The AI arena is free today

Open Superagent
Back to blog
Model Release·Technical Deep Dive

Gemini 4 Hydrogen: Argon-Class Agent Work at $1.50 and $7.50, Flat to 1M

Google released Gemini 4 Hydrogen on Nov 12, 2026 at $1.50/$7.50 per 1M, flat to 1M tokens. Beats Sonnet 5.5 on 4 of 7 Google rows, trails on terminal and computer use. Self-reported.

Jonathan Chavez
Jonathan Chavez
Co-Founder @ LLM Stats
·15 min read
Gemini 4 Hydrogen: Argon-Class Agent Work at $1.50 and $7.50, Flat to 1M

Key Numbers

Gemini 4 Hydrogen · Nov 12, 2026

$0.00 / $0.00

per 1M tokens, flat to 1M context

input / output
0.0%

vs 60.4% Claude Sonnet 5.5

Agents' Last Exam v1.1
0.0%

vs 18.2% Gemini 3.8 Flash

Terminal-Bench 4.1
0.0%

vs 3.1% Sonnet 5.5, 5.8% GPT-6.1 Terra

SWE-sweep
0.0%

than Argon and Carbon list from Dec 1 ($4 / $20)

lower

Access status

Generally available Nov 12, 2026 on the Gemini API (free and paid), Vertex AI, and the Gemini app. Default thinking: high. Scores self-reported by Google.

Scores self-reported by Google from the Gemini 4 Hydrogen model card and launch table, November 2026. Not independently verified by third parties or by LLM Stats. Banner: Google.

At a Glance

Google released Gemini 4 Hydrogen on November 12, 2026 as its execution model for everyday coding and agent work. It is adapted from Gemini 4 Carbon, priced at $1.50 input / $7.50 output per million tokens, flat across a 1,048,576-token window, with a $0.075 implicit cache rate and a free tier. Google positions Carbon and Argon as the models for deep, long-horizon reasoning, and Gemini 4 Hydrogen as the model for rapid execution across diverse agentic workflows.

On Google's self-reported launch table it splits with Claude Sonnet 5.5 (ahead on Agents' Last Exam, SWE-sweep, Vending-Bench 2, and Chartography; behind on Terminal-Bench 4.1, OSWorld-2.1, and Terminal-Bench-Science) and beats GPT-6.1 Terra on five of six shared rows. Google's own safety assessment found no material capability increase over Argon, so treat Gemini 4 Hydrogen as Argon-class work at a lower price: a worker beside Carbon or Argon, not a replacement. Default thinking is high, so default calls bill more thinking tokens than Carbon or Argon at their medium default. See Google's launch post and model card.


What Gemini 4 Hydrogen Is

Gemini 4 Hydrogen sits under Carbon and Argon in Google's Gemini 4 lineup. Google's pitch is pairing: Carbon plans or resolves hard roadblocks; Gemini 4 Hydrogen handles code edits, parallel subagent tasks, and test runs. Models can switch across turns in one session with previous_interaction_id. The Orchestrator in Google Antigravity switches between Carbon and Gemini 4 Hydrogen in one session. Google says that on Terminal-Bench 4.0 it reached near-Carbon task success rates while using 24% less quota on average and finishing 19% faster. Google gives no absolute scores for that claim. The Antigravity agent in Gemini Managed Agents now uses Gemini 4 Hydrogen by default.

Availability spans the Gemini API (free and paid) in Google AI Studio and Android Studio; Gemini for business and Vertex AI; Google Antigravity for paid Google AI Pro and Ultra subscribers; the Gemini app model picker (free users via automatic model selection); Google Search for all users; and Gemini in Google Sheets starting with AI Pro and Ultra. It is also on OpenRouter as google/gemini-4-hydrogen. Knowledge cutoff is March 2026. No parameter count is published. Gemini 3.5 Flash-Lite remains Google's pick for classification, routing, and sub-second latency; Gemini 4 Hydrogen is the recommended upgrade from Gemini 3.8 Flash and earlier Gemini 3 and 2.5 models. Note: Carbon is now on the Gemini API paid tier (it was limited to cyber defenders when that post ran). See also our Argon launch post.

