Google’s Gemini 4 Argon Goes Public, Priced to Win Back the Frontier
Key points
- Gemini 4 Argon is now available globally to paying users via the API, the Gemini app, AI Mode in Search and Antigravity, with an introductory price of $2 per million input tokens and $10 per million output tokens 1.
- Google says Argon tops 14 of 18 results in its own comparison table, but the independent Artificial Analysis Intelligence Index places it at about 53, roughly tied with GPT-6 Astra and behind Claude Opus 5.5 6.
- Internally, Argon helped teams raise agentic code submissions by more than 35% and cut rollbacks by nearly a third, and it solved all 12 simulated end-to-end cyberattack scenarios in Google's safety tests 1 3.
- The introductory API price doubles to $4 and $20 after November 30, 2026, and there is no free tier, which critics say limits mainstream appeal 12 7.
A Frontier Comeback Bid
Google has made Gemini 4 Argon available worldwide to developers, enterprises and paying subscribers, ending a limited preview that began with trusted cyber defenders in its Fairwind Program 1 6. It is the first model in the Gemini 4 family, and Gemini product lead Tulsee Doshi called it a “powerhouse” and the first of several models to come 4.
The timing matters. Argon is Google’s first major frontier release since Gemini 3 in late 2025, and it arrived after Gemini 3.5 Pro was shelved and while rivals shipped steadily 5 4. It also landed a day after Meta announced its own frontier model, Muse Wonder 5. Gizmodo noted that the launch was dampened by reports of employee skepticism about coding performance and by the departure of senior researchers 7.
Where You Can Use It
Argon is available through the API in AI Studio and the Gemini platform, to Google AI Pro and Ultra subscribers in the Gemini app and AI Mode, and to Ultra subscribers in Antigravity, Google’s agentic coding tool 1. The Gemini app gets a macOS integration of Antigravity’s coding interface, a chat-to-tasks toggle for work that continues while you are away, and connections to more than 40 apps including Instacart, OpenTable, Adobe and Shopify 2.
Google is also offering 10,000 scientists the top Antigravity usage tier free for six months, and expanding Fairwind to hundreds more partners, who get the model without the usual cyber guardrails 10 12. For developers, the model has a unified limit of 1,048,576 tokens covering input and output, three thinking levels and automatic long-output continuation in the SDK 13.
Free users get nothing, and Pro costs $19.99 a month 6 10.
Price Versus Performance
Google lists $2 and $10 per million tokens, with cached input at 95% off 1. Mashable compared that with $4 and $20 for Claude Opus 5.5 and $10 and $50 for GPT-6 Astra 6. After November 30, Argon’s rates double to the same level as Opus 5.5 12 10.
Benchmark claims deserve care. Google says Argon leads the Vals Index and tops 14 of 18 comparisons, including 51.3% on AutomationBench and 77.9% on DeepSWE v1.1, but it trails on some coding tests such as Terminal-Bench 4.0 6. Mashable also found that some competitor figures in Google’s table differ from what Anthropic reported 6. On the independent Artificial Analysis index, Argon scores 52.6 to 53, a 23-point jump over Gemini 3.1 Pro but sixth overall, with a 0.1-point gap to GPT-6 Astra 10 6. Per task, it costs about $1.99 at launch pricing and $3.98 afterward, against $5.98 for Opus 5.5 10.
Real Work, Inside Google
Google leaned on internal use. In Chrome, Argon helped separate webpage handling from a roughly 6 million line legacy folder, turning months of work into weeks 1. It sped up Gemma inference 2.85x on an A100, and cut the CPU footprint of YouTube’s comments frontend in half 1 12.
The most detailed account is the ADK run. Given a one-paragraph goal and team guidelines in .agents/skills, Argon worked unattended for 11 hours on a roughly 205,000-line Python codebase. Six optimizations were proposed, two shipped, and maintainers rejected four. The shipped changes cut in-memory session prep time by about 40% in long sessions and sped up cold imports by 21% 9. An earlier run without guidelines missed team conventions, which is a useful reminder that humans still review the output 9.
Robotics and science claims are bigger. Argon scored 87% on the ERQA embodied reasoning benchmark, and in a demonstration it helped teach a humanoid to play cornhole by watching video 8 7. Google also reports gains in quantum error correction decoding and quantum circuit cost 1.
Safety and Open Questions
Because Argon showed elevated cyber and biology capability, Google applied its frontier safeguards. The model solved all 12 simulated cyberattack scenarios, up from five previously 3. Protections include input screening, activation probes that read internal signals, and continuous monitoring, and external bio red teaming unlocked under 1% of sub-tasks 3. Shortcut attempts on coding benchmarks fell to 2%, and attacks on Gray Swan’s prompt injection test succeeded 0.7% of the time 3. Google also says it does not penalize chain-of-thought during training, because an accidental penalty made silent shortcuts more than triple 3.
These figures are Google’s own. The Next Web noted that Antigravity requires an Ultra subscription, which puts the tool behind the headline results out of reach of the $19.99 tier 11. Whether Argon wins everyday users will depend on independent testing, and on whether Google eventually offers a free tier.
