Google's first Gemini 4 model, shipped after a cancelled flagship, a leadership shake-up and seven months without a Pro-class release. It tops the Vals Index and undercuts GPT-6 Astra by 80% on price — for six more weeks.
On professional work, yes. On everything else, no — and the price advantage has a hard expiry date of December 1, 2026. Gemini 4 Argon ranks #1 on the Vals Index at 68.90%, ahead of Claude Sonnet 5.5 (67.04%), Claude Opus 5.5 (66.97%) and Claude Fable 5.1 (65.83%) — the first Gemini model ever to top that index. But on the independent Artificial Analysis Intelligence Index v4.3.2 it scores 53 (high reasoning), tying GPT-6 Astra and Fable 5.1 while trailing Claude Opus 5.5 at 58 and Sonnet 5.5 at 56. Google's own 19-row table has it winning 13, tying 1 and losing 5 — and the losses are the agentic-coding rows: FrontierSWE v2 (55.0 vs Astra's 65.5) and Terminal-Bench 4.0 (57.4 vs Opus 5.5's 66.4). The price is the real event: ~$2 per 1M input / ~$10 per 1M output with cached input 95% off through November 30, 2026, roughly one-fifth of GPT-6 Astra's and Claude Fable 5.1's $10/$50 and half of Claude Opus 5.5's and Meta Muse Wonder's $4/$20. It runs the AA index at $1.99 per task versus $3.26 for Astra. Then, on December 1, 2026, the price doubles to $4/$20 — exact parity with Opus 5.5 and Muse Wonder, with none of Opus's lead. Two more catches: Antigravity and AI Studio require AI Ultra at $99.99/month (AI Pro at $19.99 only reaches Argon in the Gemini app), and Bloomberg reported anonymous Google employees saying it underperforms on real coding work, singling out front-end design, which Google calls inaccurate. Verdict: 4.0/5 — the best model you can buy for legal, finance and long-document work at the cheapest frontier-class token price on the market, until the calendar catches up.
Gemini 4 Argon (gemini-4-argon, listed as stable) is the first frontier model in Google's Gemini 4 generation. It was announced on September 30, 2026 by Koray Kavukcuoglu, SVP of Google DeepMind and Chief AI Architect, to a deliberately tiny audience: vetted cyber defenders inside Google's Fairwind Program, plus Google's own engineers. It reached developers, enterprises, scientists and consumers on October 13, 2026 — through the paid Gemini API in AI Studio and Android Studio, in Gemini for business, in Google Antigravity for AI Ultra subscribers, and in the Gemini app and AI Mode in Search for AI Pro and Ultra subscribers. Google has said plainly that Argon is "just the beginning" of a larger Gemini 4 family.
The context that matters is what it replaced. Google's previous flagship promise, Gemini 3.5 Pro, was announced by Sundar Pichai at I/O on May 19, 2026 with a "give us until next month" that slid past June, past July and eventually into oblivion — Google has since told Reuters it no longer plans to release it. Axios reported in July that low morale inside DeepMind was part of the delay. In the middle of that, on August 5, 2026, Demis Hassabis stepped down as CEO of Google DeepMind to become the lab's chairman and Alphabet's first chief scientist, with Kavukcuoglu taking the day-to-day lead, and Jeff Dean left after 27 years to start his own company. Argon is the first model shipped under that new arrangement, which is why its launch reads less like a routine release and more like a company answering a question about itself.
Technically it is a vision-language reasoning model, and the architecture, parameter count and training details are undisclosed. What is published: a unified 1,048,576-token context window that input, thinking and output all draw from, accepting text, images, video, audio and PDF in and returning text out, with three thinking levels (low, medium, high) and a March 2026 knowledge cutoff. It is proprietary, with no open-weights variant announced. For how it sits against the wider market, see our LLM API cost guide.
Google's framing for Argon is "deep work" — long-horizon reasoning and multi-step agentic tasks rather than fast chat. Four things carry that claim.
There are three different Argon stories in circulation, and they do not agree. Google's table, the independent evaluators, and people who actually used it.
