Gemini 4 Hydrogen at $1.50/$7.50: Flash Cheaper Until 2027

– views

Source: Google Blog — Introducing Gemini 4 Hydrogen (announcement, 12 November 2026)

  • Gemini 4 Hydrogen
  • Gemini API
  • API pricing
  • model selection
  • coding agents
  • Google DeepMind
  • AI news

Google priced its new everyday model at exactly the rate its predecessor stops charging on 1 January 2027. Google announced Gemini 4 Hydrogen on 12 November 2026 as the successor to Gemini 3.8 Flash at $1.50 per million input tokens and $7.50 per million output tokens, flat across the whole 1,048,576-token context window — and no introductory promotion on the standard tier. That is not a discount from where Gemini 3.8 Flash sits today: Flash launched on 1 September 2026 with a 50% launch credit that Google’s own pricing page runs to 31 December 2026, which makes its current rates $0.75 input / $3.75 output, and its standard rates — the ones that apply from 1 January 2027 — $1.50 / $7.50. So if you are weighing a migration from 3.8 Flash on price alone, Hydrogen is 2× more expensive today and identical from January. What is different is the shape of the bill: thinking tokens now bill as output, caching is implicit-only, and four thinking levels replace the sampling knobs you may still be sending.

Disclosure: this site has no affiliate relationship with Google, Alphabet, Anthropic, OpenAI or DeepSeek, and we received nothing from any of them. Our Vultr page does carry a direct affiliate relationship, and that is the only affiliate link below. Every rate quoted here was read by us on 16 November 2026 from the pages listed in the sources block, and each table says which page a number came from. Two caveats we will not hide: blog.google and aistudio.google.com refused our fetcher on 16 November (connection reset, three attempts, both the English and the Portuguese mirror), so the official announcement text is quoted from search-engine captures of those pages, from Google’s own Chinese-language announcement published on its official WeChat account on 13 November, and from Google Cloud’s Model Garden pricing page, which we could fetch. Independent price cards (LiteLLM’s day-0 card and its model registry, plus a third-party /models listing record from 13 November) are used to cross-check every Google figure, and where Google and a gateway disagree, or where Google’s plans page and its help page disagree, we say so instead of picking one.

The rate card, tier by tier

Prices per 1M tokens. The standard tier is Google’s, confirmed by the official Chinese announcement (“在其完整 的 100 万 token 上下文窗口内实行统一计价:每百万输入 token 1.50 美元,每百万输出 token 7.50 美元”); the Batch, Flex and Priority rows are the ones we could only read from LiteLLM’s day-0 card and its model registry.

TierInputOutputCached inputSource we read
Standard (global)$1.50$7.50$0.075 (implicit only)Google announcement; LiteLLM card, 12 Nov 2026
Batch / Flex$0.75$3.75scales with tierLiteLLM card, 12 Nov 2026
Priority$2.70$13.50n/aLiteLLM card, 12 Nov 2026
Free tier (Gemini API / AI Studio)free of charge, rate-limitedfree of chargefree of chargeGoogle pricing page via third-party capture; LiteLLM FAQ
Grounding with Google Searchfirst 5,000 queries/month free (shared with Gemini 3 and newer), then $14 per 1,000 queriesLiteLLM FAQ, 12 Nov 2026
Regional (non-global) endpoints+10% over the rows aboveLiteLLM card (Gemini Enterprise Agent Platform)

Three things follow from that table. Hydrogen is now the cheapest Gemini 4 row, because Gemini 4 Argon and Gemini 4 Carbon are at $2.00 / $10.00 until 30 November 2026 and $4.00 / $20.00 from 1 December (LiteLLM, corroborated by Google’s Carbon announcement as summarised by Promptea). Batch/Flex at $0.75 / $3.75 is the only way to match today’s Flash price, and it costs you latency. And the free tier is new to this tier of the family: Hydrogen is the first Gemini 4 model listed on the Gemini API free tier, while Argon’s own entry states it “is not available on the Free tier”.

Why this is not a price cut: the Flash promo is the comparison

Model / momentInput / 1MOutput / 1MCached input / 1MAs of
Gemini 3.8 Flash — launch promo, valid to 31 Dec 2026$0.75$3.75$0.075LiteLLM day-0 card (1 Sep 2026); cache figure via a LiteLLM-sourced calculator page verified 22 Oct 2026
Gemini 3.8 Flash — standard, from 1 Jan 2027$1.50$7.50see note belowLiteLLM day-0 card; Google Cloud Model Garden lists the 31 Dec 2026 / 1 Jan 2027 split with a “50% credit” footnote
Gemini 4 Hydrogen — from 12 Nov 2026, no promo$1.50$7.50$0.075 (implicit only)Google announcement; LiteLLM card

Note on Flash’s standard cached rate: the pages we could fetch publish Flash’s promotional cached rate ($0.075, per a LiteLLM-sourced calculator page verified 22 October 2026) and LiteLLM’s statement that “cache read, batch, flex and priority tiers all get the proportional discount”. We therefore treat a standard cached rate of $0.15 from 1 January 2027 as our arithmetic, not as a Google-published number. Same caveat applies to any cached-input figure for Flash after the promo ends.

