Google made Gemini 4 Hydrogen (gemini-4-hydrogen) generally available on November 12, 2026, on the Gemini API’s free tier and at a paid rate of $1.50 per million input tokens and $7.50 per million output tokens. Gemini API changelog That is twice what Gemini 3.8 Flash costs today and exactly what 3.8 Flash will cost from January 1, 2027, when its introductory price ends. Gemini API pricing Hydrogen also adds a fourth thinking level, xhigh, above a default that is already high, and thinking tokens are billed as output. Thinking docs
S5 Labs has not run Hydrogen. The Gemini prices, limits and migration changes below are read from Google’s documentation, the Anthropic and OpenAI prices come from those companies’ own pages, and the benchmark and speed figures are Google’s unless they are marked as coming from Artificial Analysis.
Twice 3.8 Flash until December 31, level with it from January 1
Standard paid rates per 1M tokens, from the pricing page:
| Model | Input / output today | Scheduled change | Cached input |
|---|---|---|---|
gemini-4-hydrogen | $1.50 / $7.50 | None listed | $0.075 |
gemini-3.8-flash | $0.75 / $3.75 through December 31, 2026 | $1.50 / $7.50 from January 1, 2027 | $0.075, then $0.15 |
gemini-4-carbon and gemini-4-argon | $2.00 / $10.00 through November 30, 2026 | $4.00 / $20.00 from December 1, 2026 | $0.10, then $0.20 |
The page lists no price-change dates for Hydrogen. It does not call the rate introductory, and the absence of a date is not a commitment to hold it. Google’s launch post adds that the rate is “flat across its entire 1 million-token context window”. Introducing Gemini 4 Hydrogen
The next table prices one request on each model: 100,000 uncached input tokens and 10,000 output tokens (thinking included) at Standard rates. This is S5 Labs arithmetic from the rates above, and it assumes identical token counts on every model, which will not hold once thinking differs.
| Model | November 12 | From December 1 | From January 1 |
|---|---|---|---|
| Hydrogen | $0.225 | $0.225 | $0.225 |
| 3.8 Flash | $0.1125 | $0.1125 | $0.225 |
| Carbon or Argon | $0.30 | $0.60 | $0.60 |
By the same arithmetic, Hydrogen is 75% of the Carbon and Argon rate today and 37.5% of it from December 1. From January 1 the only gap left in Standard per-token rates between Hydrogen and 3.8 Flash is cached input, where Hydrogen’s $0.075 is half of Flash’s $0.15. 3.8 Flash also carries a cache storage price, $1.00 per 1M tokens per hour from January 1; Hydrogen lists none, and its caching is implicit only. Our December 1 article covers the Carbon and Argon doubling, and the Argon pricing and migration article lists Argon’s rates by tier.
Batch and Flex are $0.75 / $3.75 with cached input at $0.0375 and no free tier. Priority is $2.70 / $13.50 with cached input at $0.135. Gemini API pricing
Hydrogen does not compete on price with the cheapest tier elsewhere. Claude Haiku 5.5 (for prompts up to 100,000 tokens) and GPT-6 Luna are both $0.10 / $0.50, so Hydrogen costs 15 times as much per token as either. Claude Sonnet 5.5 at $2 / $10 is the closer comparison, and Hydrogen’s rate is 25% below it. Anthropic pricing · GPT-6 Luna model page Our Haiku 5.5 pricing article covers the cheap tier.
The free tier has no published per-model limits
On Standard, the pricing page gives Hydrogen a free tier with input and output free of charge. Google uses free-tier content to improve its products and does not use paid-tier content that way. Search and Maps grounding are not available on the free tier. Neither Carbon nor Argon is on the free tier at all, so of the three, Hydrogen is the only one with a free evaluation path through the API. Gemini API pricing
The rate limits page publishes no per-model request or token limits for Hydrogen. It gives only Batch enqueued-token caps: 3,000,000 on Tier 1, 400,000,000 on Tier 2 and 1,000,000,000 on Tier 3. Check what your own project is allowed before planning an evaluation around the free tier.