Family position · same-method rows

Argon-class work at a lower price

Gemini 4 Hydrogen $1.50 / $7.50

vs Argon and Carbon $4 / $20 from Dec 1

OSWorld-2.1 offline subset, partial

Argon 74.7%

Gemini 4 Hydrogen 77.6%

Carbon 81.5%

Vending-Bench 2

Argon $13,718

Gemini 4 Hydrogen $14,549

Carbon $16,484

Chartography (no tools)

Gemini 4 Hydrogen 66.6%

Argon 71.6%

Carbon 72.5%

Agents' Last Exam v1.1

different harness

Argon 69.0%

Gemini 4 Hydrogen 71.6%

Carbon 77.4%

Argon and Carbon values from Google's Gemini 4 Carbon comparison table (October 2026); Gemini 4 Hydrogen from Google's Gemini 4 Hydrogen table (November 2026). Same published methodology on the first three rows. Agents' Last Exam used a different harness for Gemini 4 Hydrogen (Antigravity CLI) than for Argon and Carbon (ALE-Claw), so treat that row as context, not a clean delta. All self-reported by Google.

Benchmarks

All scores below are self-reported by Googlefrom the November 2026 Gemini 4 Hydrogen model card and launch table. Gemini 4 Hydrogen scores are at the highest thinking setting except Terminal-Bench 4.1, which Google ran at high. Peer scores are as published in Google's table and may differ from each developer's own reports. Methodology notes live on Google's evals methodology page.

Google's Gemini 4 Hydrogen table · by job

Gemini 4 HydrogenSonnet 5.5

Agentic coding

SWE-sweep
+10.4 vs Sonnet 5.5

Gemini 4 Hydrogen in Google's Antigravity CLI harness; peers from the public leaderboard

Gemini 4 Hydrogen
13.5
Claude Sonnet 5.5
3.1
GPT-6.1 Terra
5.8
Terminal-Bench 4.1 (cheating-adjusted)
-6.9 vs Sonnet 5.5

Gemini 4 Hydrogen at high thinking, mini-swe-agent, average of 5 attempts

Gemini 4 Hydrogen
57.0
Claude Sonnet 5.5
leader
63.9
GPT-6.1 Terra
Not reported
Not reported

Knowledge work

Agents' Last Exam v1.1 (score)
+11.2 vs Sonnet 5.5

122 tasks, 5-hour window; Gemini 4 Hydrogen in Antigravity CLI, Sonnet in ALE-Claw run by Google; Terra from the public leaderboard

Gemini 4 Hydrogen
71.6
Claude Sonnet 5.5
60.4
GPT-6.1 Terra
69.8
Vending-Bench 2 (final bank balance)
+$3,384 vs Sonnet 5.5

Andon Labs leaderboard; dollars, own axis

Gemini 4 Hydrogen
$14,549
Claude Sonnet 5.5
$11,165
GPT-6.1 Terra
$9,107

Computer use

OSWorld-2.1 (82-task offline subset, partial score)
-6.3 vs Sonnet 5.5

Gemini 4 Hydrogen and Sonnet run by Google; Terra from its system card

Gemini 4 Hydrogen
77.6
Claude Sonnet 5.5
leader
83.9
GPT-6.1 Terra
74.2

Multimodal

Chartography (no tools)
+5.0 vs Sonnet 5.5

Sonnet from its system card

Gemini 4 Hydrogen
66.6
Claude Sonnet 5.5
61.6
GPT-6.1 Terra
63.2

Science and math

Terminal-Bench-Science 0.1
-8.5 vs Sonnet 5.5

Gemini 4 Hydrogen run by Google; peers from the public leaderboard

Gemini 4 Hydrogen
51.4
Claude Sonnet 5.5
leader
59.9
GPT-6.1 Terra
58.3
Self-reported by Google, November 2026. Gemini 4 Hydrogen at its highest thinking setting except Terminal-Bench 4.1 (high). Peer scores as published in Google's table, sourced from leaderboards and system cards, and may differ from each developer's own reports. SWE-sweep and Agents' Last Exam used different harnesses for Gemini 4 Hydrogen and its peers. OSWorld-2.1 is the 82-task offline subset, partial score. Not independently verified.

