> ## Content Index
> Fetch the complete content index at: https://claudefox.com/llms.txt
> Use this file to discover other available public pages before exploring further.

# Gemini 4 Hydrogen vs Claude Sonnet 5.5 for coding: price, speed and benchmarks
- URL: https://claudefox.com/gemini-4-hydrogen-vs-claude-sonnet-5-5/
- Published: 2026-11-18T04:16:41.000Z
- Updated: 2026-11-18T04:16:47.000Z
- Description: Gemini 4 Hydrogen costs $1.50/$7.50 per million tokens against $2/$10 for Claude Sonnet 5.5. Google's own table has each model winning different coding tests, and an independent index puts them one point apart. Here are the published numbers, with sources and dates.
- Author: hupicake
- Tags: Research

Google released Gemini 4 Hydrogen on November 12 as its fast, low-cost model for everyday coding and agent work. For people who build with Claude, the obvious comparison is Claude Sonnet 5.5, which Anthropic released on September 28 for well-scoped coding and knowledge work. Both have a 1 million token context window. Google pitches Hydrogen as a fast execution model that works alone or alongside its larger Carbon and Argon models; Anthropic describes Sonnet 5.5 as fast, cost-efficient intelligence for well-scoped coding and knowledge work.

We have not run our own head-to-head test yet. This compares published numbers from Google's and Anthropic's tables, both price pages, and Artificial Analysis, an independent benchmarking site. Each number names its source; all were read on November 18, 2026.

## Which is cheaper, Gemini 4 Hydrogen or Claude Sonnet 5.5?

Hydrogen is 25% cheaper per token at list price, and Artificial Analysis measured it at about a third of Sonnet 5.5's cost per task at each model's top setting.

- Gemini 4 Hydrogen (Gemini API, paid tier): $1.50 per million input tokens, $7.50 per million output tokens including thinking tokens, flat across the 1M context. Cached input is $0.075 per million, and only implicit caching is supported.
- Claude Sonnet 5.5 (Claude API): $2 per million input tokens, $10 per million output tokens, flat across the 1M context at standard Claude API rates (US-only inference is 1.1 times that, and regional endpoints on cloud platforms add 10%). Cache hits cost $0.10 per million; writing to the 5-minute cache costs $2.50 per million. Anthropic advertises up to 90% savings with caching and 50% with batch processing.
- Hydrogen also runs on the Gemini API free tier at no charge. Google's pricing table says free-tier traffic is used to improve its products; paid-tier traffic is not.

On equal token counts the math is simple. A coding task that reads 200,000 tokens and writes 20,000 costs $0.45 on Hydrogen and $0.60 on Sonnet 5.5, before caching. Token counts do not carry over between vendors, though: each uses its own tokenizer, and Anthropic says the tokenizer in Claude 4.7 and later models produces about 30% more tokens for the same text than its earlier one. The models also think for different lengths, which is why the per-task figures below matter more.

- Artificial Analysis, Gemini 4 Hydrogen at xhigh thinking: $3.16 per Intelligence Index task, with 11.3 billion input tokens and 304 million output tokens across the index.
- Artificial Analysis, Claude Sonnet 5.5 at max effort: $10.02 per task, with 30.6 billion input tokens and 335 million output tokens.
- Sonnet 5.5 at max effort read about 2.7 times as many input tokens as Hydrogen at xhigh. Artificial Analysis ranks the two among the most verbose models it tracks: Sonnet 5.5 91st of 91 and Hydrogen 90th.

## Which scores higher on coding benchmarks?

It depends on the benchmark: in Google's own table Sonnet 5.5 leads on terminal coding and computer use, while Hydrogen leads on bug-fixing, knowledge work and reasoning tests.

Google's Gemini 4 Hydrogen table on deepmind.google compares Hydrogen with Claude Sonnet 5.5 and OpenAI's GPT-6.1 Terra. These are Google's figures, run with Google's methodology:

- Terminal-Bench 4.1 (cheating-adjusted): Sonnet 5.5 63.9%, Hydrogen 57.0%.
- OSWorld-2.1 (offline subset, computer use): Sonnet 5.5 83.9%, Hydrogen 77.6%.
- Terminal-Bench-Science 0.1: Sonnet 5.5 59.9%, Hydrogen 51.4%.
- SWE-sweep (find and repair bugs without breaking existing code): Hydrogen 13.5%, Sonnet 5.5 3.1%.
- Agents' Last Exam v1.1: Hydrogen 71.6%, Sonnet 5.5 60.4%.
- AutomationBench-AA: Hydrogen 76.1%, Sonnet 5.5 71.8%.
- HLE-Diamond: Hydrogen 58.8%, Sonnet 5.5 53.3%.
- Chartography: Hydrogen 66.6%, Sonnet 5.5 61.6%.
- MKA (multimodal knowledge work): Hydrogen 53.4%, Sonnet 5.5 52.0%.
- Vending-Bench 2 (final bank balance in a long-running business simulation): Hydrogen $14,549, Sonnet 5.5 $11,165.

