Abner Li at 9to5Google, October 13, is the product note. Gemini 4 Argon is rolling out in the Gemini app to Google AI Pro and AI Ultra. Last month, 9to5 says, initial availability was Ultra only, the $99.99-and-up tier. Google’s line is “frontier reasoning and improved performance across complex workflows.” That is a marketing sentence. The meter is not.
The Verge puts Argon in the Gemini app and in AI mode in Search for Pro and Ultra, and on Antigravity for Ultra. It is not clear when, or if, Plus or free users get it. Ars Technica is the same rollout with the lights off: paying users, lower tiers without full access. The Next Web says Pro starts at $19.99 a month and that Google moved free users to its smallest model four days earlier.
We already met Gemini 3.8 Flash working harder. Decisions API in 150ms is a different product that still does not write the email. Mellum2.1’s SWE-bench asterisk is another vendor scoring themselves. This week is a paid Gemini SKU with two prices, two clocks, and a coding surface that still sits behind Ultra.
Pro is the chat box. Ultra is the shop
If you pay for AI Pro, 9to5 says Argon is in the Gemini app now. That is the consumer surface: questions, docs, the thing people mean when they say “Gemini.” It is not Antigravity. It is not AI Studio. 9to5 is explicit that Argon in AI Studio and in Android Studio Canary still wants AI Ultra.
Antigravity is Google’s agentic development platform in these write-ups. The Verge keeps it on Ultra. TNW repeats Ultra at $99.99. If your team heard “Argon is out” and opened the free app, you are not on the story. If you opened Pro and expected the IDE agent, you are also not on the story.
Ars says Google revealed Argon a few weeks ago and delayed a wide release over “nefarious use.” The delay did not last long. Capabilities, Ars says, did not change from the late-September benchmarks. The model is trained to refuse malicious instructions, and an automated misalignment monitor can halt a workflow. Google’s claim is that it needs that because the model is capable. Treat the monitor as a product assertion, not as an audit you ran.
Thousands of Googlers have been using it internally, per Ars. That is a dogfood claim. Useful as “it exists on real desks.” Useless as your latency budget.
$2 and $10 are a sale with two calendars
9to5 prints the API intro price: $2 per million input tokens, $10 per million output, until November 30, 2026. After that, $4 and $20. The Verge says the intro rate is around a fifth the cost of OpenAI’s GPT-6 Astra and Anthropic’s Fable 5, and that even after the double at the end of November it should still be cheaper per token. The Verge also warns that per-token stickers do not equal per-task cost if models spend tokens differently.
TNW agrees on $2 / $10 versus $4 / $20, then says the launch discount runs “at least a month with no stated end date.” That fights 9to5’s November 30. Do not pick a favorite in a spreadsheet. Put both in the footnote and ask Cloud billing which meter is live on your project. If you are signing a quarter of inference, the difference between “end of November” and “unspecified” is the whole negotiation.
TNW’s consumer floor is $19.99 Pro. 9to5’s Ultra floor is $99.99+. Those are seats, not tokens. A Pro seat that chats in the app is not an API budget. An Ultra seat that unlocks Antigravity is not 10 million output tokens. Separate the SKUs or you will explain a bill with the wrong screenshot.
Introductory API pricing is how vendors buy switchers. If you rewrite a stack onto Argon because $2 looks like a fifth of Astra, calendar the November 30 row now. Doubling is still “cheap” on The Verge’s comparison only if Astra and Fable 5 stay where they are. They will not send you a courtesy email.
The YouTube emoji story is a demo. It is also a 45% claim
9to5 quotes a speed run on an A100 via vLLM: 2.85× end-to-end inference, about 37,900 tokens per second to over 108,000 on Gemma 4 E2B, 100% token parity, zero quality drop, “a few days” instead of months. That quote is about Argon used as an optimizer on a different model, not about your chat latency in the Gemini app. Keep the nouns straight.
The YouTube comments frontend is the anecdote everyone will repeat. 9to5: Argon in Antigravity traced production profiles, blamed redundant emoji rendering for nearly 45% of server overhead, produced 12 targeted optimizations, cut server comment-load latency by 30 ms and client end-to-end by 20 ms. Ars shortens it to a 45% server-load drop from eliminating redundant emoji rendering. The Verge says Argon halved the compute for YouTube’s comments section.
Those are Google-reported engineering stories. 30 ms is real if you are YouTube. It is not a reason to buy Ultra for a 20-person app. If you want the same class of win, you need production profiles, permission to change the frontend, and someone who can tell a bad patch from a good one. Antigravity does not supply that person.