Compact table

Same seven usable rows from Google's self-reported table, with short method notes. GPT-6.1 Terra's Terminal-Bench 4.1 cell is Not reported.

BenchmarkGemini 4 HydrogenSonnet 5.5GPT-6.1 Terra
SWE-sweep13.5%3.1%5.8%
Terminal-Bench 4.157.0%63.9%Not reported
Agents' Last Exam v1.171.6%60.4%69.8%
Vending-Bench 2$14,549$11,165$9,107
OSWorld-2.1 offline partial77.6%83.9%74.2%
Chartography (no tools)66.6%61.6%63.2%
Terminal-Bench-Science 0.151.4%59.9%58.3%

Method notes: SWE-sweep used Google's Antigravity CLI for Gemini 4 Hydrogen and the public leaderboard for peers. Agents' Last Exam used Antigravity CLI for Gemini 4 Hydrogen, ALE-Claw for Sonnet (run by Google), and the public leaderboard for Terra. OSWorld-2.1 is the 82-task offline subset, partial score. Scores self-reported by Google.

MKA (Gemini 4 Hydrogen only)

Google also reports MKA (Multimodal Knowledge Agent) Climb WSR accuracy of 53.4% for Gemini 4 Hydrogen, run by Google. MKA is a new benchmark Google introduced this month. Peer results are not shown here. Score self-reported by Google.

Versus Gemini 3.8 Flash

Google's launch blog highlights Terminal-Bench 4.1 at 57% vs 18.2% for Gemini 3.8 Flash (more than tripling). Google also claims Gemini 4 Hydrogen is 2.5x faster at the high default and has more than 5x fewer fabricated claims than 3.8 Flash. No other 3.8 Flash scores are published for this comparison.

Honest reading

Versus Sonnet 5.5, Gemini 4 Hydrogen leads 4 of 7 rows (SWE-sweep +10.4, Agents' Last Exam +11.2, Vending-Bench 2 +$3,384, Chartography +5.0) and trails 3 (Terminal-Bench 4.1 by 6.9, OSWorld-2.1 by 6.3, Terminal-Bench-Science by 8.5). Versus GPT-6.1 Terra, it leads 5 of 6 shared rows and trails on Terminal-Bench-Science by 6.9; the Agents' Last Exam lead over Terra is thin (+1.8). The honest read: stronger on knowledge work, repo bug repair, and charts; behind Sonnet 5.5 in the terminal and on desktop computer use.

Google also says Gemini 4 Hydrogen is #5 on Agent Arena (arena.ai), #1 on Vals Tax Agent Bench (52.3%), and tied #1 on Vals Legal Research Bench (55.8%). Those are Google's statements about third-party leaderboards.


Pricing

Gemini API paid tier: $1.50 input, $0.075 cached input, $7.50 output per million tokens. Output includes thinking tokens. The rate is flat across the full 1M window. A free tier is available (rate-limited). Caching is implicit only for prompts of at least 4,096 tokens (95% off input). Explicit caching is not supported. Sampling parameters (temperature, top_p, top_k) are ignored.

Vertex AI global: Standard $1.50 / $0.075 / $7.50; Priority $2.70 / $0.135 / $13.50; Batch $0.75 / $0.0375 / $3.75. Non-global endpoints cost 10% more (Standard $1.65 / $0.0825 / $8.25). Vertex adds cache writes at $1.875 starting January 12, 2027. Grounding with Google Search: 5,000 free requests per month (shared across Gemini 3 and newer models), then $14 per 1,000. Full lists: Gemini API pricing and Vertex AI pricing.

Versus peers: Gemini 4 Hydrogen is 25% below Claude Sonnet 5.5 ($2 / $10) on input, output, and cache; 1.5x GPT-6.1 Terra at short context ($1 / $5), and cheaper than Terra on input (equal on output) above 272K; 25% below Argon and Carbon today ($2 / $10 through November 30) and 62.5% below them from December 1 ($4 / $20); 2x Gemini 3.8 Flash today ($0.75 / $3.75 through December 31) and equal to 3.8 Flash's list from January 1, 2027 ($1.50 / $7.50). Default thinking is high, so a default call may use more thinking tokens than Carbon or Argon at their medium default. Google warns the model may use more tokens at higher thinking levels.