Anthropic's Sonnet 5.5 table predates Hydrogen, so it has no Gemini 4 column. Its headline coding numbers are 70.6% on Terminal-Bench 4.0, 55.5% on CursorBench 4.0, and 52.1% on FrontierCode 1.1 at xhigh effort (46.2% at max). Anthropic's footnote explains the lower max score: at max effort Sonnet 5.5 more often ran Claude Code's code-review skill, which splits a review across many subagents, and in two cases Cognition examined this led to a timeout or to edits beyond the task's scope. The same table puts Sonnet 5.5 at 1844 on GDPval-AA v2.1, 1811 on AA-Briefcase v1.1 and 64.5% on Humanity's Last Exam with tools.

## Can you compare Google's and Anthropic's numbers directly?

Mostly no: the two vendors use different benchmark versions, effort settings and harnesses, so only a few rows line up.

- Terminal-Bench: Anthropic reports version 4.0 (70.6% for Sonnet 5.5); Google reports a cheating-adjusted version 4.1 (63.9% for Sonnet 5.5, 57.0% for Hydrogen). Those are different tests, not a disagreement.
- OSWorld 2.1: Anthropic lists Sonnet 5.5 at 80.1% on a partial set; Google lists it at 83.9% on an offline subset. Google's number for the Claude model is the higher one.
- Chartography: both vendors list Sonnet 5.5 at 61.6%, the one row where the figures match exactly. Anthropic labels it no tools.
- SWE-sweep, where Hydrogen leads 13.5% to 3.1%, does not appear in Anthropic's table, so there is no second measurement of Sonnet 5.5 on it.

## What does an independent benchmark say?

On the Artificial Analysis Intelligence Index the two are one point apart at their top settings, but Hydrogen finished the index's tasks in less than half the time and at about a third of the cost per task.

- Gemini 4 Hydrogen (xhigh): index score 54, ranked 23rd of 92 models; 14.8 minutes per index task; 174 output tokens per second; 42.69 seconds to the first answer token.
- Claude Sonnet 5.5 (max): index score 55, ranked 20th of 92; 32.3 minutes per index task; 147 output tokens per second; 290.29 seconds to the first answer token.
- At Hydrogen's default high setting, Google's announcement cites an index score of 51, one point above Gemini 4 Argon.

The index (version 5.0.1) combines nine evaluations, including Terminal-Bench 4.0 and HLE-Diamond. It is one site's harness and one set of settings; it does not tell you how either model behaves inside your own agent loop.

## How do context window, output limits and thinking settings compare?

Both read up to 1 million tokens; Hydrogen can produce much longer outputs through Google's long-decode mode, and both let you trade speed for reasoning depth.

- Context: Hydrogen's documented context window is 1,048,576 tokens (input and output combined). Sonnet 5.5's is 1M tokens.
- Output: Hydrogen returns up to 65,536 tokens per synchronous call by default. Going beyond that needs long decode (a background request or client-side continuation), up to 1,048,576 tokens shared with the input. Sonnet 5.5's maximum output is 128K tokens.
- Thinking: Hydrogen has four thinking levels, low, medium, high (the default) and a new xhigh. Google's docs say to start at high and keep xhigh for hard multi-step problems: it adds latency and tokens, and on simple prompts such as JSON extraction it can overthink or run out of max\_output\_tokens. Sonnet 5.5 uses adaptive thinking with a default effort of high; Anthropic's table also reports xhigh and max effort results.

## Are there free tiers or rate limits to plan for?

Hydrogen is usable on the Gemini API free tier, and paid Gemini projects start with spend-based limits that can stop a runaway agent.

- Gemini API rate limits apply per project, not per API key, and requests-per-day quotas reset at midnight Pacific time.
- A new paid project (Tier 1) has a spend-based limit of $10 per 10 minutes and a billing tier cap of $250\. Tier 2 (after $100 paid and 3 days) raises the spend limit to $200 per 10 minutes and the billing cap to $2,000\. Going over the spend limit returns a 429 RESOURCE\_EXHAUSTED error.
- Outside the APIs, Hydrogen is in the Gemini app and Google Search; Sonnet 5.5 is in Claude.ai on web, iOS and Android.

