Gemini 4 Argon launched — an 11-hour overnight coding run, and $2 per million tokens until November 30
Google launched Gemini 4 Argon on October 13, 2026 as its flagship model for long, multi-step professional work. API pricing is $2 per million input tokens and $10 per million output tokens until November 30, then $4 and $20 — the same list price as Claude Opus 5.5. In Google's showcase, Argon worked unattended for about 11 hours on a roughly 205,000-line open-source codebase, proposed six optimizations, and had two merged after human review. On the independent Artificial Analysis Intelligence Index it scores 53, against 58 for Opus 5.5.
The three lines
- Launch — October 13, 2026; API, AI Studio, Gemini app (Pro/Ultra), AI Mode in Search, Antigravity
- Price — $2/$10 per million tokens until Nov 30, then $4/$20 (same as Opus 5.5); 1M-token context
- Results — 11-hour unattended run: 6 proposals, 2 merged; independent index 53 vs Opus 5.5 at 58
Key questions
- Gemini 4 Argon price
- **$2 input / $10 output per million tokens until November 30, then $4 / $20.** | Model | Input | Output | Note | |---|---|---|---| | Gemini 4 Argon (to Nov 30) | $2 | $10 | Intro price; cached input 95% off | | Gemini 4 Argon (from Dec 1) | $4 | $20 | List price | | Claude Opus 5.5 | $4 | $20 | Anthropic API | | Meta Muse Wonder | $4 | $20 | Meta Model API | | GPT-6 Astra | $10 | $50 | OpenAI API |
- Gemini 4 Argon benchmarks
- **Strong on Google's own charts; five points behind Opus 5.5 on the independent index.** | Metric | Argon | Comparison | |---|---|---| | Artificial Analysis Index | 53 | Opus 5.5 58, GPT-6 Astra 53 | | Vals Index | #1 (Google) | Score not given | | ERQA (spatial reasoning) | 87% | Google claims above all frontier models | | Indirect prompt injection success | 0.7% | Gray Swan, 15 adaptive attempts | | Coding shortcut attempts | 2% | Google internal |
- How to use Gemini 4 Argon
- **Developers via the Gemini API and AI Studio; consumers via paid Gemini plans.** | Route | Who | |---|---| | Gemini API / AI Studio | Developers, enterprises (paid tier) | | Gemini app, AI Mode in Search | Google AI Pro and Ultra subscribers | | Antigravity | Google AI Ultra subscribers | | Fairwind Program | Cyber defenders (guardrails relaxed, by application) | | Scientists | 6 months of top Antigravity tier free (selected) |
Google's new flagship is built to work through the night. Gemini 4 Argon launched on October 13, 2026, pitched at professional work that runs for hours across many steps — coding, finance, law, cybersecurity and science. It arrived a day after Meta's first frontier model, Muse Wonder, so two big labs shipped models in the same class and price band within a week.
The pricing play is clear: half the price of Claude Opus 5.5 until November 30, the same price after that.
1. What launched
| Item | Detail |
|---|---|
| Release | October 13, 2026 |
| Model ID | gemini-4-argon |
| Context | 1M tokens, unified input-output |
| Price (to Nov 30) | $2 input / $10 output per million tokens; cached input 95% off |
| Price (from Dec 1) | $4 / $20 |
| Consumers | Google AI Pro and Ultra (Gemini app, AI Mode in Search) |
| Coding | Antigravity (Ultra) |
| Security | Fairwind Program for defenders, guardrails relaxed |
Google also credited Argon with scientific contributions: a quantum arithmetic implementation with over 100x lower cost than a published baseline, a counterexample in fusion-related mathematics, rare-disease genomics work, and helping Chrome finish a months-long refactor in weeks.
2. What the 11-hour run actually produced
Google's developer blog let Argon loose on google/adk-python, its open-source agent toolkit — about 205,000 lines, ~800 files, ~17,000 unit tests.
| Round 1 | Round 2 (overnight) | |
|---|---|---|
| Duration | ~2 hours | ~11 hours, from 9:11 p.m. |
| Instructions | One-paragraph goal | Repo rulebook + /goal mode |
| Output | Convention problems; altered a public schema; ran only 216 tests | 6 optimizations across 32 files; ran 16,892 tests |
| Human review | — | 2 of 6 merged, 4 rejected |
| Gains | — | ~21% faster import; ~40% less session prep in long sessions |
The rejected four are the instructive part: a cache that sped up a synthetic benchmark by 17% but risked stale data, and code reading a private attribute, among others. One change passed every open-source test but broke 25 internal Google test targets. The difference between round one and round two was not the model but the rules file and goal setup around it.
3. Against the field
| Model | Independent index | Input /M | Output /M | Release |
|---|---|---|---|---|
| Claude Opus 5.5 | 58 | $4 | $20 | 2026 |
| Meta Muse Wonder Full | 56 | no API | — | Oct 12 |
| Gemini 4 Argon | 53 | $2 → $4 | $10 → $20 | Oct 13 |
| GPT-6 Astra | 53 | $10 | $50 | 2026 |
| Meta Muse Wonder | 52 | $4 | $20 | Oct 12 |
On the independent index Argon trails Opus 5.5 by five points and ties GPT-6 Astra. Google's headline claims — #1 on the Vals Index, 87% on ERQA spatial reasoning — are mostly self-reported.
4. What remains open
- December price rise: whether usage built during the discount survives the move to $4/$20.
- Independent testing: little beyond Artificial Analysis so far, and that score is marked as updated.
- Release date discrepancy: Artificial Analysis lists September.
- Fairwind safeguards: selection criteria for the guardrail-relaxed version are not detailed.
Sources
- Google Blog — Gemini 4 Argon: Built to solve your hardest problems
- Google Developers Blog — Optimizing a 200,000-line Python codebase overnight with Gemini 4 Argon
- Google Cloud Blog — Gemini 4 Argon available on Google Cloud
- Artificial Analysis — Gemini 4 Argon
- Interesting Engineering — Google model cuts quantum costs 100x, tackles fusion equations