Companies mentioned: Google (GOOGL $347.91 ▼1.7%) • Meta (META $734.71 ▼3.6%) • Adobe (ADBE $237.51 ▲0.5%) • Shopify (SHOP $171.41 ▼2.9%) • Anthropic
Primary sources
Gemini 4 Argon: Built to solve your hardest problems (blog.google) – Google's primary announcement of Gemini 4 Argon: availability across the API, Gemini app, AI Mode and Antigravity, introductory pricing, a #1 Vals Index ranking, and internal case studies in Chrome, ADK, xTB, Gemma, quantum computing and cybersecurity with Wiz. The text cuts off during the nuclear fusion section. Published on blog.google.
Gemini 4 Argon comes to the Gemini app. (blog.google) – Google's post on Argon in the Gemini app: more direct answers, less hedging, Antigravity's coding interface in the macOS app, a chat-to-tasks toggle, and connections to more than 40 apps. Available to AI Pro and Ultra subscribers. Published on blog.google.
Further sources
Developing Gemini 4 Argon responsibly and safely (deepmind.google) – Google DeepMind's safety write-up. Covers the Fairwind early access, Wiz's hospital-software finding, 12 of 12 simulated cyberattack scenarios solved, layered real-time protections, a misalignment monitor, a 2% shortcut rate, the chain-of-thought training rule, and 0.7% prompt injection success. Points to a full Safety and Security Report and model card. Published on deepmind.google.Google says Gemini 4 Argon is a 'powerhouse' AI model — and it's rolling out now (businessinsider.com) – Business Insider's report quoting Tulsee Doshi, who calls Argon a powerhouse and the first of a family. It covers paid-only access starting at $19.99 a month, the cornhole robot, the chain-of-thought monitoring and competitive context including Meta's Muse Wonder and GPT-6. Also notes Google skipped Gemini 3.5 Pro.
Google finally releases Gemini 4 Argon (theverge.com) – The Verge's account of the launch. It notes Google's claims against GPT-6 Astra and Claude models, internal uses at YouTube and Wing, the competitive introductory price, and the long gap since Gemini 3 and internal shakeups.
Google releases Gemini 4 Argon, its new frontier model: Benchmarks, how to try it, and pricing (mashable.com) – Mashable's detailed explainer. It covers pricing comparisons, Google's 18-result benchmark table (top on 14), mixed coding results, discrepancies with Anthropic's reported figures, the independent Artificial Analysis score of 53 with cost per task, and safety claims including 0.7% prompt injection success. Includes an ownership disclosure about Ziff Davis.
Google’s Gemini Argon 4 Is Now Available Worldwide (gizmodo.com) – Gizmodo's critical take. It notes the pricing edge over GPT-6 Astra, employee skepticism, the delay of Gemini 3.5 Pro, Jeff Dean's departure, the robot cornhole demo, and the lack of a free tier hurting mainstream appeal.
Gemini 4 Argon for Robotics (developers.googleblog.com) – Google developers blog on robotics. Argon scores 87% on the ERQA embodied reasoning benchmark, with example prompts for controlling a Spot quadruped, a robot arm, steering a world model and building digital twins in MuJoCo. Text is truncated.
Optimizing a 200,000-line Python codebase overnight with Gemini 4 Argon (developers.googleblog.com) – Google developers blog deep dive on the overnight ADK optimization run: about 205,000 lines, an 11-hour run, six proposed optimizations of which two shipped and four were rejected in review, with roughly 40% faster long sessions and 21% faster imports. Compares an unguided first round with a guideline-driven second round. Text is truncated.
Google Pushes Gemini 4 Argon Into the Market With Steep Discounts (trendingtopics.eu) – Trending Topics breakdown of who gets Argon, pricing in dollars and euros, the doubling after November, Artificial Analysis ranking (52.6, sixth), cost per task comparisons, the 10,000-scientist offer and a one-million-token limit.
Google says Gemini 4 Argon is a ‘powerhouse’ AI model, and anyone can pay for it (thenextweb.com) – The Next Web summary: launch pricing versus standard rates, Doshi's remarks, Fairwind and Wiz, a European regulatory angle, Google's safety numbers, internal Antigravity gains and Ultra gating.
Gemini app rolling out Gemini 4 Argon to AI Pro and AI Ultra (9to5google.com) – 9to5Google's rollout report: model picker contents, Pro and Ultra availability, Ultra requirement for Antigravity, AI Studio and Android Studio, the YouTube comments optimization, the November 30, 2026 end of introductory pricing and the 10,000-scientist offer.
Developer's guide to Gemini 4 Argon | Gemini Enterprise Agent Platform | Google Cloud Documentation (docs.cloud.google.com) – Google Cloud developer guide: a unified 1,048,576-token input and output limit, long decodes with continuation tokens, LOW, MEDIUM and HIGH thinking levels, implicit caching, SDK version requirements and code quickstarts. Published on docs.cloud.google.com.
Powered by News Ranker