News · 2026-11-13

Gemini 4 Hydrogen ties the leaders on one index, but its token bill and a changing Terminal-Bench version complicate Google's speed pitch

Google DeepMind released Gemini 4 Hydrogen on November 12, 2026, the fast, execution-focused tier of the Gemini 4 family, priced at $1.50 per million input tokens and $7.50 per million output tokens. Google says its default setting scores 51 on the Artificial Analysis Intelligence Index against 50 for the larger Argon, and 54 at the new xhigh thinking level. Independent measurement by Artificial Analysis confirms the 54 but calls the model "very verbose" and slower than average per task.

Key facts

• Who and when: Google DeepMind, November 12, 2026, in the launch post; available in AI Studio, the Gemini API, Gemini for business, the Gemini app and Google Search.
• Price and window: $1.50 input, $7.50 output per million tokens, flat across a 1M-token window, with outputs up to 1M tokens.
• Independent score: 54 at xhigh versus a median of 41 for comparable models, per Artificial Analysis, which logged 304 million output tokens for the index run against a median of 72.8 million.
• Thinking levels: low, medium, high (the default) and a new xhigh.

What Google claims

Google positions Hydrogen as the execution tier, with Carbon and Argon reserved for "deep, long-horizon reasoning." Against its own Gemini 3.8 Flash, the post says tasks finish 2.5 times faster at the high setting, Terminal-Bench 4.1 climbs from 18.2% to 57%, and the model is "cutting fabricated claims by more than 5x," with no number given for the baseline. Partner speedups run 2x to 2.8x with 12% to 23% fewer tokens. Google also reports a safety evaluation: overreach in agentic tasks down up to 81%, chain-of-thought legibility of 99.9% across more than 200,000 audited turns, and more than 2.3 times better resistance on the Gray Swan indirect-prompt-injection benchmark. All of these are Google's internal numbers.

What the independent view says

Artificial Analysis puts Hydrogen at xhigh at 54, near the top, at a cost of $3.16 per index task and 13.7 minutes per task against a 13.0 median. Output speed is 188 tokens per second but the time to the first answer token is 40.55 seconds. The headline contradiction with Google's 2.5x claim is mostly a baseline difference: Google compares to Gemini 3.8 Flash at high, while Artificial Analysis measures xhigh against the median of comparable models.

A secondary source, the Latent Space AINews recap citing Artificial Analysis and others, adds that 54 is one point behind Claude Sonnet 5.5 and GPT-6.1 Sol at their maximum settings, that cost is about 32% of Sonnet's, and that Hydrogen's hallucination rate on a knowledge test is 8% against 43% and 52%. Those figures are not on the Artificial Analysis page and are secondhand. The same recap shows Hydrogen at 54% on Terminal-Bench 4.0 against 64% for Sonnet. That does not contradict Google's 57%, because Google's figure is on version 4.1. Google's own two posts disagree on the version (the Antigravity post says 4.1, the DeepMind post says 4.0), so the exact version is unresolved.

The Orchestrator bet

The more interesting product idea is the Antigravity Orchestrator. Google's Antigravity post criticizes static "model routers" that guess a task's difficulty from the first prompt, and describes a "dynamic, bidirectional" handoff within one trajectory. A stronger model such as Carbon starts read-only, explores dependencies, resolves ambiguities and writes a plan with explicit re-planning triggers; Hydrogen then writes edits, updates packages and runs tests, and hands back to the planner when it hits friction. Google claims "near-Carbon task success" with on average 24% less quota and 19% faster. For background on the alternative, see our lesson on model routing and cascades. The tier is narrower than first reported: the Carbon pairing rolls out to Google AI Ultra subscribers, while the Argon pairing also reaches paid Pro subscribers and enterprises through Gemini for business.

Caveats from testers

Single-source reports in the recap say xhigh burned 131,000 thinking tokens, about $1.08, on one image question without answering, and that one tester saw budget exhaustion on 1.8% of a text benchmark. Simon Willison reportedly calls the price only a 25% discount on Carbon and Argon, which are said to double in price on December 1; that is unverified. Commenters on Reddit described $1.50 and $7.50 as steep for a Flash-tier model and were confused by the naming.

Bottom line

Hydrogen is a split decision. It ties the leaders on the overall index and does well where the recap says hallucination is measured. It trails on terminal coding, burns many tokens, and costs more per task than its price per token suggests, a theme also covered in our piece on price per token versus price per task. The Orchestrator's headline gains are Google's own and use an inconsistent benchmark label, so they await outside testing.

Primary source, verified: read the paper →

Ground Truth. (2026, November 13). Gemini 4 Hydrogen ties the leaders on one index, but its token bill and a changing Terminal-Bench version complicate Google's speed pitch. Ground Truth. https://groundtruth.day/news/gemini-4-hydrogen-ties-the-leaders-on-intelligence-but-spends-heavily-on-tokens.html

@misc{groundtruth:gemini-4-hydrogen-ties-the-leaders-on-intelligence-but-spends-heavily-on-tokens,
  title  = {Gemini 4 Hydrogen ties the leaders on one index, but its token bill and a changing Terminal-Bench version complicate Google's speed pitch},
  author = {{Ground Truth}},
  year   = {2026},
  month  = {nov},
  url    = {https://groundtruth.day/news/gemini-4-hydrogen-ties-the-leaders-on-intelligence-but-spends-heavily-on-tokens.html}
}

Topics: google · gemini · gemini-4 · pricing · benchmarks · agents · antigravity