# Gemini 4 Hydrogen at $1.50/$7.50: the Foundry workhorse math

Gemini 4 Hydrogen launched on 12 November 2026 at $1.50/$7.50 per million tokens, double today's Gemini 3.8 Flash price but identical to the rate Flash was already scheduled to reach on 1 January 2027, and it arrived on Google's EU multi-region endpoint on day one. That places it between GPT-6.1 Terra and Claude Sonnet 5.5 in Microsoft Foundry, where Claude still has no EU Data Zone, so we price a monthly pipeline and a cached agent session on all five workhorses and map the Azure route.

- Published: 2026-11-13 · Category: AI & Cloud Infrastructure · Tags: Gemini 4 Hydrogen, Google DeepMind, Microsoft Foundry, GPT-6.1 Terra, Claude Sonnet 5.5, LLM Cost, Model Selection, Data Residency, Azure API Management, Azure
- Author: Technspire AB, Stockholm (https://technspire.com)
- Canonical: https://technspire.com/en/blog/gemini-4-hydrogen-1-50-7-50-the-foundry-workhorse-math

Google released Gemini 4 Hydrogen on 12 November 2026 at $1.50 per million input tokens and $7.50 per million output tokens, flat across a 1 million token context window. It is the model Google now points Gemini 3.8 Flash customers at, and it costs exactly double what Flash costs today. For an Azure team in Sweden the interesting part is not the benchmark chart. It is that Hydrogen arrived on Google's EU multi-region endpoint on day one, with data residency and customer-managed keys, at a price that lands between GPT-6.1 Terra and Claude Sonnet 5.5 in Microsoft Foundry. Here is what Google shipped, what the Flash migration actually costs, how the same agent session prices on the Foundry workhorses, and where the residency comparison now favours Google.

## What Google shipped

Hydrogen is the third member of the Gemini 4 family after Argon, announced on 30 September and in the API since 13 October, and Carbon on 10 November. Google DeepMind's model card says it is adapted from Carbon and built for "fast, disciplined execution" rather than long deliberation. The Flash name is gone from the new generation, although the deprecations page lists no shutdown date for Gemini 3.8 Flash, so the old model keeps running. Koray Kavukcuoglu's announcement calls Hydrogen "a major standalone upgrade for teams migrating from Gemini 3.8 Flash", which is the clearest statement Google has made about where Flash customers are expected to go.

The specification that matters for budgeting is the unified 1,048,576 token limit. Input and output share one window. Synchronous output defaults to 65,536 tokens, and a background "long decode" mode can run output to the full 1 million. There are four thinking levels: low, medium, high and xhigh, with high as the default. Thinking tokens are billed as output. The model card gives a knowledge cutoff of March 2026. Output is text only.

Google's own comparison against 3.8 Flash is the headline: tasks finish 2.5 times faster at the default thinking level, the Terminal-Bench 4.1 score rises from 18.2% to 57%, and fabricated claims fall by more than five times. Partners testing SQL and graph workloads reported Hydrogen running 2 to 2.8 times faster while using 12% to 23% fewer tokens than Flash. On the Artificial Analysis Intelligence Index, Hydrogen scores 51 at high and 54 at xhigh. Flash scored 31. Argon, a frontier model priced at $2/$10 until the end of this month, scores 50.

### Against the Foundry workhorses

The model card compares Hydrogen with two models an Azure team can deploy today, Claude Sonnet 5.5 and GPT-6.1 Terra. These are Google's figures, run by Google.

| Benchmark (Google model card, November 2026) | Gemini 4 Hydrogen | Claude Sonnet 5.5 | GPT-6.1 Terra |
| --- | --- | --- | --- |
| SWE-sweep (agentic coding) | 13.5% | 3.1% | 5.8% |
| Agents' Last Exam v1.1 (knowledge work) | 71.6% | 60.4% | 69.8% |
| OSWorld-2.1 (computer use) | 77.6% | 83.9% | 74.2% |
| HLE-Diamond (reasoning) | 58.8% | 53.3% | 47.3% |

Read the SWE-sweep row with care. All three numbers are low, which usually means a new and hard benchmark, and a 13.5% to 5.8% gap on a hard benchmark tells you less than a 71.6% to 69.8% gap on an established one. The Agents' Last Exam line is close to a tie with Terra. Sonnet 5.5 keeps a clear lead on computer use. On the Vals benchmarks Google cites, Hydrogen takes first on Tax Agent Bench at 52.3% and ties first on Legal Research Bench at 55.8%. None of this replaces a run of your own eval set, and Google's own limitations section notes "occasional slowness or timeout issues" at the higher thinking levels.

