Gemini 4 Hydrogen: Default Is High, $1.50/$7.50 Is 2x Flash's Promo, and Sync Caps at 64K

Google opened Gemini 4 Hydrogen on November 12, 2026, as the everyday sibling of Gemini 4 Carbon and Gemini 4 Argon. The callable id is gemini-4-hydrogen. It is live on the free and paid tiers of the Gemini API and Google AI Studio, and it is already rolling out in the Gemini app and in Search. The official HN thread was at 551 points this morning; the Antigravity Orchestrator thread was at 451. The model card says Hydrogen is adapted from Carbon. List price is a flat $1.50 / $7.50 per million input / output tokens across the whole 1M window — the same rate Gemini 3.8 Flash moves to on January 1, 2027, but twice Flash's current introductory $0.75 / $3.75, which lasts through December 31, 2026.

Swapping "gemini-3.8-flash" for "gemini-4-hydrogen" is not the whole job. Hydrogen's API default thinking_level is high, not the medium default on 3.8 Flash, Carbon, and Argon. Artificial Analysis's default Hydrogen page is the xhigh variant (Index 54), not the API default. Synchronous output still caps at 65,536 tokens per hop — the 1M long-decode path needs background=true and a raised max_output_tokens. Explicit caching is gone. temperature, top_p, and top_k are ignored. thinking_level: "minimal" returns an error.

What shipped

Vendor docs, compressed:

Google's announcement says Hydrogen finishes tasks 2.5× faster than 3.8 Flash at the default high setting, and more than triples Terminal-Bench 4.1 (57% vs 18.2%). The same model card table also lists Claude Sonnet 5.5 at 63.9% on that bench, so "triples Flash" is not "beats the class." Partner tests Google cites saw 2× to 2.8× faster runs and 12% to 23% fewer tokens versus 3.8 Flash. Those are vendor figures, on their harnesses, not something I reproduced. The rest of this post is about what changes when you switch, the list-price math, and what Artificial Analysis reported at each thinking level this morning.

Migration gotchas — default high, ignored sampling, and the 64K cap

A first Interactions call, as a shape only — this box did not execute it. Pin thinking_level yourself:

from google import genai client = genai.Client() interaction = client.interactions.create( model="gemini-4-hydrogen", input="Outline a safe rollback plan for a failed FastAPI migration, in six bullets.", generation_config={"thinking_level": "high"}, ) print(interaction.output_text) 

If you omit thinking_level, the API default is high — not the medium default a 3.8 Flash, Carbon, or Argon client is probably assuming. Docs say xhigh raises latency and token use and can overthink simpler prompts; eval all four levels, then pick the lowest that meets the bar.

Other changes that will surprise a Flash-era client:

1. minimal is an error. Map any leftover minimal to low or medium. A prompt line such as "answer directly without deliberating" is not a substitute for setting the field.

2. Sampling parameters are ignored, not applied. Remove temperature, top_p, and top_k. The server ignores custom values if you still send them, so a request can look like it has a tighter sampler and not.

3. Explicit caches have to come out. Hydrogen does not support cachedContents or caches.create. Remove cached_content fields. Implicit caching is on by default for prompts of at least 4,096 tokens and discounts cache-hit input by 95%. 3.8 Flash still has an explicit cache with a storage fee; that path does not carry over.

4. The 1M output is not the synchronous default. Hydrogen shares one 1,048,576-token window, unlike 3.8 Flash's split 1M input / 64K output budgets, but the per-hop cap is still 65,536 (streaming or not). max_output_tokens caps combined thinking and output. For anything past 64K, set background to true and raise max_output_tokens above 65,536 (docs suggest 131,072 or 262,144, up to 1,048,576 minus input). Leave the default cap and the server stops at 65,536 even in the background. Raise or remove hardcoded low caps such as 8,192. Versus Carbon and Argon, Hydrogen allows 65,536 tokens per non-streaming hop on REST and the SDK; those two cap a non-streaming REST hop at 32,768.

5. Declare tools or a schema. Do not prompt the model, via system instructions, to emit custom raw-text JSON tool calls with zero declared tools — docs say that can return MALFORMED_FUNCTION_CALL. Use native function calling or structured outputs. On the Interactions API, pass previous_interaction_id for multi-turn and tool sessions; in a client-managed generateContent history, return prior model turns and thought signatures unchanged. Do not copy a thought signature across parallel functionCall parts.

You can already switch models mid-session: route planning turns to gemini-4-carbon and execution turns to gemini-4-hydrogen with previous_interaction_id. No request-schema change is required when you already speak Gemini 4.

