Model · Google · released September 30, 2026 (limited release) · checked October 3, 2026

Gemini 4 Argon: benchmarks, independent tests, price and when you can use it

Google’s first Gemini 4 model claims more top scores than any other frontier model. What it scores, what it will cost, who can use it now, and when to pick it.

The short answer

Google’s own tables put it ahead; the first independent tests put it level with GPT-6 Astra and behind Claude Opus 5.5. It is the most careful about facts, uses many more tokens per task, and almost nobody can use it yet.

Google’s announcement (opens in a new tab)

1 · Performance · vendor-reported unless marked

Google’s numbers and independent tests

Independent index (Artificial Analysis): 53—level with GPT-6 Astra (53) and behind Claude Opus 5.5 (58); 8th of 224 models; $1.99 per task in their run, versus $3.26 for Astra.Independent
Factual reliability (AA-Omniscience): The lowest hallucination rate of any model scoring 45 or more: 15% vs 51% for Astra—but fewer right answers: 50% vs 63%.Independent
Tokens per task: About 62,000 output tokens per task vs about 27,000 for Astra—2.3× more, which eats into the price advantage.Independent
Software engineering (DeepSWE v1.1): 77.9% vs 74.2% for Claude Opus 5.5 and 74.1% for GPT-6 Astra.Vendor-reported
Business workflows (AutomationBench): 51.3% vs 42.5% for Opus 5.5 and 41.4% for Astra.Vendor-reported
Legal agent (Harvey): 19.6% vs 5.4% for Astra and 3.8% for Opus 5.5.Vendor-reported
Finance agent (Vals Finance Agent v2): 65.4% vs 58.6% for Opus 5.5 and 53.5% for Astra.Vendor-reported
Long video (LVBench): 91.7% vs 87.5% for Astra and 83.7% for Opus 5.5.Vendor-reported
Finding vulnerabilities (CWE-bench v1): 68%, tied with Astra; Opus 5.5 at 67%.Vendor-reported
Prompt-injection attacks (Gray Swan): 0.7% succeeded vs 1.0% for Opus 5.5 and Fable 5.1 and 8.5% for Astra. Lower is better.Vendor-reported
Where it trails: FrontierSWE v2: 55.0% vs Astra’s 65.5%. Terminal-Bench Science: 57.6% vs 68.1%. Terminal-Bench 4.0: 57.4% vs Opus 5.5’s 66.4%.Vendor-reported

Rows marked Independent come from Artificial Analysis (checked October 3, 2026; Astra and Opus figures as reported by The Next Web). The rest are Google’s own evaluations as compiled by VentureBeat, which counts Argon leading or tying on 13 of 18. Bloomberg reports that some Google staff doubt its coding and front-end design and worry about benchmark-chasing; that is reporting, not a test.

2 · Price and task cost

What it costs

Introductory API input: $2 per 1M tokens
Introductory API output: $10 per 1M tokens
Cached input: 95% off the input price
After the introductory period: $4 input / $20 output per 1M tokens. Google has not said when the introductory period ends.
Longest response: Up to 1M output tokens, up from 64K

3 · Where you can use it

Availability

Now: Cyber defenders in Google’s Fairwind Program—more than 650 partners, including governments, critical-infrastructure operators and core technology platforms—plus U.S. government pre-release testing.
Next: Paid Gemini API customers and Google AI Ultra subscribers, “as soon as possible.” No date.
Not yet: Not on Google’s Gemini API pricing page as of October 1, 2026, and no public model ID.

4 · Our read · opinion

When to use it

Worth waiting for

Long, multi-step business work—legal, finance, automation—and defensive security, where Google’s numbers lead most clearly.

Keep GPT-6 Astra or Opus 5.5 for now

Hard science and terminal-heavy coding, where Google’s own tables put Argon behind—and you can use them today.

Compare at the later price

Introductory $2 / $10 matches GPT-6.1 Sol and Sonnet 5.5; the later $4 / $20 matches Opus 5.5. Test your own task at the price you will actually pay.

The Google desk →GPT-6.1 Sol →Claude Opus 5.5 →Claude Fable 5.1 →