Gemini 4 Argon Is Out for Paying Users: Exactly Who Gets It, What It Costs Until November 30, How to Turn It On in Cursor, and Why It Isn't Showing Yet

Every way to get it, with the price

What you're getting

• A 1M-token window shared between input and output — 'an industry-leading 1M unified input-output token limit.' Default synchronous output is 65,536 tokens; up to 1,048,576 via Google's new long-decode 'continuation protocol.'
• Thinking levels low, medium, high (default medium on Vertex; Antigravity runs high). Inputs: text, image, video, audio, PDF. Output: text only — no image or audio generation, no Live API. Computer use in preview.
• Knowledge cutoff March 2026, per the model card published October 13; 'large scale sparse mixture-of-experts.'
• Caching is implicit only, at 95 percent off input. No long-context surcharge — one price across the window, unlike GPT-6 Astra's step-up past 272K.
• Logan Kilpatrick's launch note: 'battle tested by over 100,000 people at Google over the last month' and 'This is just the start of the Gemini 4 era.' Reuters reports Google 'no longer planned to release Gemini 3.5 Pro.'

Where it stands, as of today

• Independent, general: Artificial Analysis Intelligence Index 53 at high effort, eighth of 216 — behind Claude Opus 5.5 (58), Sonnet 5.5 (56), and Muse Wonder Full (56), level with Fable 5.1 and GPT-6 Astra. AA's September note still holds: a 15 percent hallucination rate against 51 percent for Astra, and about $2 per Index task at intro pricing.
• Independent, where it leads: first on the Vals Index at 68.9 percent (Sonnet 5.5 67.0, Opus 5.5 67.0), first on Vals' Finance Agent and IOI; first on the Arena text leaderboard at 1525 (Muse Wonder 1513, Opus 5.5 1508).
• Coding, independent: Vals Terminal-Bench 4.0 57.6 percent, sixth; not yet on tbench.ai (where Opus 5.5 via Claude Code leads at 64.8). Google's own table lists Argon at 57.4 — 'self computed' — against Opus 5.5's provider-reported 66.4, which is Anthropic's harness and not comparable.
• CursorBench, read carefully. Cursor's post says '52.8% on CursorBench at default effort, on par with Opus 5.5.' On Cursor's evals page that's Argon at medium (52.8) against Opus 5.5 at medium (52.5). Opus at max scores 57.8 and Argon at high 52.9, so the parity is at the default setting, not at the ceiling — true, and narrower than it sounds. GPT-6 Astra isn't on CursorBench or in Cursor.
• Google's table (vs Astra, Fable 5.1, Opus 5.5) is the September 30 table unchanged — competitor numbers are 'sourced from providers' self reported numbers' at max settings, and it dates its benchmarks 'as of September, 2026.' Our Argon breakdown walks the rows.

The robot, as a vendor claim

The launch's most-shared demo is DeepMind's: 'Gemini 4 Argon taught a robot to play cornhole just by watching a video… and taught Apptronik's Apollo humanoid to sink 7 throws in a row.' Google's blog describes the pipeline — watch a raw video, 'mapped the spatial coordinates of the environment, and calculated the required whole-body joint movements,' then 'autonomously built the simulation code,' transferred to the robot — and calls it 'Video-to-Sim-to-Real.' There is no paper, no count of attempts, no distance, and no sim-to-real detail; the companion robotics post covers a different, tabletop experiment (a photo to a 'calibrated MuJoCo simulation twin,' raising a real-robot success rate from 56 to 75 percent). Apptronik hasn't posted its own account. Impressive if true; it's a two-minute clip until someone publishes the method.

If it isn't showing

• Check the plan, not the app. Pro and Ultra only. Plus ($4.99) gets an upgrade prompt; Free gets nothing. A Pro plan does not unlock Argon in AI Studio or the API — that's 'not available… on the Google AI Pro plan' by design.
• Give it a few hours. On launch day, Pro users posted 'I'm a Pro user and I do not have it' at 16:12 UTC and 'OK, now I have it too (2–4 hours later)' the same afternoon. Google's text says 'rolling out today globally' and 'select countries'; it never says 'over the coming days,' but that's how it behaved.
• Antigravity says you need Ultra because you do. 'Only Ultra' there is the policy, not a bug.
• macOS app workaround, user-reported: one person got the selector to populate by clicking the Antigravity button in the top right and switching back. Unofficial.
• Cursor and Copilot: Settings → Models in Cursor; Copilot's rollout is 'gradual,' and Copilot Pro isn't included.
• No help-center article exists yet. Google hasn't replied to the complaints we found.

What changes in our rankings

Gemini moves from third to second in AI chatbots, above Claude, and its rating rises from 4.7 to 4.8. The hold we placed on September 30 — 'pending app availability' — is lifted because the thing we were waiting for happened: a frontier-class model on the $19.99 plan, in the app most people use, with the strongest free tier in the category underneath it. ChatGPT stays first: scale, Dots, the plugin layer, and GPT-6 Astra on Pro. Claude stays the writer's and coder's choice, and Opus 5.5 is still ahead of Argon on every independent coding harness — which is why the code editors ranking doesn't move, though Antigravity's note now reads 'Ultra-only for Argon.' The thing we'll watch: whether Pro's 'expanded access' means caps that bite, and what the Gemini 4 'family' Logan Kilpatrick teased looks like.