Google published 19 rows covering Argon, GPT-6 Astra, Claude Fable 5.1 and Claude Opus 5.5. Argon leads outright in 13 rows against the whole field, ties one (CWE-bench v1, a three-way tie with Astra and Grok 4.7) and loses five. The headline-count argument you will see on social media — 12, 13, 14 or 15 wins — comes from which rivals you count. Nine rows are public leaderboards scored for every model, which is the fairest subset; Argon leads seven of those.
| Benchmark | Gemini 4 Argon | GPT-6 Astra | Claude Opus 5.5 | Who measured it |
|---|---|---|---|---|
| Vals Index | 68.9% | 63.1% | 67.0% | Vals AI |
| AutomationBench | 51.3% | 41.4% | 42.5% | Zapier |
| Vals Finance Agent v2 | 65.4% | 53.5% | 58.6% | Vals AI |
| Harvey Legal Agent | 19.6% | 5.4% | 3.8% | Vals AI |
| DeepSWE v1.1 | 77.9% | 74.1% | 74.2% | Mixed harnesses |
| FrontierSWE v2 | 55.0% | 65.5% | 62.3% | Proximal leaderboard |
| Terminal-Bench 4.0 | 57.4% | 58.2% | 66.4% | Mixed |
| OSWorld-2.0 (computer use) | 69.2% | 72.6% | n/r | Mixed |
| GraphWalks 256K–1M | 84.2% | 71.8% | 66.8% | Google, all models |
| LVBench (long video) | 91.7% | 87.5% | 83.7% | Google, all models |
| CWE-bench v1 (cyber) | 68.0% | 68.0% | 67.0% | Public leaderboard |
Two clarifications that change how you read that table. First, the losses cluster in terminal-heavy agentic coding: FrontierSWE v2 and Terminal-Bench 4.0 are both last-place finishes on leaderboards that scored every model, and if your workload is repository-scale work in a terminal, those two rows matter more than the 13 wins. Second, the methodology is not uniform even inside Google's table: Argon ran at its highest thinking setting against rivals at maximum reasoning, LVBench gave each model a different number of video frames (Argon 1 FPS, Astra 800 frames, Opus 600), and the OSWorld figure is Argon's best of three runs rather than an average. Google also discloses that Argon ran Agent's Last Exam and OSWorld with safety filters on, where flagged responses returned empty strings — a handicap, if anything.
The honest synthesis: Argon is the best model in the world at reading and reasoning over very long professional documents, the cheapest frontier-class model on the market during the promotional window, and not the best model for terminal-heavy agentic coding. If your work looks like legal research, financial analysis, long-context synthesis or 3D/spatial reasoning, Argon is a genuine first choice. If your work looks like a repository-scale terminal agent, Terminal-Bench and FrontierSWE both say Opus 5.5 or Astra, and the GPT-6 Astra review covers that case in detail.
Google has published two price books in one launch, and the second one arrives on a schedule. Every figure below is approximate, in US dollars, checked against Google's own launch post, the Gemini API pricing documentation and launch-day coverage on October 14, 2026 — always confirm current pricing on Google's official pages before you build or subscribe, because the promotional rate has a stated end date and regional variation is normal.
Gemini API — the promotional rate, through November 30, 2026 (gemini-4-argon)
~$2.00 per 1M input tokens · ~$10.00 per 1M output tokens · ~$0.10 per 1M cached input (95% off) · ~$1.00 / $5.00 for Batch or Flex traffic (50% off)
Thinking tokens bill at the output rate. Caching is implicit only — Argon does not support explicit context caching, which matters if you were counting on caching a large fixed corpus. There is no free tier for this model on the API.
Gemini API — the standard rate, from December 1, 2026
~$4.00 per 1M input tokens · ~$20.00 per 1M output tokens · ~$0.20 per 1M cached input · ~$2.00 / $10.00 for Batch or Flex traffic
Google's footnote is explicit: "after the introductory period expires on November 30, 2026, the price of $4 per 1M input tokens and $20 per 1M output tokens will apply." That is exactly double, and it lands Argon on the same $4/$20 rate card as Claude Opus 5.5 and Meta Muse Wonder while still trailing Opus 5.5 on the AA Intelligence Index by five points. It remains 60% cheaper than the $10/$50 cards of GPT-6 Astra and Claude Fable 5.1.
Consumer and subscription routes
Free tier: no Argon — Google has restricted free users to Flash-Lite · Google AI Plus: not listed as including Argon · Google AI Pro ~$19.99/month: Argon in the Gemini app and AI Mode in Search (from October 13, 2026), plus Argon code generation in Canvas mode · Google AI Ultra ~$99.99/month (5x Pro) or ~$199.99/month (20x Pro): everything in Pro, plus Argon in Antigravity and AI Studio
Google has declined to say when, or whether, Argon will reach Antigravity on cheaper tiers. If you need Argon as an agent rather than a chat model, the subscription floor is $99.99/month.
The cost-per-task trap. Comparing rate cards understates what the December change does to a bill. Artificial Analysis measured Argon at $1.99 per index task during the promotion and $3.98 at standard rates — the same model, the same prompts, double the money — because Argon generates about 62K output tokens per task against Astra's 27K. Argon's advantage came from price, not from efficiency, and it has now been scheduled away. If you are evaluating Argon today, model your unit economics at $4/$20 and treat anything measured at $2/$10 as a six-week promotion. Our LLM API cost guide has the wider table.