List price rails · USD per 1M

output scale to $20

Current outputLater list (ghost)

GPT-6.1 Terra

$1.00 / $5.00 · $2.00 / $7.50 above 272K input

$5.00

out / 1M

in $1.00

Gemini 4 Hydrogen

$1.50 / $7.50 · flat to 1M

$7.50

out / 1M

in $1.50

Claude Sonnet 5.5

$2.00 / $10.00

$10.00

out / 1M

in $2.00

Gemini 4 Argon and Carbon

$2.00 / $10.00 · through Nov 30

then $4.00 / $20.00 from Dec 1

$10.00

out / 1M

in $2.00

Gemini 3.8 Flash

$0.75 / $3.75 · through Dec 31

then $1.50 / $7.50 from Jan 1, 2027

$3.75

out / 1M

in $0.75

List prices per 1M tokens from Google, Anthropic, and OpenAI pricing pages, November 16, 2026. Output includes thinking tokens. Cache reads: Gemini 4 Hydrogen $0.075, Sonnet 5.5 $0.10, Terra $0.05.

Thinking Levels

Gemini 4 Hydrogen supports four thinking levels: low, medium, high (default), and the new xhigh. Requests with minimalreturn an error. Google's guidance: start with high; reserve xhigh for hard multi-step problems; xhigh adds latency and tokens and can cause overthinking on simpler prompts. Carbon and Argon default to medium and cap at high. See Google's what's new notes and the developer blog.


Should You Switch?

Short decision guide from Google's migration notes and the self-reported table above.

FromGuidance
Gemini 3.8 FlashSwitch. Unified 1M window; default high instead of medium; minimal rejected; remove explicit caching and sampling parameters; declare tools natively; background execution for outputs past 64K; budget 2x per token until Dec 31, parity from Jan 1, 2027.
Claude Sonnet 5.5Consider Gemini 4 Hydrogen for knowledge work, document and chart-heavy agents, and repo bug repair at 25% less. Stay on Sonnet for terminal-heavy coding and computer use.
GPT-6.1 TerraGemini 4 Hydrogen scores higher on five of six shared rows for 1.5x the short-context price; above 272K the gap narrows.
Carbon or ArgonKeep them for planning and the hardest problems; move worker and verifier turns to Gemini 4 Hydrogen (no schema change; set thinking_level explicitly).

Safety

Versus Gemini 3.8 Flash (Google's launch blog): overreach in agentic tasks down up to 81%, under-service down up to 79%, misaligned actions and task shortcutting cut 12x across tool-use and coding evaluations, up to 2x less likely to bypass permission boundaries on impossible or blocked tasks, 94% on Google's internal agent honesty evaluation, 99.9% chain-of-thought legibility with zero deceptive reasoning edits across more than 200,000 audited turns, more than 2.3x more resilient on the Gray Swan indirect prompt injection benchmark, and HealthBench Pro false refusals under 0.4%.

Model card automated evals vs 3.8 Flash: text safety +0.80 points, multilingual +1.10, image-to-text +7.00, refusal tone -0.65, unjustified refusals +5.57 points (worse; Google says particularly for multimodal prompts).

Frontier Safety Framework: no material capability change vs Argon, so no full critical capability suite; targeted CBRN and cyber testing found the critical capability levels not reached, but the alert thresholds reached for biological and cyber risk, with mitigations deployed. The full Gemini 4 Hydrogen safety report is coming soon. Known limitations on the card: hallucinations, occasional slowness or timeouts, and more tokens at higher thinking levels.