Our arithmetic on those published rates, for a workload of 1M input tokens and 100K output tokens:

  • On Flash today: 1 × $0.75 + 0.1 × $3.75 = $1.125.
  • On Hydrogen at default high: 1 × $1.50 + 0.1 × $7.50 = $2.25 — 2.0×.
  • On Hydrogen with 800K of the input served from implicit cache: 0.8 × $0.075 + 0.2 × $1.50 + 0.1 × $7.50 = $1.11. Note the cache read is billed at $0.075 for both models, so on cache-saturated traffic the gap narrows to the output rate.
  • On Hydrogen at xhigh with thinking roughly tripling billed output to 300K tokens: $1.50 + 0.3 × $7.50 = $3.75, i.e. 3.3× the Flash bill for the same visible answer.

VentureBeat’s own estimate of cost per benchmark task lands in the same place: $1.87 at high and $3.16 at xhigh. That is the outlet’s estimate, not Google’s, and we did not reproduce it.

What actually controls the bill: four thinking levels, and thinking bills as output

Set with the thinking_level field: low, medium, high (default) and a new xhigh. Google’s advice is to start at the default and move only when your own evals justify it. Two mechanical catches come with it:

  • Thinking tokens are billed at the $7.50 output rate, so xhigh costs more even when the visible answer is the same length. LiteLLM maps reasoning_effort to thinking_level, and as of 12 November its xhigh passthrough was still open in PR #47086 — worth checking if you route through a gateway.
  • max_output_tokens caps thinking and output together. Google’s guidance is to raise an old low cap (Apidog gives 8,192 as the example) to 65,536 and lower the thinking level instead, otherwise answers get cut short. Per synchronous request Google returns up to 65,536 output tokens against the 1,048,576 window shared by input, thinking and output.
  • Google ignores temperature, top_p and top_k on Hydrogen. If your config sets them, delete them; they are silently inert rather than an error.
  • Caching is implicit only, from 4,096 prompt tokens. Explicit caching is not supported. Google’s own caching doc describes explicit caching as a paid feature billed on cached tokens and TTL storage; on Hydrogen you cannot buy a guaranteed cache, so your hit rate is not a lever you control. Compare with DeepSeek, where cached input is the headline number: deepseek-flash (DeepSeek-V4.1-Flash-1017) bills cache hits at $0.003 off-peak / $0.006 peak and cache misses at $0.15 / $0.30, with output at $0.60 / $1.20 (DeepSeek API docs, fetched 16 November 2026).

Cross-family price context, 16 November 2026

ModelInput / 1MOutput / 1MNotesSource date
Gemini 4 Hydrogen$1.50$7.50cheapest Gemini 4 row; free tier available12 Nov 2026
Gemini 4 Argon / Carbon$2.00 → $4.00 on 1 Dec$10.00 → $20.00same promo window for both10–12 Nov 2026
Claude Fable 5.2$8.00$40.00down from Fable 5.1’s $10 / $50; 1M context9 Nov 2026
Gemini 3.8 Flash (promo, to 31 Dec)$0.75$3.75standard rate returns 1 Jan 20271 Sep 2026
DeepSeek deepseek-flash$0.15 / $0.30 (miss) · $0.003 / $0.006 (hit)$0.60 / $1.20peak = off-peak × 2; version DeepSeek-V4.1-Flash-1017docs fetched 16 Nov 2026
Gemini Decision 1 (scoring only)$0.10$0.00classifies ≤256K input; writes no text11 Nov 2026

The pattern across the same eight days is documented in our listing-check story — where we also recorded Artificial Analysis putting Gemini 4 Carbon (high) at 59 — and in our Haiku price-cliff note. Hydrogen’s Artificial Analysis Intelligence Index score is 51 at high and 54 at xhigh (the figures LiteLLM and Notebookcheck both read off Google’s comparison material), which puts it one point above the larger Gemini 4 Argon (50) and far below Carbon — Carbon (high) sat at 59 on the same index when we recorded it a few days earlier. In other words Google sold Hydrogen on speed-per-dollar, not on being the smartest row in the family.