Argon is sold through four different doors with four different limits. Here is the practical version.
| Route | What you get | What it costs | Who it is for |
|---|---|---|---|
| Paid Gemini API | Argon with full control of thinking level and 1M unified context; Long Decode for million-token outputs | ~$2/$10 per 1M until Nov 30, then ~$4/$20. No free tier. | Developers with their own harness, pipeline or product. The only route with per-token control and no subscription floor. |
| Google Antigravity | The full agentic platform: slash commands, multi-agent loops, custom skills | AI Ultra only — from ~$99.99/month | Teams that want Google's agent harness rather than their own. Pro subscribers are excluded with no announced date. |
| Gemini app + AI Mode in Search | Argon as a chat and analysis model, plus Canvas coding | AI Pro ~$19.99/month or Ultra | Professionals who want Argon's document and reasoning strength without building anything. |
| Fairwind Program | Argon without cyber guardrails, standalone or inside CodeMender | Vetted access, no public price | Governments, national cyber authorities, critical infrastructure operators and security vendors that pass vetting. |
Two consequences fall out of that. First, the cheapest path to Argon for a professional is AI Pro at $19.99, not the API — if your use is analysis rather than automation, a Pro subscription buys more Argon than $19.99 of tokens usually will. Second, the $99.99 Antigravity requirement is the real gate for agentic use, and it is the same platform whose forced auto-compaction and slow usage reporting developers complain about most. Google also has two science programs running: 10,000 scientists can apply for Antigravity's highest usage tier free for six months, and Fairwind expansion opened to hundreds of additional partners on launch day. Both are subject to eligibility, capacity and selection — and neither has published selection criteria.
| Model | Price (per 1M in / out) | Vals Index | AA Index | Best for |
|---|---|---|---|---|
| Gemini 4 Argon | ~$2 / ~$10 until Nov 30, then ~$4 / ~$20 | 68.90% | 53 | Long-document legal, finance and tax work; 1M-token context; cheapest frontier tokens while the promotion runs |
| Claude Sonnet 5.5 | ~$2 / ~$10 | 67.04% | 56 | Vibe coding and code migration, at Argon's post-December rate today |
| Claude Opus 5.5 | ~$4 / ~$20 | 66.97% | 58 | Best overall intelligence and terminal-heavy agentic coding |
| Claude Fable 5.1 | ~$10 / ~$50 | 65.83% | 53 | Creative work, tax agent tasks and design-heavy pipelines |
| Meta Muse Wonder | ~$4 / ~$20 | 64.57% (Full) | 52 (Full: 56) | Cross-conversation memory and long-horizon agent work |
| GPT-6 Astra | ~$10 / ~$50 | 63.1% | 53 | Repository-scale agentic coding and computer use |
| OpenAI Dots | No standalone price (~$100/mo Pro gate) | — | 53 (via Astra) | Always-on agents with their own cloud computer |
No. There is no free tier for Gemini 4 Argon anywhere. On the developer side Google states there is no free tier for the model on the Gemini API, so every call is billed, and the cheapest published rate is two dollars per million input tokens and ten dollars per million output tokens through November 30, 2026. On the consumer side the free Gemini app does not get Argon at all: since Google restricted free users to Flash-Lite, free accounts are left with the lighter models, and the cheaper AI Plus tier is not listed as including Argon either. The lowest monthly door is Google AI Pro at about nineteen dollars and ninety nine cents per month, which includes Argon in the Gemini app and in AI Mode in Search from October 13, 2026. Antigravity and AI Studio need Google AI Ultra at ninety nine dollars and ninety nine cents per month or more. Prices are in US dollars, vary by region, and may change, so confirm the current figure on Google's own pages before you buy.
There are two price books and they expire. Through the Gemini API, until November 30, 2026, Argon costs about two dollars per million input tokens and about ten dollars per million output tokens, with cached input at about ten cents per million, a ninety five percent discount, and batch or Flex traffic at half price, roughly one dollar and five dollars. Thinking tokens bill at the output rate, and Argon supports implicit caching only. From December 1, 2026 the standard rate applies: about four dollars per million input tokens, about twenty dollars per million output tokens, about twenty cents per million cached input, and about two dollars and ten dollars for batch or Flex. Subscription routes are separate: Google AI Pro at about nineteen dollars and ninety nine cents per month, and Google AI Ultra at about ninety nine dollars and ninety nine cents per month or about one hundred and ninety nine dollars and ninety nine cents for the twenty times tier. Confirm current numbers on Google's own pages, because these are launch day prices with regional variation.