Artificial Analysis lists Hydrogen only at xhigh
Google says Hydrogen scores 51 on the Artificial Analysis Intelligence Index at its default high setting, against 50 for Argon, and 54 at xhigh. Introducing Gemini 4 Hydrogen Artificial Analysis lists only the xhigh variant, at 54, and lists Argon (High) at 50. The 51 at high has no entry there, so it remains Google’s number. Artificial Analysis, Gemini 4 Hydrogen · Artificial Analysis, Gemini 4 Argon
Artificial Analysis’s Hydrogen page reports that the xhigh variant generated 304M tokens to run the index, against a median of 72.8M, and calls that “very verbose”. Argon (High) generated 113M and 3.8 Flash (High) 126M. Artificial Analysis, Gemini 3.8 Flash We are not quoting the cost-per-task figures on those pages: they did not reconcile with the pages’ own token counts and prices when we checked the arithmetic, and the Flash page uses the introductory price while the Argon page uses the December one.
The rest is vendor-reported, and we found no independent reproduction. Against 3.8 Flash, Google reports 57% versus 18.2% on Terminal-Bench 4.1, tasks finished 2.5 times faster at its high default setting, and fabricated claims cut by more than 5x. It says unnamed partners saw, in early testing, runs 2x to 2.8x faster with 12% to 23% lower token usage than 3.8 Flash on complex graph, SQL and data-heavy enterprise workflows. A separate result concerns Orchestrator in Google Antigravity, which switches between models such as Carbon and Hydrogen within a session: near-Carbon task success on Terminal-Bench 4.0 while, on average, using 24% less quota and finishing 19% faster, with no baseline stated. That is a system result and says nothing about Hydrogen’s speed alone. Introducing Gemini 4 Hydrogen
Correction and update, November 17, 2026: This paragraph originally called the Orchestrator figure a result “on a different benchmark version” from Hydrogen’s Terminal-Bench 4.1 score. Google’s two November 12 posts disagree on that. The launch post cited above says Terminal-Bench 4.0, in its text and in the alt text of its chart, and still did when we re-read it today; the page is marked as last updated November 16. The Antigravity team’s post reports the same 24% and 19% figures “On Terminal-Bench 4.1”. Gemini 4 Hydrogen and Orchestrator We cannot tell which version is right, so we have removed the claim. Neither post says what the 24% and 19% are measured against, and both numbers are Google’s.
The Antigravity post also says who gets Orchestrator. Google AI Ultra subscribers get Orchestrator (Carbon) and Orchestrator (Argon), paid Google AI Pro subscribers get Orchestrator (Argon), and Google says more models and plans will be supported later. Enterprises “can access both in Gemini for business”. As an example of the switching, the post has Carbon exploring and planning in a read-only phase, Hydrogen writing the edits and running tests, and Hydrogen handing back to Carbon when it “hits friction”. Gemini 4 Hydrogen and Orchestrator
Google’s model card, published the same day, shows unjustified refusals up 5.57 percentage points against 3.8 Flash, where lower is better, and says this is particularly the case for multimodal prompts. Its refusal-tone measure is down 0.65 percentage points, which the card describes as similar to 3.8 Flash. Google adds that a manual review of flagged content found the losses in its automated safety evaluations were overwhelmingly false positives or not egregious. Gemini 4 Hydrogen model card
The card also assesses Hydrogen against Critical Capability Levels (CCLs), which Google’s Frontier Safety Framework defines as capability levels at which a model, absent mitigations, “may pose heightened risk of severe harm”. Google says it could not rule out that Argon and Carbon had reached the CCLs for CBRN (biological) uplift and cybersecurity uplift, so it ran targeted tests on Hydrogen in both domains. For both, the card reads “CCL not reached, following additional targeted testing. CCL alert threshold reached.” The full Safety & Security Report is listed as coming “in the coming days”. Gemini 4 Hydrogen model card
What changes when you move from 3.8 Flash
From Google’s migration guide:
| Area | On Hydrogen | What to do |
|---|---|---|
| Default thinking | high, where 3.8 Flash defaulted to medium; minimal is rejected with an error | Set thinking_level in generation_config explicitly |