$1.50/$7.50 is 2x Flash's promo — and cheaper than Carbon

I ran prices.py on the official list rates (arithmetic, not an invoice and not a measured task cost). For a made-up turn of 25,000 input + 5,000 output tokens, with no cache, no tools, and no grounding:

hydrogen $0.0750 flash-3.8-promo $0.0375 # through Dec 31, 2026 flash-3.8-from-2027 $0.0750 # starting Jan 1, 2027 carbon-argon-promo $0.1000 # through Nov 30, 2026 carbon-argon-from-dec $0.2000 # starting Dec 1, 2026 hydrogen / flash promo (thru Dec 31): 2.00x hydrogen / flash from Jan 1, 2027: 1.00x (same list rate) hydrogen / carbon-argon promo: 0.75x hydrogen / carbon-argon from Dec 1: 0.375x implicit cache hit on the same 25k+5k: $0.0394 cache read vs base input hydrogen: 0.0500x ($0.075 / MTok) 

Equal-token arithmetic is the only comparison I can make without a key. Real sessions will not hold token counts still, and moving from Flash's default medium to Hydrogen's default high will usually spend more output tokens than this table shows. Through the end of 2026, a straight swap at the same token counts is twice Flash's introductory rate; on January 1 the list rates meet. Against Carbon and Argon, Hydrogen is 75% of the current $2 / $10 promo and 37.5% of the $4 / $20 rate that starts December 1. Implicit cache hits are the official lever for a long, stable prefix; there is no explicit cache write you can pin. I am not claiming a Batch price: the paid-tier overview mentions a Batch API at 50% off, but Hydrogen's own pricing table does not list a batch row.

Default thinking is high — and AA's default page is xhigh

The API default is high. Docs say start there for most complex work and reserve xhigh for hard, multi-step problems where high falls short. AA's default Gemini 4 Hydrogen page is the xhigh variant, so a casual look at artificialanalysis.ai/models/gemini-4-hydrogen is not a look at the API default.

I fetched AA's four thinking-level pages this morning (2026-11-13, America/Toronto). These are AA's figures, on their harness (Intelligence Index v5.0.1), not a run from this box:

On that table, high already posts 51, the number Google cites when it says Hydrogen surpasses Argon's 50 at the default setting. Stepping to xhigh adds three Index points and almost triples AA's Index output tokens (110M → 304M), with cost going $1.87 → $3.16 and time 7.9 → 13.7 minutes. AA calls the xhigh page "very verbose" and "slower than average per task," and lists time to first answer token as 40.55s (based on Google's API, per AA). Medium — Flash, Carbon, and Argon's default — is a different model of the work: Index 41, $0.80, 3.7 minutes. "It scored 54 on the Index" is not a reason to pin xhigh in production; if you skip the field, you are already on high.

Orchestrator is a product switch, not a new API schema

In Google Antigravity, Hydrogen is the new default for the Managed Agents / Antigravity agent. A separate dropdown, Orchestrator, switches models inside one session: Carbon explores and writes a plan; Hydrogen edits files, updates packages, and runs tests; friction hands control back to Carbon. The Antigravity post reports that on Terminal-Bench 4.1, Orchestrator reached near-Carbon task success rates while, on average, consuming 24% less quota and completing 19% faster. (The main Gemini announcement cites the same 24% / 19% figures against Terminal-Bench 4.0. I am not reconciling the version labels; both are vendor-reported.) Access, per that post: Hydrogen is rolling out to paid Google AI Pro and Ultra; Orchestrator (Carbon) and Orchestrator (Argon) are for Ultra; Pro gets Orchestrator (Argon); enterprises can use both in Gemini for business. You can already pair the two ids in your own code with previous_interaction_id — you do not need the dropdown.

What this box actually did

preflight 2026-11-13 ~09:07–09:08 America/Toronto NO GEMINI API KEY / NO GEMINI API CALL / NO WEIGHT DOWNLOAD HTTP 200: Google announcement; What's new in Gemini 4 Hydrogen; Gemini Developer API pricing; Hydrogen model card; Antigravity Orchestrator post; developer post HTTP 200: AA /models/gemini-4-hydrogen (xhigh), -high, -medium, -low, and /models/releases/gemini-4-hydrogen HN Firebase: items 50471431 (551/341) and 50472260 (451/137) Parsed: $1.50/$7.50 flat; Flash promo $0.75/$3.75 thru Dec 31, 2026; Carbon/Argon $2/$10 thru Nov 30, then $4/$20; default thinking high; xhigh new; minimal errors; no explicit cache; 65,536 sync cap Local run: prices.py list-price arithmetic only NOT tested: authenticated completion; implicit cache hits; long decode / background hops; mid-conversation model switches; Antigravity or Orchestrator behavior; Vertex AI availability (not in the official launch distribution list) 

For a Toronto product team on 2026-11-13, the rational first move is pin gemini-4-hydrogen with thinking_level set explicitly (start at high for hard turns, measure medium before you spend the Flash-promo gap), strip explicit caches and sampling fields, raise or drop low max_output_tokens caps, and treat xhigh as a ceiling test — not a new default. Keep Carbon or Argon for the planning turns that actually need a frontier model, especially after their $4 / $20 rate starts on December 1.

YongBo Yu — Toronto AI engineer (agent systems / LLM workflows). Site: yongbo-yu.vercel.app · GitHub: YongBoYu1