It doubles. Google's own footnote says that after the introductory period expires on November 30, 2026 the price of four dollars per one million input tokens and twenty dollars per one million output tokens will apply, which is exactly double the launch rate. Cached input goes from about ten cents to about twenty cents per million tokens and batch or Flex pricing goes from about one dollar and five dollars to about two dollars and ten dollars. That matters more than a normal price rise because of where it lands. At four and twenty dollars, Argon costs exactly the same per token as Claude Opus 5.5 and Meta Muse Wonder, the two models it is measured against most often, and Opus 5.5 scores 58 against Argon's 53 on the Artificial Analysis Intelligence Index. Argon still undercuts GPT-6 Astra and Claude Fable 5.1, which both sit at ten and fifty dollars, by sixty percent rather than eighty percent. If you are building a workflow around Argon, run your own unit economics at the December rate before you commit, and treat any saving measured at two dollars and ten dollars as temporary.
It depends which benchmark you trust, and the two headline sources disagree. On the Vals Index, which weights finance, coding, legal and tax tasks by each sector's share of United States GDP, Gemini 4 Argon ranks first at 68.90 percent, ahead of Claude Sonnet 5.5 at 67.04, Claude Opus 5.5 at 66.97 and Claude Fable 5.1 at 65.83, and it is the first Gemini model to top that index. On the Artificial Analysis Intelligence Index it scores 53 at high reasoning, level with GPT-6 Astra at max effort and Claude Fable 5.1, but behind Claude Opus 5.5 at 58 and Claude Sonnet 5.5 at 56. Google's own nineteen row table has Argon leading 13 rows outright against the whole field, tying one and losing five, including FrontierSWE v2 at 55.0 against Astra's 65.5, Terminal-Bench 4.0 at 57.4 against Opus 5.5's 66.4 and OSWorld-2.0 computer use at 69.2 against Astra's 72.6. Bloomberg also reported anonymous Google employees saying the model does less well on real coding work than its benchmarks suggest, which Google disputed. The honest summary is that Argon is the best model available for professional document, legal, finance and long-context work, and not the best model for terminal-heavy agentic coding.
For the coding platforms, yes. In Google Antigravity and AI Studio, Argon is restricted to Google AI Ultra subscribers, who pay about ninety nine dollars and ninety nine cents per month for the five times tier or about one hundred and ninety nine dollars and ninety nine cents for twenty times. Google has not said when or whether lower tiers will get Antigravity access. Google AI Pro subscribers at about nineteen dollars and ninety nine cents per month do get Argon in the Gemini app and in AI Mode in Search, and can generate code with it in Canvas mode, but not through Antigravity's multi threaded agent workflows, which Google appears to treat as too compute intensive for the cheaper plan. Developers do not need a subscription at all: the paid Gemini API tier serves Argon in AI Studio and Android Studio with the model identifier gemini-4-argon, billed per token. The free tier and the cheaper AI Plus plan are excluded from every route.
Fairwind is Google's limited access program for cyber defenders, launched on September 2, 2026 with more than 650 participating partners worldwide and expanded to hundreds of additional partners on October 13, 2026. What made it contentious is one sentence in Google's announcement: for trusted defenders and Google's own internal teams, Google released Argon without cyber guardrails, while everyone else gets a build designed to refuse requests that would enable cyber attacks. Inside a partner organisation only security, incident response and penetration testing staff using phishing resistant multi factor authentication may use the unguardrailed build, and every use is tracked. Fairwind also runs through the AI security firm Wiz, whose Scan for Good initiative used Argon to find a vulnerability exposing personal data in healthcare software used by hospitals worldwide. The criticism is that the guardrail has become a service tier rather than a safety mechanism: Google publishes no membership criteria, no application route and no appeals process, and the model family sits in the same line as a Google system that accessed protected systems at three real companies during a May 2026 evaluation. If you are a defender, Fairwind is the only way to get the unrestricted build; if you are anyone else, treat the guardrailed version as the product.
Honest accounting software reviews: Wave, QuickBooks, Xero & free options.
🎮 Gaming Gear ReviewsEsports hardware decoded: gaming mice, keyboards, headsets & budget picks.
📄 PDF Merge OnlineCombine multiple PDF files into one — free, no upload, files stay on your device.
📷 ID Photo StudioPassport & visa photos with AI background changer, compressor & converter.
🧮 HK Salary Tax CalculatorEstimate Hong Kong salaries tax for 2026/27 in seconds — see what you owe before you file your tax return.
🎁 HK Benefits Quick CheckGovernment & benefits eligibility checker — OAA, CSSA, housing & more.
📅 HK Calendar 香港月曆Public holidays, lunar dates & solar terms for 2020-2036 — bilingual, free, no sign-up.
★ Our own product 🏠 Home Expense BookkeepingFamily bookkeeping: income, spending, net worth and monthly bills. Data stays on your device. HK$98 one-time.