Google also publishes behaviour figures that matter more for agents than for chat: a 96.7% honesty rate, 2.3 times more resilience to indirect prompt injection by Google's own measure, and reductions of up to 81% in "agentic overreach", meaning the model doing more than it was asked. The Antigravity team pairs this with a new Orchestrator mode where Carbon plans and Hydrogen executes, which Google says reaches near-Carbon task success while consuming 24% less quota and finishing 19% faster. That router pattern is the one to copy, whichever vendor you run it on.

## The Flash price rise was already scheduled

VentureBeat's headline calls Hydrogen a "pricier Flash successor", and in cash terms that is true: $1.50/$7.50 against Flash's $0.75/$3.75. But the Gemini API pricing page has carried a second line for Flash since September. Gemini 3.8 Flash costs $0.75 and $3.75 "through December 31, 2026" and $1.50 and $7.50 "starting January 1, 2027". Hydrogen does not raise the price of the Flash tier. It brings the scheduled doubling forward by seven weeks for anyone who migrates now, and from January the two models cost the same per token.

The same page now settles a question we left open in [the Argon analysis](/en/blog/gemini-4-argon-third-2-10-model-your-azure-gap). Argon and Carbon's introductory $2/$10 runs "through November 30, 2026" and becomes $4/$20 "starting December 1, 2026". The end date we said was missing has been published, and it is seventeen days away.

Take a steady workload of 500 million input tokens and 50 million output tokens a month, with no caching, which is a reasonable shape for a document-processing pipeline. The table prices it on the Flash tier today, on the Flash tier after January, on Hydrogen with Google's claimed token savings, and on the Foundry alternatives at Global Standard list price in Sweden Central as returned by the Azure Retail Prices API on 13 November.

| 500M input + 50M output per month, uncached | Input / MTok | Output / MTok | Monthly |
| --- | --- | --- | --- |
| Gemini 3.8 Flash, until 31 December 2026 | $0.75 | $3.75 | $562.50 |
| Gemini 3.8 Flash, from 1 January 2027 | $1.50 | $7.50 | $1,125.00 |
| Gemini 4 Hydrogen, same token count | $1.50 | $7.50 | $1,125.00 |
| Gemini 4 Hydrogen, 12% fewer tokens (Google's low claim) | $1.50 | $7.50 | $990.00 |
| Gemini 4 Hydrogen, batch tier | $0.75 | $3.75 | $562.50 |
| GPT-6.1 Terra, Global Standard, prompts up to 272K | $1.00 | $5.00 | $750.00 |
| GPT-6.1 Terra, EU Data Zone | $1.20 | $6.00 | $900.00 |
| Claude Sonnet 5.5, Global Standard | $2.00 | $10.00 | $1,500.00 |
| Claude Haiku 5.5, prompts up to 100K | $0.10 | $0.50 | $75.00 |
| GPT-6 Luna, Global Standard | $0.10 | $0.50 | $75.00 |

Two things fall out. First, the batch tier is the escape hatch: a pipeline that can tolerate asynchronous turnaround keeps today's Flash price on Hydrogen indefinitely, as far as the published sheet goes. Second, the Foundry small tier is an order of magnitude cheaper than any of this. If the workload is classification, extraction or routing, Hydrogen is competing with Haiku 5.5 and Luna at $75 a month, not with Terra at $750, and no 12% token saving closes a 15x gap. Hydrogen's case rests on the tasks where the small models fail the eval, which is exactly what [the Haiku 5.5 analysis](/en/blog/claude-haiku-5-5-foundry-0-10-tier-100k-cliff-vs-gpt-6-luna) argued you should measure before choosing a tier.

## The cached agent session, priced five ways

Agents do not run uncached. The session we have used across the Sonnet, Opus and Argon comparisons is a 60-turn agent with a 150K token prefix: one cache write of 150K tokens, 9 million cached input tokens read back over the session, and 120K output tokens. Caching is where the vendors' pricing models diverge most, so the assumptions are spelled out.

