AI/Nov 16 edition/Reporting & analysis
ModelsAgentsBusinessCoding

ModelsArchitectures & capability

Google launches Gemini 4 Hydrogen, a fast-tier model that edges Gemini 4 Argon on Artificial Analysis's index at $1.50/$7.50 per million tokens

Google's Gemini 4 Hydrogen, the successor to Gemini 3.8 Flash, scores 51 on Artificial Analysis's Intelligence Index at its default setting, edging the larger Gemini 4 Argon, but its $1.50/$7.50 per-million-token pricing places it well above rival small models.

blog.google (article, date unverified): At its default high thinking level, Hydrogen scores 51 on Artificial Analysis's Intelligence Index, and 54 at xhigh. [1] [2] [3]

SOURCES[1] Introducing Gemini 4 Hydrogen (Google) blog.google[2] Artificial Analysis: Gemini 4 Hydrogen artificialanalysis.ai[3] VentureBeat: faster, smarter, pricier Flash successor venturebeat.com

AI-assisted, not human-edited. Methodology Corrections Who owns this site

KEY FACTS
What happened
At its default high thinking level, Hydrogen scores 51 on Artificial Analysis's Intelligence Index, and 54 at xhigh. [1] [2] [3]
Original source
blog.google article, company source
Brief by
AT Pulse research desk AI-assisted; publisher and contact on the desk page
When
Original publication date unverified
Evidence
9 sources across 8 sites
71/100StrongHow it’s measured

Hacker News · 572 likes · 353 comments

Reach blends public engagement (views, likes, comments, shares, stars) from the platforms where the story or its original source lives: YouTube, Hacker News, GitHub, Substack and similar. A video and the article written from its transcript share one score, taken from the video. Platforms with no data are left out rather than counted as zero, and the age of a story never lowers its score.

Illustration from Google: Google launches Gemini 4 Hydrogen, a fast-tier model that edges Gemini 4 Argon on Artificial Analysis's index at $1.50/$7.50 per million tokens
Image: Google — Original article ↗
THE CORE IDEAS4 TAKEAWAYS
01

At its default high thinking level, Hydrogen scores 51 on Artificial Analysis's Intelligence Index, and 54 at xhigh. That edges the larger Gemini 4 Argon (50) and is well above Gemini 3.8 Flash (31). Artificial Analysis's own listing matches Google's figures. [1] [2] [3]

02

Hydrogen costs $1.50/$7.50 per million input/output tokens, with one flat price across its 1M-token window. That is roughly 15 times the cheapest rival small models. Gemini 3.8 Flash moves to the same rates on January 1, 2027, which removes its cost advantage. [1] [3]

03

Capability depends heavily on thinking level: 27 at low, 41 at medium, 51 at high and 54 at xhigh. But xhigh is slow, averaging about 14.9 minutes per index task. It also added no gain on Terminal-Bench 4.0 while raising cost, so the setting needs tuning for each task. [2] [3]

04

Google positions Hydrogen as the executor in a three-tier Gemini 4 lineup, running worker and verifier agents while Carbon plans and coordinates. Alongside it, Google launched Orchestrator in its Antigravity coding tool to switch between models within one session. [1] [7]

WHY IT MATTERS

the index scores, published prices and scheduled Flash price change are documented. Implication: teams moving off Flash should budget by cost per completed task and tune thinking levels rather than defaulting to xhigh.

Executive brief

Google's new fast-tier model, Gemini 4 Hydrogen, outscores its larger sibling Gemini 4 Argon on Artificial Analysis's Intelligence Index: 51 vs. 50 at the default setting, and 54 at its top "xhigh" setting (Google). Artificial Analysis lists the same 51 and 54 (Artificial Analysis). Hydrogen launched on November 12, 2026 as the successor to Gemini 3.8 Flash. It costs $1.50/$7.50 per million input/output tokens. Small models from rival vendors list at about $0.10/$0.50 (VentureBeat). For buyers, that makes it a mid-priced execution model rather than a cheap one.

What changed and event timeline

  1. Carbon debuts in limited access

    Gemini 4 Carbon, Google's frontier model, first went to a small group. Search-result summaries say that group was cybersecurity defenders. Carbon later became the model Hydrogen is said to be distilled from ().

  2. Carbon reaches paying customers

    Carbon became generally available to paying users of the Gemini API and Gemini for business. Its introductory price of $2/$10 rises to $4/$20 after November 30 ().

  3. Hydrogen launches globally

    Google released Hydrogen across AI Studio, the Gemini API, Gemini for business, the Gemini app and Search. It also launched Orchestrator, a feature in its Antigravity coding tool that switches between models within one session ().

  4. Consumer plans reshuffled

    Hydrogen became the headline model of the $4.99 Google AI Plus plan, and Pro models were reported removed from that plan. Google's own subscription pages disagree on this ().

  5. (scheduled): Flash's price rises

    Gemini 3.8 Flash's introductory $0.75/$3.75 pricing ends, and Flash moves to Hydrogen's $1.50/$7.50 rates ().

Capabilities and access

  • Model ID: gemini-4-hydrogen (VentureBeat).
  • Context and pricing: 1M-token input, up to 1M output tokens, one flat price across the whole window.
  • Caching: cached input costs $0.075/M (DataCamp).
Read the full section
  • Model ID: gemini-4-hydrogen (VentureBeat).
  • Context and pricing: 1M-token input, up to 1M output tokens, one flat price across the whole window. Four thinking levels: low, medium, high (default) and xhigh (Google).
  • Caching: cached input costs $0.075/M (DataCamp).
  • Free users: cannot pick Hydrogen manually. Auto-selection routes most prompts to Flash-Lite and sends harder ones to Hydrogen (Notebookcheck).
  • Antigravity: limited to paid AI Pro and Ultra subscribers (Google developers blog).