| Context window | One 1,048,576-token window shared by input, thinking and output; 3.8 Flash had separate 1M input and 64K output budgets | Budget input and output together |
Explicit caching (cachedContents, caches.create) | Not supported | Remove cached_content fields. Implicit caching runs on prompts of at least 4,096 tokens and takes 95% off cached input |
temperature, top_p, top_k | The server ignores custom values | Remove them |
| Output over 64K | Synchronous responses support up to 65,536 tokens, and a background request without max_output_tokens also stops there | Set background=true and a higher max_output_tokens, or use client-side continuation in the latest SDKs |
| Hardcoded low caps such as 8,192 | A low limit can cut answers short | Remove them or raise to 65,536; to cut tokens, lower the thinking level instead |
| Raw-text JSON tool calls with no declared tools | Can return MALFORMED_FUNCTION_CALL | Use native function calling or structured outputs |
| Conversation state | Thought signatures are part of the history | Return prior model turns unchanged in generateContent histories, or use previous_interaction_id in the Interactions API |
The caching, sampling and output-limit rows repeat what we listed for Argon in October. Output is text only, and the model page lists audio generation, image generation and the Live API as unsupported. The changelog also makes Hydrogen the default model for the Antigravity agent in Gemini Managed Agents, so check which model your agent runs now use.
Update, November 17, 2026: A Google post dated November 17 announces a Robotics plugin for Antigravity, with skills for robot bring-up, MuJoCo simulation, and perception and control; Antigravity itself is for paid Google AI Pro and Ultra subscribers. In the same post Google recommends that developers using Gemini Robotics ER 2 try Hydrogen for those workloads, through gemini-4-hydrogen in the Interactions API. We have not tested either. Gemini 4 Hydrogen: a faster model for general purpose robotics
Nothing in the changelog forces the move. No entry deprecates 3.8 Flash or routes it to Hydrogen, and the November 2 entry sends traffic toward 3.8 Flash instead: from December 1, gemini-2.5-pro requests route to gemini-3.8-flash and are billed at its rates, so that traffic meets the January price step too. Our December 1 article covers that reroute, and the auto-routing article covers the October reroutes of 3.7 and 3.5 Flash.
Price xhigh by the thinking tokens it adds
Thinking tokens are billed with output at $7.50 per million, on the full count the model generates even though the API returns only a summary. usage.total_thought_tokens reports the count. max_output_tokens includes thinking: if the limit is hit mid-reasoning the status is incomplete, the output may be truncated or empty, and the thinking tokens are still billed. Thinking docs
Take a request with 20,000 input tokens and 2,000 tokens of visible output. It costs $0.045 before any thinking, and every 10,000 thinking tokens add $0.075. The token counts are illustrative and the arithmetic is ours. Google’s docs give no figure for how many more tokens xhigh uses than high, so measure it on your own prompts.
Google’s guidance is to reserve xhigh for hard multi-step problems where high falls short. The migration guide says it “can cause overthinking on simpler prompts”, and the thinking docs warn that it can exhaust max_output_tokens on simple tasks such as text or JSON extraction. Carbon, Argon and 3.8 Flash default to medium and have no xhigh, so a workload moved from 3.8 Flash without an explicit thinking_level gets a higher default and, until December 31, double the token price.
Run your evaluation set at all four levels with a high max_output_tokens. For each level record total_thought_tokens, output tokens, latency, the share of incomplete responses and your pass rate, then divide spend by accepted results, the comparison we described for Sonnet 5.5’s cost per finished task. Pick the lowest level that meets your quality bar, which is also what Google recommends. If you are on 3.8 Flash, run the same set there before December 31, while its introductory price still applies. From January 1 the two list prices match, and the choice comes down to tokens per accepted result and cache hits.
