SIMD code in a Rust port of the libgav1 video decoder and made it 2.7x faster, with identical output.11

Safety: the disclosures to read closely

Argon's cybersecurity focus brings real risks, and Google says so. The model card says Google cannot rule out that Argon reaches its Critical Capability Levels for biological uplift and for cyberattack uplift. In both cases Google says it has applied safeguards that bring risk below those levels.12 On "loss of control," Argon passed 14 of 15 stealth and situational-awareness challenges, so Google also cannot rule out that alert threshold.12

The company's response is to split access. A separate gemini-4-argon-cyber-permissive model, which Google describes as being without cyber guardrails, is limited to vetted defenders in its Fairwind Program.1114 Cloud documentation lists it as allowlist-only, with no retirement date announced.142 This is a clear statement of a dual-use policy: the most capable version goes only to a small, screened group. How well that holds up will depend on vetting that outsiders can't easily check.

Carbon arrives two weeks later

Argon's run as Google's top model lasted about two weeks. On October 28, Google announced Gemini 4 Carbon, which builds on Argon and is pitched as its strongest model for cybersecurity and long-horizon work.1 Google says Carbon is first on the Artificial Analysis Intelligence Index v5.0 with a score of 59. Its cyber benchmarks were run without guardrails, and access starts with defenders in the Fairwind Program.81

Reactions to Carbon differ. Ars Technica calls it essentially a slightly better Argon. It notes that Carbon beats Argon by small margins across the board and could be a post-training upgrade. It also points out that all frontier models land within a few points of each other.4 A separate write-up stresses what's missing: no general-availability date, no context window spec, and no Carbon model card. It also notes that the safety report Google linked actually covers Argon.1 Both outlets agree that Carbon keeps Argon's $2/$10 introductory pricing.14

Why it matters

Our view is that Argon matters less for any single benchmark and more for how Google is releasing models. Google dropped its planned Gemini 3.5 Pro, a model that was reported months late after coding results fell short of internal targets.917 Since then it has moved to fast, staged releases: restricted early access for cyber defenders, then paid developers, then consumers. Argon took about two weeks to go from announcement to general availability.20 Carbon now looks set to follow the same path.4

The rest of the field isn't slowing down either. October roundups also list OpenAI's GPT-6 Pro Parallel Agents, Anthropic's preview of the background agent Claude Duet, and Meta's Muse Spark Open.3 One analysis lists GPT-6.1 Sol, Claude Opus 5.5, Gemini 4 Argon and Grok 4.7 as the leading models by late October, which it calls about two generations in six months.3 With results this close, enterprise buyers will likely care more about price, context length and access policy than about benchmark rankings. On the first two, Argon is now competitive.