What Google released
Google DeepMind announced Gemini 4 Hydrogen on 12 November 2026, which was 13 November in New Zealand. Google positions it as the fast, cheaper model in the Gemini 4 family, alongside the frontier models Gemini 4 Carbon and Gemini 4 Argon. The launch post describes Hydrogen as a model for everyday coding and agent work. It can run on its own or act as the execution model in a multi-model setup. VentureBeat reports it as the successor to Gemini 3.8 Flash.
The model ID is gemini-4-hydrogen. According to Google's developer notes, it's available on the free and paid tiers of Google AI Studio and the Gemini API, through both GenerateContent and the Interactions API. Google Cloud lists it as a short-term generally available release with a minimum of 45 days' notice before retirement. Consumers get it in the Gemini app and Google Search. Free app users only reach it through automatic model selection.
Pricing
On the paid Standard tier, the Gemini API pricing page lists Hydrogen at US$1.50 per million input tokens, US$7.50 per million output tokens (thinking tokens included) and US$0.075 per million cached input tokens. The rate stays the same across the whole 1M-token context window, so long prompts don't hit a higher price band. Batch and Flex both cost half the Standard rate. Batch targets a 24-hour turnaround. Flex runs individual requests at lower priority with a target response of 1 to 15 minutes, according to VentureBeat. An invite-only Contributor tier also costs half price, and Google uses the data from those requests to improve its products. The pricing page says free-tier traffic is used the same way. Grounding with Google Search is free for the first 5,000 requests a month, then costs US$14 per 1,000.
The comparison with Gemini 3.8 Flash matters for budgeting. Flash's introductory price is US$0.75 input and US$3.75 output until 31 December 2026, and its scheduled price from 1 January 2027 is exactly Hydrogen's rate. Moving to Hydrogen doubles per-token cost today. From January, Flash costs the same per token as Hydrogen. VentureBeat also points out that Hydrogen's standard rates are about 15 times the cheapest standard rates for Claude Haiku 5.5, GPT-6 Luna and Meta's Muse Ember. Hydrogen is a mid-priced model, well above the budget tier.
Benchmarks and their caveats
Google's headline figures are an Artificial Analysis Intelligence Index score of 51 at the default high thinking level and 54 at xhigh, against 50 for Gemini 4 Argon. Google also reports 57% on Terminal-Bench 4.1, a benchmark of agent tasks in a command-line environment, where Gemini 3.8 Flash scored 18.2%. Google says Hydrogen finishes tasks about 2.5 times faster than 3.8 Flash and makes more than five times fewer fabricated claims.
Some of these numbers need care. VentureBeat notes that Artificial Analysis ran its own Terminal-Bench 4.0 test and got 54% at both high and xhigh, so the extra reasoning added cost without improving the score. Google's 4.1 figure and Artificial Analysis's 4.0 figure come from different benchmark versions and can't be compared directly. Artificial Analysis estimates the cost per benchmark task at US$1.87 on high and US$3.16 on xhigh, against US$7.77 for 3.8 Flash on high. Those are weighted estimates for that benchmark suite and won't match your bill. On the separate Vals AI index, Hydrogen ranks sixth at 68.49% accuracy and US$9.84 per test, and Wccftech describes it as the cheapest and fastest model in that top six. The safety figures in the launch post, including an 81% reduction in overreach on agent tasks, come from Google's internal evaluations and haven't been checked independently.
What changes in your code
Several API changes will break or quietly alter existing Flash integrations. The developer notes say Hydrogen drops temperature, top_p and top_k. In-range values are ignored, and out-of-range values return a 400 error. The Cloud model page lists the fixed values as temperature 1.0, topP 0.95 and topK 64. Any pipeline that relied on temperature 0 for repeatable output will need structured outputs or stronger evaluation checks instead.
Thinking is set with generation_config.thinking_level, which accepts low, medium, high and a new xhigh. The default is high, where 3.8 Flash defaulted to medium, so a direct model swap will spend more thinking tokens per request. The minimal level and the thinking_budget parameter both return errors. Explicit caching through cachedContents isn't supported. Implicit caching kicks in automatically for prompts of at least 4,096 tokens and gives a 95% discount on cache hits.
Input, thinking and output tokens share one context window of 1,048,576 tokens. A single synchronous request can produce at most 65,536 output tokens. Longer output, up to the full window, needs background: true together with a max_output_tokens value above 65,536. Google advises removing old low caps such as 8,192, because max_output_tokens counts thinking and answer tokens together. Google also says tools should be declared with native function calling. Imitating tool calls with raw JSON in the system instructions can return MALFORMED_FUNCTION_CALL.
generation_config = {"thinking_level": "medium"}Setting the level explicitly, as in the snippet above, is the simplest way to stop a migration from silently raising your token spend.
Limits for New Zealand teams
Data residency is the main constraint for local builders. The Google Cloud model page lists ML processing locations as the us and eu multi-regions only, with pay-as-you-go also available on the global endpoint. Neither Sydney nor any Asia-Pacific location is listed. Teams with contracts or privacy assessments that require Australian processing can't use Hydrogen on Vertex for that work yet. The free tier of the Gemini API is a poor fit for client data, because Google uses free-tier content to improve its products.
The Cloud page also lists features Hydrogen doesn't support: RAG Engine, tuning, the Gemini Live API and Fixed quota. Hydrogen does accept Office and OpenDocument files directly, along with up to about 45 minutes of video with audio or 8.4 hours of audio per prompt.
Where it fits
Hydrogen makes most sense as the worker model in agent loops with many tool calls and document-heavy steps. Box's early-access results, reported by VentureBeat, showed the biggest gains on table reading and calculations. For high-volume classification or extraction, the budget models still cost far less per task. Before switching, we'd rerun your own evaluation set at medium and high, compare cost per completed task against 3.8 Flash at its January price, and check that removing temperature doesn't hurt output consistency. Members who benchmark Hydrogen on local workloads can share the results on the ASAI projects page.
Sources
- Introducing Gemini 4 Hydrogen (Google blog)
- What's new in Gemini 4 Hydrogen (Google AI Studio docs)
- Gemini API pricing (Google AI Studio docs)
- Gemini 4 Hydrogen model page (Google Cloud docs)
- Google launches Gemini 4 Hydrogen, a faster, smarter, pricier Flash successor (VentureBeat)
- Google's Flash-class Gemini 4 Hydrogen is now the cheapest, fastest model in Vals AI's top 6 (Wccftech)
- ASAI Insights feed
- New AI Models, November 2026 LLM Releases (LLM Gateway)
Build with people working on the same things
Join the Auckland Society for AI: NZ$599 a year with a ForgeBox and the members' projects area. Our Auckland meetups are open to everyone.