3-Line TL;DR
- Gemini 4 Argon matches Astra on AA's AI performance index and has appealing launch prices
- Earlier Antigravity frustrations and Gemini 3 fixes that felt slow still make me hesitate
- Once my account has access, I would try one month and hold off on annual payment
The performance is worth a try
- My earlier Gemini 4 post covered rumors; the September 30 announcement brought official results and prices
- In Artificial Analysis's AI performance test, Argon High and Astra Max both scored 53 (AA evaluation)
| Google's tests | Argon | Astra | Opus 5.5 |
|---|---|---|---|
| Extended coding — DeepSWE v1.1 | 77.9% | 74.1% | 74.2% |
| Business task execution — AutomationBench | 51.3% | 41.4% | 42.5% |
| Command-line tasks — Terminal-bench 4.0 | 57.4% | 58.2% | 66.4% |
Source: Google DeepMind — Google's published tests, separate from AA; chart by HSL
- Argon leads two of these tests, while Opus leads command-line work
- The coding and automation results make me want to try it on my own work
The low price depends on a launch discount
- These are developer API charges, separate from a monthly subscription
| Argon API pricing | Input per 1M tokens | Output per 1M tokens |
|---|---|---|
| Launch discount | $2 | $10 |
| After the discount | $4 | $20 |
Source: Google — the discount's end date is unspecified
- In AA's test, Argon cost $1.99 per task at the discount price, versus $3.26 for Astra Max
- At regular prices, the same Argon usage would cost $3.98 and exceed Astra's cost in that test
Source: Artificial Analysis — its test workload; $3.98 applies regular prices to the same usage, rather than a new test
Antigravity is why I would wait on annual payment
- In my coding-tool comparison, Antigravity felt fast and good value but less finished overall
- The 2.0 app I used lacked multilingual support and the remote-server workflow I wanted
- Gemini 3 fixes also felt slow to me
- A February 23, 2026 report of repeated Gemini 3.1 output received a same-day staff escalation (Google forum)
- Another user reported similar problems on March 25, though the thread does not establish a shared cause or the exact fix date
- Improvements have continued: the July 13 changelog includes overload retries and hang fixes (Official changelog)
- I want a tool that works comfortably every day, as well as a model with good scores
When can ordinary users try it?
| Access stage | Announced plan |
|---|---|
| Initial access | Selected cyber defenders through Fairwind and trusted testers |
| Next | Paid API customers and Google AI Ultra subscribers first |
| Broader access, including Pro and free accounts | No confirmed date announced |
Source: Google's rollout plan — checked October 1, 2026
- Ultra-first access does not mean Pro is permanently excluded
- Before paying, I would check that my account has Argon and that the price and limits suit my work (Plan guide)
- Model access and limits vary by plan, and limits can change with server capacity
- A month sounds worth trying; a full year is more trust than I have yet
Community Reactions
- The launch thread includes excitement about the scores and requests to see real results after public access (Reddit)
- Pro users worry about Ultra getting access first, while replies caution that this does not mean permanent exclusion (Reddit)
- An older Antigravity thread has both error complaints and successful web-app reports, rather than evidence of today's reliability (Reddit)
Q&A (Field Notes)
- Q. Do these API prices tell me what a subscription costs?
- No, check the actual monthly fee and included limits separately
- Q. Have I tried Argon?
- No, the performance figures come from published tests, and my experience concerns earlier Antigravity versions
- Q. What would I check during a one-month trial?
- Use the same tasks as my current tools and compare results, retries, quota use and recovery after an interruption