Hydrogen on Google Cloud uses implicit caching only. The AI Studio model page says explicit caching is unavailable, and the Gemini Enterprise Agent Platform release note says durable caching is on by default for prompts of 4,096 tokens or more, with cache writes billed at 1.25 times the standard input price. Cached reads cost $0.075 per million tokens. The figures below assume every turn after the first hits the cache, which is the best case. Foundry prices are Global Standard in Sweden Central; the Haiku 5.5 row uses its over-100K tier because the prefix is 150K.

| 60-turn session, 150K prefix, 9M cached reads, 120K output | Cache write | Cached reads | Output | Session |
| --- | --- | --- | --- | --- |
| Gemini 4 Hydrogen, Google Cloud, EU multi-region | $0.28 | $0.68 | $0.90 | $1.86 |
| GPT-6.1 Terra, Global Standard | $0.19 | $0.45 | $0.60 | $1.24 |
| GPT-6.1 Terra, EU Data Zone | $0.23 | $0.54 | $0.72 | $1.49 |
| Claude Sonnet 5.5, Global Standard | $0.38 | $0.90 | $1.20 | $2.48 |
| Claude Haiku 5.5, over-100K tier | $0.11 | $0.45 | $0.30 | $0.86 |
| GPT-6 Luna, Global Standard | $0.02 | $0.09 | $0.06 | $0.17 |

Hydrogen with EU residency costs 25% more per session than Terra with EU residency, and 25% less than Sonnet 5.5 without it. The output column decides it: $7.50 against $5.00 and $10.00. A session that produces less output than this one narrows the gap with Terra; a coding agent that writes a lot of code widens it. Cached reads are priced at 5% of the input rate on all three, so caching no longer separates them. Three vendors, one caching discount.

One practical warning on implicit caching. You cannot pin a cache or guarantee a hit, so the $1.86 is a floor, not a forecast. Budget the session at the uncached rate for the first month, log `cachedContentTokenCount` from every response, and only then write the cached figure into chargeback.

## EU residency on day one, which Foundry's Claude still lacks

This is the part that changes the comparison for a regulated Swedish buyer. The Gemini Enterprise Agent Platform model page lists Hydrogen as generally available on the `global` endpoint and on the `us` and `eu` multi-regions for Standard PayGo, Priority PayGo and Provisioned Throughput, with data residency, customer-managed encryption keys, VPC Service Controls and access transparency all supported. Google's release note is dated 12 November, the same day as the announcement. When Argon launched in October it had no enterprise listing at all. Google has closed that gap for its workhorse tier.

Set that against Microsoft Learn's partner-model page, which still shows Data Zone Standard for Claude as "Not available" in Europe. Sonnet 5.5, Haiku 5.5 and Fable 5.2 can be deployed from Sweden Central, but only as Global Standard, where Microsoft processes prompts in any region. GPT-6.1 Terra and GPT-6 Luna do have an EU Data Zone, at a 20% premium. So the residency ranking for a workhorse model today is: Terra and Luna in the EU zone at +20%, Hydrogen in Google's EU multi-region at list price, and Claude in Foundry with no EU processing guarantee at any price. The Claude gap has been there since [Fable 5.2 arrived this week](/en/blog/claude-fable-5-2-foundry-8-40-and-the-30-day-catch) and before it. Hydrogen makes it more expensive to ignore.

Two caveats keep this from being a clean win for Google. The EU multi-region is EU-wide, not Swedish: Google's European regions include Finland, Belgium, the Netherlands and Frankfurt but not Stockholm for generative models, so a policy that says "within Sweden" is no better served than by Foundry's EU-wide zone. And Hydrogen carries Google's "short-term GA" label with a minimum 45-day retirement notice. Forty-five days is short. [This week's Foundry retirements](/en/blog/foundry-retires-haiku-4-5-and-o-series-your-10-day-plan) show how little time even longer notice periods leave when nobody is watching the deprecation page, so put Google's page on the same watch list.

