Gemini 4 Hydrogen: Pricing, API Changes and Migrating From 3.8 Flash

Google released Gemini 4 Hydrogen on November 12, 2026. It is the fast, everyday model in the Gemini 4 family and replaces the "Flash" line. It costs $1.50 per million input tokens and $7.50 per million output tokens, which is twice Gemini 3.8 Flash's current rate. Google says it scores slightly above Gemini 4 Argon on a major intelligence index. If you build on Gemini 3.8 Flash, the upgrade is mostly a model-ID change, but a few parameters now fail or get ignored. Re-check your cost assumptions too, because Hydrogen thinks harder by default.

This guide covers what Hydrogen is, where you can use it, what it really costs, what breaks when you migrate, and when it beats Carbon, Argon or a cheaper model.

What Gemini 4 Hydrogen is and where it runs

Google's announcement, written by Google DeepMind head Koray Kavukcuoglu, calls Hydrogen "our daily driver for agentic workflows." Gemini 4 Carbon and Gemini 4 Argon stay the frontier models for deep, long-horizon reasoning. Hydrogen is meant for speed and volume: coding, tool calls, data work and everyday chat.

It is the third Gemini 4 model in just over six weeks. SiliconANGLE notes that Argon and Carbon both started with Google's vetted Fairwind cybersecurity program before wider access. Hydrogen skipped that stage and opened to everyone on launch day.

Where you can use it from November 12:

The AI Studio developer guide says the model is available through both the GenerateContent and Interactions APIs. It also says the managed Antigravity agent (antigravity-preview-09-2026) now uses Hydrogen by default.

Benchmarks: what Google claims, and what to discount

Google's headline figure is 51 on the Artificial Analysis Intelligence Index at the default "high" thinking level, against 50 for Argon. The new "xhigh" level raises it to 54. Other numbers Google reports:

• Terminal-Bench 4.1: 57%, against 18.2% for Gemini 3.8 Flash.
• Speed: tasks finish about 2.5x faster than 3.8 Flash at the default setting. Early partners reported 2x–2.8x faster runs with 12%–23% fewer tokens on SQL, graph and data-heavy workflows.
• Other evals: 58.8% on HLE-Diamond, 52.3% on Vals Tax Agent Bench, 55.8% on Vals Legal Research Bench, and #5 on Agent Arena.
• Safety: up to 81% less overreach in agentic tasks, up to 2x less likely to bypass permission boundaries, and 96.7% on Google's internal agent honesty evaluation.

Three caveats come out of the independent coverage.

1. Most of these numbers are vendor-reported. VentureBeat points out that much of the data comes from Google's own launch comparison or early-access testing. In Google's own Terminal-Bench 4.1 table, Claude Sonnet 5.5 scores 63.9%, ahead of Hydrogen.
2. xhigh isn't always worth paying for. On Artificial Analysis's Terminal-Bench 4.0 run, Hydrogen scored 54% at both high and xhigh. The estimated cost per task still rose from $1.87 to $3.16.
3. Benchmark versions differ. Google's Antigravity post gives the Orchestrator result on Terminal-Bench 4.1, while SiliconANGLE and VentureBeat say 4.0. Treat cross-version comparisons loosely.

One independent data point is useful. Box CEO Aaron Levie shared that Box's agent went from 74% to 79% accuracy on its internal Complex Work Eval after moving from Flash to Hydrogen, with less than half the run time and about 12% fewer tokens, as reported by VentureBeat.

Pricing: the per-token rate doubles, but per-task cost may not

Hydrogen is priced flat across its 1M-token window. There is no surcharge for long prompts.

VentureBeat also reports that Batch (24-hour turnaround) and Flex (1–15 minutes, limited availability under load) cost half the standard rate. There is also an invite-only "Contributor" tier, where Google can use your request data for training.

Here is a worked example: an agent step with 20,000 input tokens and 2,000 output tokens (including thinking), run 100,000 times a month, with no caching.

• Hydrogen: $0.045 per call, about $4,500/month
• 3.8 Flash (2026 pricing): $0.0225 per call, about $2,250/month
• Argon (intro): $0.06 per call, about $6,000/month. After November 30: about $12,000/month

That arithmetic assumes the same token counts, and they won't be the same. Hydrogen defaults to a higher thinking level than Flash, which adds output tokens. On the other hand, partners report it finishes tasks with fewer tokens overall. Artificial Analysis's per-task estimates, quoted by VentureBeat, show this clearly: Hydrogen at high cost about $1.87 per benchmark task, and 3.8 Flash at high cost $7.77. A higher price per token can still mean a lower bill per finished job if the model takes fewer turns and fewer retries. Measure tokens per completed task, not per call.

Implicit caching matters too. It is on by default for prompts of at least 4,096 tokens and gives a 95% discount on cache hits. Agents that resend a long system prompt and tool list on every turn will pay far less than the table above suggests. To model your own traffic, plug real token counts into the LLM API Cost Calculator, or measure your prompts with the Token Counter.

