Gemini 4 Hydrogen vs 3.8 Flash: What Breaks When You Migrate

Gemini 4 Hydrogen vs 3.8 Flash: What Breaks When You Migrate

Google shipped Gemini 4 Hydrogen on November 12, and it's already the default model for the Antigravity agent in Gemini Managed Agents. The model ID is gemini-4-hydrogen, it's generally available, and Google calls it a major standalone upgrade for teams migrating from Gemini 3.8 Flash.

Before you swap one string for another, read the migration notes. Some of your old request parameters now get ignored, and one now throws an error. And for the next seven weeks, Hydrogen costs twice what 3.8 Flash does.

Here's what changed, what it costs, and when switching actually pays off.

Gemini 4 Hydrogen Key Features and Technical Breakdown

Hydrogen is Google's fast tier, sitting below the frontier models Gemini 4 Carbon and Gemini 4 Argon. It takes text, image, video, audio, and PDF in, but returns text only. No Live API, no audio generation, no image generation.

The headline specs:

• Context: a unified 1,048,576-token limit.
• Thinking levels: low, medium, high (the default), and xhigh.
• Tools: code execution, computer use, file search, function calling, Search and Maps grounding, structured outputs, and URL context.
• Free tier: yes, rate-limited. Free-tier content is used to improve Google's products, and there's no Search or Maps grounding on it.

Gemini 4 Hydrogen pricing

Per Google's pricing page, the standard rate is $1.50 per million input tokens and $7.50 per million output tokens. It's flat across the full context window, so there's no surcharge past some long-context threshold.

Grounding is extra on paid tiers. You get 5,000 free grounded requests a month, shared across Gemini 3 and newer models, then it's $14 per 1,000.

The benchmarks (Google's numbers)

Every figure here comes from Google's launch post. Google hasn't published methodology for them, so treat them as claims until independent runs land.

• Terminal-Bench 4.1: 57%, up from 18.2% for Gemini 3.8 Flash.
• Artificial Analysis Intelligence Index: 51 at high, 54 at xhigh. Google lists Argon at 50.
• HLE-Diamond: 58.8%.
• Speed: finishes tasks 2.5 times faster than 3.8 Flash at its default high setting.

That Terminal-Bench jump is the number to watch. If it holds up outside Google's tests, Hydrogen is a different class of agent model than the Flash you're running now.

Gemini 4 Hydrogen vs Gemini 3.8 Flash: The Breaking Changes

This is the part that matters for existing code. Google's own What's new in Gemini 4 Hydrogen page lists changes that will bite code written for 3.8 Flash.

That last row deserves a closer look. Output really can reach 1,048,576 tokens, but only through long decode: set background: true and a max_output_tokens above 65,536. Set background: true without max_output_tokens and you still stop at 65,536.

Also, thinking tokens are billed and count toward max_output_tokens. With the default bumped to high, an unchanged prompt will likely burn more output tokens than it did on 3.8 Flash.

For multi-turn agents, carry state with previous_interaction_id. Don't copy a thought signature across parallel functionCall parts.

The Price Math Nobody's Doing Yet

Here's the catch on cost. Gemini 3.8 Flash is $0.75 input and $3.75 output today. Hydrogen is exactly double that.

That gap closes on January 1, 2027, when 3.8 Flash's price doubles to match Hydrogen's. So for the next seven weeks, Flash is the cheaper model. After that, you're paying the same rate for the older model.

Take an agent step with 200K input tokens and 20K output tokens:

That table assumes the same token counts on every model, which won't hold in practice. Hydrogen's high default probably generates more thinking tokens, while its claimed 2.5x speed and better agent scores could mean fewer retries. Measure on your own workload.

Carbon and Argon double their price on December 1, too. If you were planning to route cheap steps to a frontier model at the current $2/$10 rate, that window shuts on November 30.

Real-World Use Cases

Coding agents. The Terminal-Bench 4.1 score and the Antigravity default point here. Antigravity's Orchestrator switches between Carbon and Hydrogen inside a session, which suggests a split worth testing: Carbon for hard reasoning, Hydrogen for the bulk of steps.

Long-document jobs. Flat pricing across 1M tokens makes big-context runs predictable. Pair it with Batch at $0.75/$3.75 for overnight work.

Grounded search assistants. Search and Maps grounding work on paid tiers. Watch the $14 per 1,000 requests after the free 5,000, plus the 1,500-per-day cap on Search grounding.

Not for: voice agents or image generation. It's text out only.

Hard Limitations and Flaws

• No published rate limits or knowledge cutoff. The model page says Latest update: November 2026 and that's it.
• No explicit caching. If your cost model depended on pinned caches, you're relying on implicit caching now.
• No sampling control. Teams that tuned temperature for deterministic outputs lose that lever.
• The Gemini xhigh thinking level cuts both ways. Google itself warns it increases latency and token use and can cause overthinking on simpler prompts.
• Benchmarks are self-reported. No methodology, and partner speed claims come from early testing.
• Free-tier content is used to improve Google's products. Don't send customer data through it.

Verdict: Your Gemini 4 Hydrogen Migration Plan

3.8 Flash isn't deprecated. Google still lists it as stable. So you don't have to move this week, and for pure cost, you shouldn't move before January.

Here's the plan I'd follow:

1. Grep your code for the parameters that break or get ignored. Fix those first:

grep -rnE "minimal|cachedContents|caches\.create|temperature|top_p|top_k" src/

1. Pin the thinking level. Start with medium to match 3.8 Flash behavior, then test high.
2. Run your eval suite on both models and compare cost per successful task, not per token.
3. Use Batch for anything that isn't latency-sensitive. It halves the bill.
4. Switch production by late December, before 3.8 Flash's January 1 price bump removes its price advantage.

Short version? Hydrogen looks like the stronger agent model. Prove it on your evals before you pay double for it.