Gemini's cheap model now beats the flagship
14 Nov 2026 · News
Google's new everyday model scores higher than its frontier model on the main independent index. It costs less per task and finishes in less time. On 1 December the frontier model's price doubles.
Gemini 4 Hydrogen, the model Google built to replace its cheap Flash tier, scored 51 on Artificial Analysis's Intelligence Index at its default setting. Gemini 4 Argon, one of Google's two frontier models, scored 50. Hydrogen did it for $1.87 per scored task against Argon's $4.97, in 8.4 minutes against 29.4. For most work the frontier model is now the upgrade you pay extra for, and starting 1 December it costs twice as much.
The numbers
Google launched Hydrogen on 12 November and cited the index result itself. Artificial Analysis, an independent outfit that runs the same nine tests on every model, has the full breakdown on its comparison page:
| At default settings | Hydrogen | Argon | |
|---|---|---|---|
| Intelligence Index | 51 | 50 | |
| Cost per index task | $1.87 | $4.97 | |
| Minutes per task | 8.4 | 29.4 | |
| Price per 1M tokens in / out | $1.50 / $7.50 | $2.00 / $10.00 | |
| Price from 1 December | unchanged | $4.00 / $20.00 |
The price change is on Google's own API pricing page. Argon and Carbon, the frontier pair, cost $2 per million input tokens and $10 per million output tokens through 30 November, then $4 and $20. Cached input doubles too. If Argon uses the same tokens next month, the same index task costs about $9.94. That is our arithmetic, and it puts the flagship at roughly five times Hydrogen's cost per task.
Turn Hydrogen up to its new top setting and it scores 54 for $3.16 a task, still cheaper than Argon today.
Where the flagship still wins
Argon is better at long agent jobs. It leads Terminal-Bench 4.0, a test of working in a command line, 57.1% to 54.0%, and AutomationBench, a test of business workflows, 77.5% to 74.4%. The gap is widest on a science version of Terminal-Bench, 52.9% to 41.4%. Hydrogen wins the coding suite FrontierCode 2.0, 36.3% to 25.7%, and a physics research test called CritPt.
That split is the useful part. An index score averages nine tests, so a team that runs long, multistep agent work in a terminal should check whether Argon's lead shows up on its own jobs before it pays double.
Then and now
A week ago this desk wrote that Inkling 2 Small bought an 18-point index jump with four times the tokens, so its cost per task rose almost sixfold. In October Meta's Muse Wonder showed that the leaderboard can say frontier while the meter says stop. The lesson both times was that cost per task matters more than price per token.
Hydrogen shows the same lesson from the other side. At default settings it used about as many output tokens as Argon to run the whole index, and it charges less for each one. It cost less per task, scored higher and finished faster. Last week the better model was more expensive than its sticker. This week the cheaper model is better than its sticker.
Our call
By 31 March 2027, Google will not have rolled back the 1 December rise. Argon and Carbon will still list at $4 and $20 per million tokens or more. And Hydrogen at its default setting will still cost less than half as much per index task as Argon on Artificial Analysis. A cheap model that beats the flagship is no longer a one-off, and Google is pricing the frontier like a specialist tool. Most teams should default to the cheap model and move a job up only when their own results show it needs the flagship.