The Verge adds two more internal uses: finding and patching cybersecurity vulnerabilities, and debugging the drone physics simulator at Alphabet’s Wing. Ars says the Wing simulator bugs had been in the codebase since 2018, and that Argon showed a step up in embodied reasoning for robotics. Company examples. Not your drone.
TNW’s science claims are larger and should stay in quotes: counterexamples to a 1967 conjecture about fusion-reactor magnetic fields; a humanoid learning cornhole from a video. Do not put either in a procurement deck without the paper.
Fairwind, Wiz, and the Vals Index are not your eval
TNW says the first users were not developers. Argon went to a small group of cyber defenders without its guardrails, under a program Google calls Fairwind, now growing toward hundreds of partners. Wiz, TNW writes, used the model to scan critical infrastructure and says it found a bypass exposing a master key in healthcare software. That is Wiz’s finding as reported. It is not a CVE you should paste without Wiz’s own write-up.
Tulsee Doshi, Gemini product lead at Google DeepMind, called Argon a “powerhouse” and the first of a family, TNW says, talking to Business Insider. Google says it ranks first on the Vals Index for finance, coding, legal, and tax work. Company-reported leaderboard. Run your own golden files. We already watched a vendor self-report SWE-bench. Same reflex.
The Verge says Google claims Argon outperforms GPT-6 Astra, Claude Fable 5.1, and Claude Opus 5.5 across coding, cybersecurity, science, finance, and law. That is a press list. If those names are your actual alternatives, bake off on your prompts, not on Google’s slide.
Ars notes nothing about capabilities changed from the September reveal. If you already decided Argon was too hot to deploy, the October 13 news is distribution, not a new model card. If you were waiting for Pro, the news is the Gemini app. If you were waiting for a free tier, TNW’s “smallest model” move is the answer you did not want.
What to actually turn on this week
Chat users on Pro: open the Gemini app and check whether Argon is selectable. If it is not, you are still on last month’s Ultra-only gate. Do not assume Search’s AI mode has it because the app does. The Verge lists both; your account may not.
Developers: hit the API on a throwaway key and log input/output tokens on ten real tasks. Price them at $2/$10 and again at $4/$20. If the November price kills the use case, you do not have a use case. You have a sale.
Antigravity shoppers: budget Ultra. Confirm Argon is actually attached to your org, then pick one repo with tests. The YouTube story started from a profile, not from a vibe. If you do not have profiles, you will get refactors you cannot measure.
AI Studio and Android Studio Canary: 9to5 says Ultra. Canary is not your release train. Do not demo Argon to a VP on a Canary build and then act shocked when stable Android Studio does not have it.
Security reviewers: Ars’s refuse-and-monitor story is the residual risk from the delayed launch. Ask Google what the monitor logs, who sees a halt, and whether Fairwind-style ungated access exists in any of your contracts. TNW flags cyber offense as a systemic-risk category in Europe. If you are in that jurisdiction, your counsel already has a list. Add Argon to it by name.
Free-tier teams: you are not late. You are on a different product. TNW says the small model is what you have. Build there, or pay. There is no secret Argon toggle.
Token math that survives November
Take a boring internal task: 20,000 input tokens of repo context and 4,000 output tokens of a patch, fifty times a day. At intro rates that is 1 million in and 0.2 million out, or $2 plus $2, four dollars. At $4 / $20 it is $4 plus $4, eight dollars. Neither number will impress a CFO. Now make the agent retry. Five samples per task, same prompt, is 5× output. Intro: $2 + $10 = $12. Standard: $4 + $20 = $24. That is a day. A month of that loop is where “a fifth of Astra” stops being a slide.
The Verge’s warning about token-hungry models is the line to paste under those rows. If Argon spends twice the output tokens of Astra to reach the same patch, the fifth-the-price claim is already gone before November 30. Log tokens on your tasks. Do not log “it felt cheaper.”
9to5’s vLLM quote is a throughput story on Gemma 4 E2B, not a discount on your chat. 108,000 tokens per second on an A100 is an infra team brag. If you do not run vLLM, you cannot spend it. If you do run vLLM, read the pull requests they mention (attention tuning, speculative decoding) instead of pasting 2.85× into a board deck.
Search’s AI mode is a different surface from the Gemini app. The Verge lists both for Pro and Ultra. A lawyer who only uses Search will not see Antigravity. An engineer who only uses AI Studio Canary will not see Search. When someone says “we turned on Argon,” make them say which of the four doors: app, Search, API, Antigravity. Then ask which SKU paid for it.
The story on October 13 is not that Google trained a new brain. It is that the paid doors opened, the API is on a clock, and the coding agent still costs a hundred dollars a month. Pro will make Argon feel ubiquitous in chat. It will not make Antigravity cheap. Keep those tabs separate when someone forwards The Verge and writes “we should switch” in Slack.