Gemini 4 Hydrogen lands as Google's launch prices end. Here's what a Gemini coding agent costs from December 1

Published November 13, 20264 min read

Google announced the model on its blog and published a model card the same day. The price changes are on the Gemini API pricing page, which we checked on November 13. We haven't run Hydrogen on our own code, so the benchmark numbers below are Google's.

What shipped

• Model ID: gemini-4-hydrogen, the ID used in Google AI Studio and on the pricing page.
• Price: $1.50 input and $7.50 output per million tokens, flat across the full context window. Cached input is $0.075. Batch and Flex are half price.
• Context: up to 1 million tokens, shared between input and output, per the model card. Synchronous output is capped at 64K tokens by default.
• Thinking levels: low, medium, high (the default), and a new xhigh tier.
• Knowledge cutoff: March 2026.
• Where: the Gemini API and AI Studio on all tiers, Android Studio, and Google Antigravity for paid AI Pro and Ultra subscribers.

Google's blog post doesn't mention Gemini CLI, Jules, or Vertex AI pricing. If you reach Gemini through one of those, check its own model list before assuming Hydrogen is there.

The prices that change

Gemini 4 Argon went GA on October 13 at launch pricing with an end date. That date is now close. The pricing page lists these standard-tier rates per million tokens, input and output:

Caching, Batch, Flex, and Priority rates double on the same dates. Context cache storage for 3.8 Flash goes from $0.50 to $1.00 per million tokens per hour. Google gives no reason on the pricing page, and we found none elsewhere.

Note the last two rows. From January, 3.8 Flash costs exactly what Hydrogen costs today. Google is moving its cheap tier up, and Hydrogen is the model it wants you on at that price.

What Hydrogen does on coding

Google's claims are mostly about speed. At its default high setting, Hydrogen finishes tasks about 2.5x faster than 3.8 Flash. Early partners saw 2x to 2.8x faster runs and 12% to 23% fewer tokens.

Two things stand out. The jump from 3.8 Flash on Terminal-Bench is large, so Hydrogen isn't a light refresh. And Sonnet 5.5 still leads on Terminal-Bench, the benchmark closest to what a terminal coding agent does all day.

The model card reports no SWE-bench or LiveCodeBench numbers. Google also says Hydrogen uses 24% less quota than Carbon in Antigravity's orchestrator, with near-Carbon task success, but gives no success rate to check.

What an agent run costs

Take one agent task that reads 2 million input tokens and writes 200,000 output tokens, with no caching. That's a mid-sized refactor in a terminal agent. At standard rates:

Caching changes the picture a lot. Agents resend the same repo context on every turn, so most input tokens can hit the cache. At Hydrogen's $0.075 cached rate, 1.8 million cached input tokens cost $0.135 instead of $2.70. If you aren't getting cache hits today, fix that before you compare models.

Hydrogen only supports implicit caching, per Google's docs as summarized by APIYI, with a 4,096-token minimum prefix. Keep your system prompt and repo context at the start of the request and stable between turns.

What to do before December 1

1. Find your Argon and Carbon traffic. Pull last month's Gemini API usage by model. Double the Argon and Carbon lines; that's your December bill if nothing changes.
2. Trial Hydrogen on the Carbon workloads first. Google pitches it as near-Carbon on agent tasks. Run ten of your own closed tickets through both and compare diffs, not just pass rates.
3. Keep Argon for the hard work. If Hydrogen fails a task type, paying $4 / $20 for Argon on that slice is cheaper than a failed run you redo by hand.
4. Move batch jobs to Batch or Flex. Code review sweeps, test generation, and doc updates rarely need an answer in seconds, and those tiers are half price.
5. Re-check other vendors. At $4 / $20, Argon costs the same per token as Claude Opus 5.5 does in Cursor's published API list prices. If you picked Gemini on price, that reason expires on December 1.
6. Watch your editor's model list. Tools that bill Gemini at API list price, such as Cursor, may pass the increase on. None had announced changes when we checked.

For a sense of what Argon does on long unattended runs, see our write-up of Google's 11-hour Argon run.

Sources

• Google: Introducing Gemini 4 Hydrogen (November 12, 2026): price, thinking levels, speed claims, benchmarks, and availability.
• Google DeepMind: Gemini 4 Hydrogen model card: context window, output limits, knowledge cutoff, and the Sonnet 5.5 and GPT-6.1 Terra comparisons.
• Gemini API pricing (checked November 13, 2026): current and future rates for Hydrogen, 3.8 Flash, Argon, and Carbon.
• APIYI: Gemini 4 Hydrogen vs Flash, Argon, and Carbon pricing: implicit caching and minimum prefix.
• Cursor docs: Models: Claude Opus 5.5 API list price for comparison.