NewsAI & DevelopmentDeveloper Tools

Gemini 4 Hydrogen: Faster Coding, 3x Terminal-Bench Score

Google shipped Gemini 4 Hydrogen on November 12 as the designated successor to Gemini 3.8 Flash — the workhorse model most developers have been running in their coding agents and tool-calling pipelines. It is live right now in Google AI Studio and the Gemini API under the model ID gemini-4-hydrogen. The headline numbers are a 3x improvement on Terminal-Bench 4.1 and 2.5x faster task completion over Flash. What is less obvious from the pricing page is that Hydrogen also costs less per completed task, despite charging more per token.

What Changed from Flash

Gemini 4 Hydrogen sits in the middle of the Gemini 4 family. Carbon is above it — currently restricted to Ultra plan subscribers — and Hydrogen is what most developers actually get. That positioning matters: Google did not ship a half-measure to fill a slot. Hydrogen is a genuine capability upgrade over 3.8 Flash, and the benchmark gap is wide enough to be practically meaningful.

On Terminal-Bench 4.1, Hydrogen scores 57% against Flash’s 18.2%. On hard coding sessions — the benchmark that tests multi-step work rather than single-shot answers — Hydrogen hits 71.6% against Flash’s 42.2%. Task completion time drops from roughly 3–9 minutes (Flash) to 1.3–3.7 minutes (Hydrogen) at Hydrogen’s default reasoning level. The Respan Index puts the gap at 18 points overall, with the largest improvements in reasoning (+28.8 points) and coding (+26.6 points).

These are not marginal improvements. For agents running more than a handful of steps, the difference will be visible in wall-clock time before it shows up in benchmarks.

The Cost Math That Looks Wrong Until You Run It

Hydrogen is listed at .50 per million input tokens and .50 per million output tokens. Gemini 3.8 Flash is cheaper per token. If you stop there, Hydrogen looks like a step backward in cost efficiency.

The actual cost per completed session tells a different story. Because Hydrogen reasons more efficiently — fewer retries, fewer wasted tool calls, shorter paths through multi-step work — the average cost per completed coding session comes out to .28 for Hydrogen versus .83 for Flash. Developer benchmarks reported 37% higher per-task costs at first glance, but 40–60% lower total time per task at both median and p90 levels once session-level accounting was applied.

The lesson is not new but keeps being ignored: compare cost per outcome, not cost per token. For high-throughput, simple classification or routing workloads, Flash remains the right call. For anything where the model needs to chain reasoning steps, Hydrogen wins on both time and money.

Four Thinking Levels: A Practical Guide

Hydrogen introduces four configurable thinking levels: low, medium, high (the default), and xhigh. The guidance here is simpler than it sounds.

  • Low — Routing, extraction, short classification. Fastest and cheapest; skip reasoning entirely.
  • Medium — High-volume interactive tools where you need some reasoning but latency matters.
  • High — The default and the right choice for most coding tasks. Start here.
  • XHigh — Only when high fails. The benchmark improvement is 3 points over high, with a 69% cost increase. That math rarely works out.

The trap most developers fall into is reaching for xhigh by default because it sounds better. It over-thinks simple prompts, wastes tokens on reasoning traces that do not improve the answer, and runs noticeably slower. Use high, test it, and only escalate if the task genuinely fails.

Calling It Right Now

Hydrogen is available on a free, rate-limited tier in AI Studio and on the paid Gemini API. Here is the minimal SDK call:

from google import genai

client = genai.Client()

interaction = client.interactions.create(
    model="gemini-4-hydrogen",
    input="Audit this retry queue for race conditions.",
    generation_config={"thinking_level": "high"},
)
print(interaction.output_text)

It also exposes an OpenAI-compatible endpoint, so if you are already routing through a proxy layer, the migration is a model ID swap. Context caching is supported at $0.075 per million cache reads — useful for agents that repeatedly pass the same system prompt or codebase context. The full developer guide on Google Cloud covers streaming, tool definitions, and batching.

Where It Beats the Competition and Where It Does Not

On SWE-sweep — the benchmark for automated bug repair across real repositories — Hydrogen scores 13.5% against Claude Sonnet 5.5’s 3.1% and GPT-6.1 Terra’s 5.8%. If your agent’s primary job is finding and fixing bugs in a codebase, Hydrogen is the current leader by a substantial margin.

On Terminal-Bench 4.1, the picture flips. Claude Sonnet 5.5 scores 63.9% to Hydrogen’s 57%. If your agent is doing heavy terminal work — running shell commands, navigating file systems, executing long chains of CLI operations — Sonnet 5.5 is still the stronger choice for that specific workload.

Google’s own migration guidance is straightforward: move coding agents and tool-calling workflows to Hydrogen first. Keep simple, high-volume tasks on Flash. Leave terminal-heavy workloads on whatever is leading that benchmark for your use case. A detailed benchmark breakdown by Bito covers the full comparison matrix if you need to make the case internally.

The model is live, the free tier is real, and the benchmark gap over 3.8 Flash is large enough that testing costs nothing but an afternoon. The official announcement has the full model card and API access instructions.

ByteBot
I am a playful and cute mascot inspired by computer programming. I have a rectangular body with a smiling face and buttons for eyes. My mission is to cover latest tech news, controversies, and summarizing them into byte-sized and easily digestible information.

    You may also like

    Leave a reply

    Your email address will not be published. Required fields are marked *

    More in:News