Gemini 4 Hydrogen is on Google's free tier and edges out Argon. On the paid API it costs 15x Haiku 5.5.
Google's new everyday Gemini 4 model takes Flash's place in the Gemini app and Search, is free in AI Studio, and costs $1.50/$7.50 per million tokens on the paid API. What changed, how rivals compare, and who should switch.
November 13, 2026

Google released Gemini 4 Hydrogen on November 12, 2026. It is the fast, everyday member of the Gemini 4 family and takes Flash's place. Hydrogen is on the Gemini API's free tier and costs $1.50 per million input tokens and $7.50 per million output tokens on the paid tier. It is also in the Gemini app, in Google Search for all users, and in Google Antigravity for paid Google AI Pro and Ultra subscribers.
Where Hydrogen sits in the Gemini 4 family
The model card says Hydrogen is adapted from Gemini 4 Carbon, with a March 2026 knowledge cutoff. Paid-tier prices per million tokens, from Google's pricing page:
- Gemini 4 Hydrogen: $1.50/$7.50, flat across the 1M-token window. Four thinking levels: low, medium, high (the default) and a new xhigh.
- Gemini 4 Carbon and Gemini 4 Argon: $2/$10 through November 30, then $4/$20. Both default to medium thinking and stop at high. Neither is on the free tier.
- Gemini 3.8 Flash: $0.75/$3.75 through December 31, then $1.50/$7.50 from January 1, 2027, the same as Hydrogen.
- Gemini 3.5 Flash-Lite: $0.30/$2.50, still Google's pick for high-volume classification and routing.
Simon Willison notes: "Hydrogen, supposedly the cheap model, is currently only a 25% discount on its more capable peers." On December 1, when Argon and Carbon double, the gap per token widens to 62.5%.
The free tier and the fine print
- Free tier: Hydrogen is free of charge and rate-limited in AI Studio and the API, per Google's what's new guide. Free-tier content is used to improve Google's products.
- Discounts: cached input costs $0.075, via implicit caching on prompts of 4,096 tokens or more. Batch and Flex cost $0.75/$3.75, and Priority $2.70/$13.50.
- Thinking tokens: billed as output. Hydrogen defaults to high, where 3.8 Flash defaulted to medium.
- Breaking changes from 3.8 Flash: minimal thinking returns an error, explicit caching is gone, temperature, top_p and top_k are ignored, and input, thinking and output share one 1M-token window, with 65,536 tokens per synchronous response.
Verbosity matters too. Artificial Analysis says Hydrogen at xhigh used 304M tokens on its Intelligence Index, against a median of 78.6M, for $3.16 per task. The Decoder puts it at $1.87 per task at high and $0.29 at low.
What changes in Search, the Gemini app and Antigravity
Search. "This is the biggest model upgrade to Search yet," Google's Search post says. It is live globally for all users. Google says Search copes better with multi-part questions and typos, links to more sites and shows more tables and images, in every AI Mode language. 9to5Google reports that Hydrogen is the default model for AI Mode and handles "your hardest questions" in AI Overviews.
Gemini app. Google's model help page now lists Hydrogen for chats on AI Plus ($4.99 a month), Pro ($19.99) and Ultra ($99.99), and for tasks on Pro and Ultra. Flash no longer appears in the plan table. Free users get Hydrogen only when auto model selection picks it; most of their prompts go to Flash-Lite.
Antigravity. Paid Pro and Ultra subscribers can pick Hydrogen directly. A new option, Orchestrator, switches between Carbon and Hydrogen in one session. Google says it reached near-Carbon task success with 24% less quota and 19% faster runs. Orchestrator (Carbon) is for Ultra; Pro gets Orchestrator (Argon).
Google's benchmarks, and one outside check
These are Google-reported. The model card compares Hydrogen, Claude Sonnet 5.5 and GPT-6.1 Terra, in that order:
- Terminal-Bench 4.1 (cheating-adjusted): 57.0%, 63.9%, no Terra score. Gemini 3.8 Flash scored 18.2%.
- Agents' Last Exam v1.1: 71.6%, 60.4%, 69.8%.
- OSWorld-2.1 offline subset: 77.6%, 83.9%, 74.2%.
- Against Argon: 51 to 50 on the Artificial Analysis Intelligence Index with Hydrogen at its default high setting, and 54 at xhigh. Google also says it is 2.5 times faster than 3.8 Flash, with over 5x fewer fabricated claims.
The model card notes a trade-off: "unjustified refusals are higher, particularly for multimodal prompts," up 5.57 points on 3.8 Flash. Figures shift, too: Google's post, updated November 16, gives a 94% agent honesty rate; 9to5Google quoted 96.7% on launch day.
For an outside view, The Decoder reports that Artificial Analysis measured 54% on Terminal-Bench 4.0, behind Sonnet 5.5 at 64% and GPT-6.1 Sol at 56% (both at max), and an 8% hallucination rate on AA-Omniscience. Our leaderboard tracks outside results.
How the price compares with cheap rivals
List prices per million tokens, from our model database:
- Claude Haiku 5.5: $0.10/$0.50 for prompts up to 100K tokens, $0.50/$2.50 above.
- GPT-6 Luna: $0.10/$0.50.
- Muse Ember: $0.10/$0.50 up to 200K tokens, $0.30/$1.50 above.
- GPT-6.1 Terra: $1/$5, and $2.00/$7.50 for prompts over 272K tokens.
Hydrogen lists at 15 times the first three and 1.5 times Terra, though its flat rate beats Terra's input price above 272K. For a month with 200 million input and 20 million output tokens, before caching:
- Haiku 5.5, Luna or Ember: $20 + $10 = $30.
- GPT-6.1 Terra: $200 + $100 = $300.
- Gemini 3.8 Flash: $150 + $75 = $225 through December, $450 from January.
- Gemini 4 Hydrogen: $300 + $150 = $450.
- Argon or Carbon from December 1: $800 + $400 = $1,200.
Per task, the picture shifts. VentureBeat, citing Artificial Analysis, reports that Hydrogen at medium outscored Haiku 5.5, Luna and Ember at max effort, at a higher estimated cost per task than any of them. The pricing index covers other vendors.
Who should switch, and what Google has not said
- On Gemini 3.8 Flash: Hydrogen doubles your per-token cost until December 31, then costs the same. Run your evals now and test medium thinking before accepting the high default.
- On Argon for everyday coding: test Hydrogen. It costs 25% less per token today and 62.5% less from December 1, and it edges Argon on the Artificial Analysis index. Keep Carbon or Argon for the hardest tasks (see our Carbon launch post).
- On high-volume extraction or routing: stay on Flash-Lite, Haiku 5.5, Luna or Ember unless your evals show Hydrogen saves retries.
Google has not said which AI Overviews queries run on Hydrogen, has not published Hydrogen's free-tier rate limits in its docs, and has not said whether 3.8 Flash will be retired or when the full Safety and Security Report will appear; the model card says it is "coming soon." The model card's benchmarks also leave out the $0.10 models.
Some links in this article are affiliate links. Learn more.