What Google claims, what outlets reported, and what nobody documented

ClaimStatusWhere it lives
2.5× faster task completion than Gemini 3.8 Flash at the default high setting; fabricated claims cut by more than 5×; Terminal-Bench 4.1 57% vs 18.2%Documented by Google (Google’s own testing)Google announcement
2–2.8× faster and 12–23% fewer tokens than 3.8 Flash on graph, SQL and heavy-data workflows, from “partners”Documented by Google, third-party conditions unpublishedGoogle announcement (Chinese mirror, fetched 16 Nov)
5th on Agent Arena; HLE-Diamond 58.8%; 1st on Vals Tax Agent Bench (52.3%); tied 1st on Vals Legal Research Bench (55.8%); leads AutomationBench-AA classDocumented by Google, no independent re-run we could fetchGoogle announcement (Portuguese and Chinese mirrors)
Present in the live gemini/ /models listingIndependently recorded by a third partycopilotkit/aimock drift record, detected 13 Nov 2026; LiteLLM day-0 support 12 Nov
“Pricier Flash successor”Reported by outletsVentureBeat headline, ETV Bharat
Google AI Plus $4.99/mo now lists Hydrogen as its headline model; the Pro model left the Plus column and starts at $19.99 (which includes Gemini 4 Argon)Reported by an outlet from Google’s plans page, and Google’s help page still disagreed on 13 NovNotebookcheck, 13 Nov 2026
Free users cannot select Hydrogen; it arrives only through auto model selection, while most prompts still go to Flash-LiteReported by an outlet, quoting Google’s help pageNotebookcheck + Gemini Apps help page
Rate limits (RPM / TPM / RPD) for Hydrogen on each paid tierNot disclosed to us — ai.google.dev refused the fetch; we quote no numbers—
Google’s own benchmark runs for the models it compares againstNot disclosed—

If you run on Gemini 3.8 Flash today: what we would actually do

  1. Do not migrate for cost. On 16 November 2026, Flash is half the price on input and output and the same $0.075 on cache reads. The migration case is Terminal-Bench, the fabricated-claim drop and the 2.5× wall clock — all Google’s own numbers. Run your own eval set before you treat them as yours.
  2. Put 31 December 2026 and 1 January 2027 in the calendar. Flash’s 50% credit ends on the 31st and its standard rate starts on the 1st; from then the two models are priced the same, and “should we move” stops being a price question. Carbon and Argon step up a day earlier, on 30 November.
  3. Re-measure, do not re-use, your token budget. Thinking billed as output plus a shared max_output_tokens cap means an old config tuned for Flash can silently cost 2–3× more or truncate. Fix the cap first, then pick the thinking level per route.
  4. If caching is your cost model, this is the weakest part of the migration. Implicit-only from 4,096 prompt tokens, with explicit caching unsupported, means your cache hit rate is Google’s decision. If your product is a long system prompt replayed thousands of times, price the DeepSeek-style cached-input rows too, and read our note on a KV-cache leak between tenants before assuming a shared cache is free of other people’s bytes.
  5. If your plan, not your API bill, is what you buy: Google’s paid consumer entry for Hydrogen is AI Plus at $4.99 per the plans page as reported on 13 November, and free-tier users reach Hydrogen only through auto model selection. On gemini.google.com you can check which model answered a reply: menu under the response → Show thinking steps → the model name is at the bottom.
  6. And if you are renting GPUs to escape per-token pricing, that comparison only closes against an open-weights row like DeepSeek’s, not against Hydrogen — our Vultr page is the affiliate page in this article, and the priceVerifiedDate field on that page reads 25 September 2026, so treat the GPU numbers there as stale and re-check before you buy anything.

What we will watch next

  1. 30 November 2026 — Gemini 4 Carbon and Argon move from $2/$10 to $4/$20. Hydrogen’s price advantage over the rest of the Gemini 4 family roughly doubles overnight.
  2. 31 December 2026 / 1 January 2027 — the Flash 50% credit ends. We will re-fetch both rate cards on 1 January and publish the diff; that is the date the migration math changes, not the launch date.
  3. Whether Google publishes Flash → Hydrogen parity numbers for tool-calling agents, and whether explicit caching or a xhigh-through-gateway path shows up in an AI Studio what’s-new page.
  4. 17 November 2026 — NVIDIA reports under an interim CEO, per our boardroom story; Anthropic’s IPO pricing is reported for 18 November. Both feed the compute line item that every per-token price below them is built on.
  5. Whether the free tier survives contact with demand. Google lists Hydrogen free with no published RPD number we could fetch, and free users cannot select the model directly.

We fetched Google’s announcement mirrors, Google Cloud’s Model Garden pricing page, Google’s plans and help pages, three LiteLLM artifacts and a third-party model-listing record on 16 November 2026; the pages that refused us are named at the top rather than quietly dropped. Every dollar figure above is a copy of what the named page said on that date, and every benchmark number is attributed to whoever ran it. We will update this entry with an updatedDate when Google publishes a rate change, when the 30 November and 1 January step-ups land, or when we get an independent benchmark run we can cite.

More coverage