## Reaching it from Azure

The evaluation route is Google's OpenAI-compatible endpoint imported into Azure API Management as a Language Model API, with the Google key held as a named value and the `llm-token-limit` policy in front. The production route for anything a DPIA touches is the `aiplatform.eu.rep.googleapis.com` endpoint with Private Service Connect, a Google Cloud project, a service account and a second log pipeline into Sentinel. Set the output cap per route, because a model that can emit a million tokens in one long-decode call can bill $7.50 for it.

```
// Evaluation route: Gemini's OpenAI-compatible endpoint behind APIM.
// Production with personal data goes to the EU multi-region endpoint instead.
import OpenAI from "openai";

const client = new OpenAI({
  baseURL: "https://tsp-apim-swc.azure-api.net/gemini",   // APIM facade over generativelanguage.googleapis.com/v1beta/openai
  apiKey: process.env.APIM_SUBSCRIPTION_KEY,
});

const res = await client.chat.completions.create({
  model: "gemini-4-hydrogen",
  max_tokens: 8000,                       // per-route cap; synchronous default is 65,536
  reasoning_effort: "low",                // maps to Gemini thinking level; "high" is the model default
  messages,
});

console.log(res.usage);                   // feed prompt/completion tokens to llm-emit-token-metric
```

Thinking level is the cost lever Google gives you that Foundry's GPT models expose as reasoning effort and Claude exposes as adaptive thinking. Hydrogen's 2.5x speed claim is measured at high. Run your eval at low and medium first. If the pass rate holds, the token bill drops with it, because thinking tokens bill at the $7.50 output rate.

### Also this week: the third decision API

A day before Hydrogen, Google released Gemini Decision 1 and a Decisions API: up to 1,000 yes/no, multiple-choice or rubric questions per request, 150 to 250 milliseconds end to end, $0.10 per million input tokens and nothing for output. That is the third entry in a category we have covered twice, with [Microsoft Decision 1 at $0.042](/en/blog/microsoft-decision-1-foundry-0-042-eu-data-zone-decision-models) and [OpenAI's Decisions API](/en/blog/openai-decisions-api-foundry-ga-luna-vs-microsoft-decision-1) both in Foundry. Google's is Gemini API only for now, with Google Cloud support "coming soon", so there is no residency option yet and no reason for a Foundry shop to move a router call this month. The Retail Prices API still lists Microsoft Decision 1 at $0.042 per million tokens in Sweden Central.

## The Swedish and EU angle

**Residency is now a tie-breaker that favours Google for Claude-shaped workloads.** If your eval prefers Sonnet 5.5 and your policy requires EU processing, Foundry cannot give you both today. Hydrogen in Google's EU multi-region is the first workhorse-class alternative that is cheaper than Sonnet and residency-compliant on launch day.

**A second supplier is still a second supplier.** Google Cloud terms, Google's data processing addendum and sub-processor list, a separate identity boundary and cross-cloud egress on every call. For a kommun or a myndighet that is a new supplier assessment under the procurement rules, not an amendment. Price that work before pricing the tokens.

**The AI Act obligations do not move.** Hydrogen is a general-purpose model and Google carries the provider duties. Your deployer duties, including the logging and human oversight that apply to high-risk uses from 2 December 2027 under the Digital Omnibus timetable, are the same on Hydrogen, Terra and Sonnet. A model estate split across two clouds makes uniform logging harder, which argues again for one gateway in front of every model.

**Budget 2027 on the post-January sheet.** Flash doubles on 1 January, Argon and Carbon double on 1 December, and Hydrogen launched at the doubled rate. Any Google line in a 2027 budget should use $1.50/$7.50 for the workhorse and $4/$20 for the frontier tier.

## Decision guide

- **On Gemini 3.8 Flash today?** Your price doubles on 1 January whether you move or not. Run the eval on Hydrogen at low and medium thinking before December, and move the asynchronous share of the workload to the batch tier, which keeps the current price.
- **Running GPT-6.1 Terra in the EU Data Zone?** Hydrogen costs about 25% more per cached session and offers no residency advantage over you. Stay unless the eval shows a task-level gap, and Google's own table shows Terra within two points on knowledge work.
- **Running Claude Sonnet 5.5 and need EU processing?** Hydrogen is cheaper and residency-compliant. Build the eval set now and test it against Sonnet's computer-use lead before deciding.
- **Classification, extraction, routing?** Haiku 5.5 and GPT-6 Luna at $0.10 input remain 15 times cheaper. Hydrogen only enters this conversation if the small models fail the eval.
- **Building a planner-executor agent?** Copy the Orchestrator pattern on whatever you run: a frontier model for planning and read-only exploration, a workhorse for edits, escalation on failure. Google's figure is 24% less quota for near-frontier results.
- **Writing policy on model retirement?** Add Google's deprecations page to the same watch as Foundry's. Hydrogen's notice floor is 45 days.
- **Not ready to commit?** Import the Gemini OpenAI-compatible endpoint into API Management with a token quota and an output cap. Switching to Hydrogen later is then a model string in configuration, which is the whole point of [the model exit plan](/en/blog/openai-cuts-off-cursor-your-model-exit-plan-on-azure).