## Which should you pick if you build with Claude?

Google's own table favors Sonnet 5.5 for terminal-heavy agent work and computer use, so keep it there, and consider Hydrogen for high-volume, cost-sensitive or multimodal jobs where its lower price and faster task times matter.

- Lean toward Sonnet 5.5 for long tool-using sessions in a terminal: it leads Hydrogen on Google's own Terminal-Bench 4.1 and OSWorld rows. Anthropic's table has no Hydrogen column, so this rests on Google's numbers.
- Lean toward Hydrogen for batch work billed per token, for jobs with video or audio input (both are listed as supported inputs on its model page), or when you want a free tier to prototype on.
- Our suggestion, not a vendor recommendation: neither table shows how the models behave on your codebase, so try both on a handful of your own past tasks at the settings you would actually use, and record cost as well as whether the result was usable. Use a paid Gemini project for private code, since free-tier traffic is used to improve Google's products.

Hydrogen is a Gemini API model, so in Claude Code you would call it as an outside tool; our setup guide, linked below, covers the routes Google documents. Within Google's lineup, Hydrogen is 25% cheaper per token than Gemini 4 Carbon until November 30; from December 1, when Carbon moves to $4 and $20, Hydrogen's per-token price is 62.5% lower.

## FAQ

How much does Gemini 4 Hydrogen cost? On the Gemini API paid tier, $1.50 per million input tokens and $7.50 per million output tokens including thinking, with cached input at $0.075 per million. It is free on the free tier.

What are the context and output limits? Hydrogen: 1,048,576 tokens shared by input and output, 65,536 output tokens per synchronous call by default. Sonnet 5.5: 1M context and 128K maximum output.

Is Gemini 4 Hydrogen better than Claude Sonnet 5.5 at coding? Not across the board. Google's own table has Sonnet 5.5 ahead on Terminal-Bench 4.1 (63.9% to 57.0%) and Hydrogen ahead on SWE-sweep (13.5% to 3.1%).

## Related on claudefox

- [Gemini 4 Hydrogen costs $1.50/$7.50, and it only gets much cheaper than Carbon on December 1](https://claudefox.com/gemini-4-hydrogen-pricing-and-orchestrator/)
- [How to use Gemini 4 Hydrogen alongside Claude Code: gemini-api CLI, Docs MCP and skills](https://claudefox.com/gemini-4-hydrogen-claude-code-setup/)
- [Gemini 4 Carbon vs Claude Opus 5.5 and Fable 5.2: benchmarks and pricing compared](https://claudefox.com/gemini-4-carbon-vs-claude-opus-5-5-fable-5-2/)
- [The 'Haiku builds, Opus advises' setup for Claude Code](https://claudefox.com/the-haiku-builds-opus-advises-setup-for-claude-code/)

## What we have not tested

We have not measured either model ourselves. Every number here was published by Google, Anthropic or Artificial Analysis, and vendors choose which benchmarks to show. If we run our own comparison, we will update this page and date the change.

Source: [Introducing Gemini 4 Hydrogen (Google, Nov 12, 2026)](https://blog.google/innovation-and-ai/models-and-research/gemini-models/introducing-gemini-4-hydrogen/?ref=claudefox.com), [Gemini models: performance tables (Google DeepMind)](https://deepmind.google/models/gemini/?ref=claudefox.com), [Gemini API pricing (Google)](https://ai.google.dev/gemini-api/docs/pricing?ref=claudefox.com), [Gemini 4 Hydrogen model page (Google AI Studio docs)](https://aistudio.google.com/docs/models/gemini-4-hydrogen?ref=claudefox.com), [Gemini thinking levels (Google AI Studio docs)](https://aistudio.google.com/docs/thinking?ref=claudefox.com), [Gemini API rate limits (Google AI Studio docs)](https://aistudio.google.com/docs/rate-limits?ref=claudefox.com), [Claude Sonnet 5.5 (Anthropic)](https://www.anthropic.com/claude/sonnet?ref=claudefox.com), [Claude models overview (Anthropic docs)](https://platform.claude.com/docs/en/about-claude/models/overview?ref=claudefox.com), [Claude API pricing (Anthropic docs)](https://platform.claude.com/docs/en/about-claude/pricing?ref=claudefox.com), [Gemini 4 Hydrogen (Artificial Analysis)](https://artificialanalysis.ai/models/gemini-4-hydrogen?ref=claudefox.com), [Claude Sonnet 5.5 (Artificial Analysis)](https://artificialanalysis.ai/models/claude-sonnet-5-5?ref=claudefox.com)

This story first appeared in the [claudefox daily briefing for November 18, 2026](https://claudefox.com/daily-2026-11-18/).