Technical analysis for researchers and developers

  • Architecture: Hydrogen is described as distilled from Carbon (DataCamp). The model card excerpt does not describe the architecture or the distillation method.
  • Cost and latency: at xhigh, Artificial Analysis measured 173 output tokens/s, 14.9 minutes per task, 304M output tokens across its index, and $3.16 per index task.
  • Thinking levels: scores rise sharply by level: 27 (low), 41 (medium), 51 (high), 54 (xhigh). Teams should tune the thinking level for each task.
Read the full section
  • Architecture: Hydrogen is described as distilled from Carbon (DataCamp). The model card excerpt does not describe the architecture or the distillation method. It points to a separate methodology page for its evaluations.
  • Cost and latency: at xhigh, Artificial Analysis measured 173 output tokens/s, 14.9 minutes per task, 304M output tokens across its index, and $3.16 per index task. It rates the model slower than average per task (Artificial Analysis).
  • Thinking levels: scores rise sharply by level: 27 (low), 41 (medium), 51 (high), 54 (xhigh). However, xhigh showed no gain on Terminal-Bench 4.0 while raising cost (VentureBeat). Teams should tune the thinking level for each task.

Claims and evidence

  • Intelligence Index of 51/54: Google's figure, matched by Artificial Analysis's own listing (AA). Gemini 3.8 Flash scored 31 on the same index (VentureBeat).
  • Terminal-Bench: Google reports 57% on version 4.1 (Google). Artificial Analysis measured 54% on version 4.0, behind Claude Sonnet 5.5 at 64% (DataCamp).
  • Vendor-only claims: the remaining results are Google-reported and not independently corroborated.
Read the full section
  • Intelligence Index of 51/54: Google's figure, matched by Artificial Analysis's own listing (AA). Gemini 3.8 Flash scored 31 on the same index (VentureBeat).
  • Terminal-Bench: Google reports 57% on version 4.1 (Google). Artificial Analysis measured 54% on version 4.0, behind Claude Sonnet 5.5 at 64% (DataCamp). The benchmark versions differ, so the two numbers are not directly comparable.
  • Vendor-only claims: the remaining results are Google-reported and not independently corroborated. These include 2.5x faster than 3.8 Flash, more than 5x fewer fabricated claims, HLE-Diamond 58.8%, MKA 53.4%, Orchestrator's 24% quota savings, and partner reports of 12–23% fewer tokens (Google).

Context and prior work

The Gemini 4 lineup has three tiers: Carbon for the hardest frontier work, Argon as a large general-purpose model, and Hydrogen as the fast tier (DataCamp). Carbon coordinates the work, while Hydrogen runs worker and verifier agents (Google). Rival small models include Claude Haiku 5.5, GPT-6 Luna, Meta Muse Ember and DeepSeek V4.1 Flash (VentureBeat).

Read the full section

The Gemini 4 lineup has three tiers: Carbon for the hardest frontier work, Argon as a large general-purpose model, and Hydrogen as the fast tier (DataCamp). Hydrogen follows the pattern of earlier Flash models: a smaller model distilled from the frontier one and given the broadest distribution. What's new this generation is the explicit planner/executor design. Carbon coordinates the work, while Hydrogen runs worker and verifier agents (Google). Rival small models include Claude Haiku 5.5, GPT-6 Luna, Meta Muse Ember and DeepSeek V4.1 Flash (VentureBeat).

Limitations, safety and contested findings

  • Safety figures (Google internal): up to 81% less overreach in agentic tasks, 12x fewer misaligned actions, 2.3x more resilience on the Gray Swan prompt-injection benchmark, and 99.9% readable chain-of-thought across more than 200,000 audited turns (Google).
  • Conflicting honesty figures: Google's post gives a 94% agent-honesty rate.
  • Terminal-heavy work: Artificial Analysis says Hydrogen is not the best model for terminal-heavy work (DataCamp).
Read the full section
  • Safety figures (Google internal): up to 81% less overreach in agentic tasks, 12x fewer misaligned actions, 2.3x more resilience on the Gray Swan prompt-injection benchmark, and 99.9% readable chain-of-thought across more than 200,000 audited turns (Google).
  • Conflicting honesty figures: Google's post gives a 94% agent-honesty rate. DataCamp reports 96.7% for the same evaluation, plus an 8% hallucination rate on AA-Omniscience (DataCamp).
  • Terminal-heavy work: Artificial Analysis says Hydrogen is not the best model for terminal-heavy work (DataCamp).

Business and practitioner implications

  • Budget: Hydrogen costs about 15x more per token than the cheapest rival tiers.
  • Flash users: the January 2027 Flash price rise removes Flash's cost advantage over Hydrogen, so migrating is the default path.
  • Routing: pairing Carbon with Hydrogen (manually or through Orchestrator) is the intended pattern.
Read the full section
  • Budget: Hydrogen costs about 15x more per token than the cheapest rival tiers. Even at medium thinking, its estimated cost per task beats Haiku, Luna and Ember's top settings on capability but exceeds their cost (VentureBeat).
  • Flash users: the January 2027 Flash price rise removes Flash's cost advantage over Hydrogen, so migrating is the default path.
  • Routing: pairing Carbon with Hydrogen (manually or through Orchestrator) is the intended pattern. Measure cost per completed task, not cost per token, given Hydrogen's long xhigh run times (AA).
FOLLOW THE EVIDENCE

The source trail.

Sources (9)
A LITTLE LESS NOISE. A LOT MORE CONTEXT.

Stay curious.
Follow the evidence.

Independent perspectives, the original sources, and room for the questions that don't have easy answers.

How we build the brief
Connect with us

Find us where you already read.