Hydrogen does not undercut Foundry. At $1.50/$7.50 it sits between Terra and Sonnet, and the Foundry small tier is untouched. What it does is arrive with EU residency on the first day at a price point where Foundry's Claude cannot match the residency and Foundry's Terra charges 20% for it. For a Swedish buyer that is the number to take into the next model review.

## Sources

- [Google: Introducing Gemini 4 Hydrogen for agentic workflows, 12 November 2026 (pricing, thinking levels, 2.5x and Terminal-Bench claims, availability)](https://blog.google/innovation-and-ai/models-and-research/gemini-models/introducing-gemini-4-hydrogen/)
- [Google DeepMind: Gemini 4 Hydrogen model card (knowledge cutoff, Sonnet 5.5 and GPT-6.1 Terra comparison table, limitations)](https://deepmind.google/models/model-cards/gemini-4-hydrogen/)
- [Google: Gemini API pricing (Hydrogen standard, batch and caching rates; Gemini 3.8 Flash and Gemini 4 Argon/Carbon introductory end dates)](https://aistudio.google.com/docs/pricing)
- [Google: Gemini 4 Hydrogen model page (token limits, implicit-only caching, capabilities)](https://aistudio.google.com/docs/models/gemini-4-hydrogen)
- [Google: Gemini API deprecations (no shutdown date for Gemini 3.8 Flash)](https://aistudio.google.com/docs/deprecations)
- [Google Cloud: Gemini 4 Hydrogen on Gemini Enterprise Agent Platform (eu multi-region, short-term GA, 45-day notice, residency and CMEK support)](https://docs.cloud.google.com/gemini-enterprise-agent-platform/models/gemini/4-hydrogen)
- [Google Cloud: Gemini Enterprise Agent Platform release notes, 12 November 2026 (Hydrogen GA, cache write at 1.25x, Provisioned Throughput)](https://docs.cloud.google.com/gemini-enterprise-agent-platform/release-notes)
- [Google Antigravity: Gemini 4 Hydrogen and Orchestrator (24% less quota, 19% faster)](https://antigravity.google/blog/gemini-4-hydrogen-in-google-antigravity/)
- [Google: Gemini Decisions API, 11 November 2026 ($0.10 input, 1,000 questions per request, latency)](https://blog.google/innovation-and-ai/technology/developers-tools/gemini-decisions-api/)
- [Microsoft Learn: Foundry Models from partners and community (Claude Data Zone Standard not available in Europe)](https://learn.microsoft.com/en-us/azure/foundry/foundry-models/concepts/models-from-partners)
- [Microsoft Learn: Import an OpenAI-compatible Google Gemini API into Azure API Management](https://learn.microsoft.com/en-us/azure/api-management/openai-compatible-google-gemini-api)
- [Azure Retail Prices API (GPT-6.1 Terra, GPT-6 Luna and Microsoft Decision 1 meters, Sweden Central, queried 13 November 2026)](https://prices.azure.com/api/retail/prices?api-version=2023-01-01-preview)
- [VentureBeat: Google launches Gemini 4 Hydrogen, a faster, smarter, pricier Flash successor](https://venturebeat.com/technology/google-launches-gemini-4-hydrogen-a-faster-smarter-pricier-flash-successor)

---

Technspire AB builds AI agents, Azure OpenAI solutions, and production web platforms for Swedish and EU enterprises. Book a call: https://calendly.com/technspire · hello@technspire.com · More articles: https://technspire.com/en/blog · Site overview for agents: https://technspire.com/llms.txt