Hydrogen is not the cheap option in absolute terms. VentureBeat notes its standard rate is 15x the lowest standard rates of Claude Haiku 5.5, GPT-6 Luna and Meta's Muse Ember, which start at $0.10 input and $0.50 output. Even at medium thinking, Hydrogen's estimated per-task cost is higher than those three models at max settings, though it outscores them.

Migrating from Gemini 3.8 Flash: what actually breaks

The AI Studio migration guide is short, but several changes will throw errors or quietly change behaviour. Work through them in this order:

1. Change the model ID to gemini-4-hydrogen.
2. Remove minimal thinking. Hydrogen rejects it with an error. The valid thinking_level values are low, medium, high (default) and xhigh. VentureBeat also reports that the thinking_budget parameter is rejected, so switch any budget-based config to thinking_level.
3. Pick a thinking level on purpose. Flash defaulted to medium and Hydrogen defaults to high. If you send no level, latency and cost will change. Run your evals at medium and high and choose the lowest level that meets your quality bar. Google gives the same advice.
4. Remove sampling parameters. The server ignores temperature, top_p and top_k. If your app relies on temperature for determinism or variety, you'll need another approach, such as structured outputs or prompt-level constraints.
5. Remove explicit caching. cachedContents, caches.create and cached_content fields aren't supported. Rely on implicit caching instead, and put stable content (system prompt, tool declarations) at the start of the prompt so it hits the cache.
6. Raise low output caps. max_output_tokens now covers thinking plus output combined. A hardcoded 8,192 cap can cut off answers once the model reasons at high. Google suggests raising caps to 65,536.
7. Use background mode for very long outputs. A synchronous hop tops out at 65,536 tokens. For more, set background: true and a max_output_tokens above 65,536. Without the cap, background jobs still stop at 65,536.
8. Declare tools natively. If you prompt for raw-text JSON tool calls without declaring tools, you can get MALFORMED_FUNCTION_CALL. In multi-turn sessions, use previous_interaction_id, or return prior turns and thought signatures unchanged.
9. Re-baseline billing. Google says billing includes thinking tokens, internal system instructions and function declarations.

Moving from Carbon or Argon needs no schema changes. The differences are defaults: Carbon and Argon default to medium and cap at high, while Hydrogen defaults to high and adds xhigh. Hydrogen also allows non-streaming hops of up to 65,536 tokens, against 32,768 on Carbon and Argon REST.

The guide lists no deprecation date for 3.8 Flash. You don't have to move today, but the VentureBeat-reported January 1 price change would remove Flash's cost advantage.

Orchestrator: splitting work between Carbon and Hydrogen

The Antigravity post introduces Orchestrator, a model option that switches models within one task. Carbon explores and plans in a read-only phase. Hydrogen writes edits, updates packages and runs tests. If Hydrogen gets stuck, it hands back to Carbon to revise the plan. Google says this gets close to Carbon's task success while using about 24% less quota and finishing about 19% faster.

Access is split by plan. Orchestrator (Carbon) is for Google AI Ultra subscribers, Orchestrator (Argon) is for Pro, and enterprises get both through Gemini for business.

You don't need Antigravity to copy the pattern. The developer guide says you can route planning turns to Carbon and execution turns to Hydrogen in one API session using previous_interaction_id. If your agent spends most of its tokens on edit-and-test loops rather than planning, a planner/executor split like this is one of the easier cost savings available right now. Compare it with other AI coding assistants before you commit.

What Hydrogen in Search means for sites and launches

Google's Search post calls this "the biggest model upgrade to Search yet." Hydrogen is now the default model for AI Mode, and 9to5Google reports it also handles the hardest queries in AI Overviews. Google says Hydrogen widens "query fan-out", meaning it runs parallel searches across related subtopics. Google also says answers will show more images, comparison tables and a broader set of links, "including sites users may not have found before."

Google gives no traffic figures, so treat the "more links" claim as unproven until you see it in your own Search Console data. Long, multi-part queries are where fan-out pulls in niche pages. Clear, specific pages with comparison-friendly structure (tables, explicit specs, pricing) are the likeliest to be picked up.

What to do now

• On 3.8 Flash in production: run your eval set on Hydrogen at medium and high. Compare cost per completed task, not cost per token, and fix the breaking parameters above before switching traffic.
• On Argon for routine work: Hydrogen at high scores about the same on Google's index at a lower price, and Argon's introductory pricing ends November 30. Consider moving high-volume calls to Hydrogen and keeping Argon or Carbon for planning.
• Cost-sensitive, simple tasks: classification, extraction and short replies may still be much cheaper on Haiku-class models or open-weight options. Don't upgrade by default.
• Watch next: independent benchmarks beyond Google's tables, published free-tier rate limits, whether the Flash price change on January 1, 2027 is confirmed in Google's own pricing docs, and whether explicit caching returns.