FAQ

  • Google released Gemini 4 Hydrogen on November 12, 2026. The model id is gemini-4-hydrogen. It rolled out the same day on the Gemini API (free and paid), Vertex AI, Google Antigravity, the Gemini app, and Google Search.
  • On the Gemini API paid tier: $1.50 input, $0.075 cached input (implicit), and $7.50 output per million tokens, flat across the full 1,048,576-token window. A free tier is available (rate-limited). Vertex AI global Standard matches that list; Priority is $2.70 / $0.135 / $13.50; Batch is $0.75 / $0.0375 / $3.75. Non-global Vertex endpoints cost 10% more. That is 25% below Claude Sonnet 5.5 and 62.5% below the $4 / $20 that Gemini 4 Argon and Carbon move to on December 1.
  • Context is a unified 1,048,576-token window shared by input, thinking, and output. Synchronous responses return up to 65,536 tokens (streaming or not). Long decode extends output up to 1,048,576 tokens minus input via background: true or client-side continuation. Carbon and Argon allow 32,768 tokens per non-streaming REST hop.
  • low, medium, high (default), and the new xhigh. Requests with minimal return an error. Google says start with high, reserve xhigh for hard multi-step problems, and expect more latency and tokens at higher levels. Carbon and Argon default to medium and cap at high.
  • On Google's self-reported table, Gemini 4 Hydrogen leads 4 of 7 usable rows (Agents' Last Exam 71.6% vs 60.4%, SWE-sweep 13.5% vs 3.1%, Vending-Bench 2 $14,549 vs $11,165, Chartography 66.6% vs 61.6%) and trails on Terminal-Bench 4.1 (57.0% vs 63.9%), OSWorld-2.1 (77.6% vs 83.9%), and Terminal-Bench-Science (51.4% vs 59.9%). List price is 25% below Sonnet 5.5. See our Claude Sonnet 5.5 launch post.
  • On Google's self-reported table, Gemini 4 Hydrogen leads 5 of 6 shared rows and trails only on Terminal-Bench-Science (51.4% vs 58.3%). The Agents' Last Exam lead is thin (71.6% vs 69.8%). Terra's Terminal-Bench 4.1 cell is not reported. At short context Gemini 4 Hydrogen costs 1.5x Terra ($1.50 / $7.50 vs $1 / $5); above 272K input Terra bills the full request at $2 / $7.50. See our GPT-6.1 Terra launch post.
  • Google's Frontier Safety assessment found no material capability increase over Gemini 4 Argon. On the three same-method rows Google published for both, Gemini 4 Hydrogen is ahead of Argon on OSWorld-2.1 and Vending-Bench 2, and behind both Argon and Carbon on Chartography. Carbon still leads the family rows. Treat Gemini 4 Hydrogen as a cheaper worker beside Carbon or Argon, not a replacement. See our Carbon and Argon posts.
  • Yes for most agent and coding work. Gemini 4 Hydrogen more than triples Terminal-Bench 4.1 vs Gemini 3.8 Flash (57.0% vs 18.2%, self-reported by Google), uses a unified 1M window, and matches 3.8 Flash's own list price from January 1, 2027 ($1.50 / $7.50). Until then it is 2x today's 3.8 Flash promo. Default thinking moves from medium to high; remove explicit caching and sampling parameters; minimal is rejected.
  • Yes. The Gemini API free tier is available (rate-limited; content may be used to improve products). Paid access is also on Vertex AI, Google Antigravity for AI Pro and Ultra, the Gemini app (free users via automatic model selection), Google Search, and OpenRouter as google/gemini-4-hydrogen.
  • Gemini API in Google AI Studio and Android Studio; Gemini for business and Vertex AI; Google Antigravity for paid Google AI Pro and Ultra subscribers; the Gemini app model picker; Google Search for all users; Gemini in Google Sheets starting with AI Pro and Ultra; and OpenRouter as google/gemini-4-hydrogen. The Antigravity agent in Gemini Managed Agents now uses Gemini 4 Hydrogen by default.

Outlook

Gemini 4 Hydrogen is Google's everyday execution SKU: Argon-class agent work at $1.50 / $7.50, flat to 1M, generally available with a free tier. Watch three calendar items. The full Gemini 4 Hydrogen safety report is still coming soon. Carbon and Argon move to $4 / $20 on December 1, which widens the price gap inside the family. Gemini 3.8 Flash moves to $1.50 / $7.50 on January 1, 2027, which removes the per-token premium for Flash users and makes Gemini 4 Hydrogen the default price point in Google's lineup. Until then, budget for the high default thinking level, and keep Carbon or Argon on the hard planning turns.

Continue Reading