AI Model Timeline

Tracking the accelerating release frequency of frontier AI models.

Gemini 4 Hydrogen

Google·

LLMsmidcloud

Overview

Fast, lower-cost tier of the Gemini 4 family, adapted from Gemini 4 Carbon per Google's model card and aimed at everyday coding and agentic workflows. Generally available on 2026-11-12 (API id gemini-4-hydrogen); default model of the Antigravity agent and rolling out in the Gemini app and Google Search. Google's migration guide moves developers from gemini-3.8-flash to this id. Unified 1,048,576-token context for input plus output, thinking levels low to xhigh with high as default. $1.50 / $7.50 per MTok, flat across the context window, cached input $0.075. Google's model card carries none of the tracked academic benchmarks. It reports Terminal-Bench 4.1 at 57.0 from its own harness and cites HLE-Diamond 58.8 from Artificial Analysis, a different subset from the HLE tracked here. Artificial Analysis Intelligence Index v5.0.1: 51 at high, 54 at xhigh.

Capabilities and innovations

Text, image, video, audio and PDF input; text output1M-token unified context (input plus output)Thinking levels low, medium, high (default) and xhighFunction calling, code execution, computer use, Search and Maps groundingBatch, Flex and Priority tiersFirst fast tier of the Gemini 4 generationUnified 1M context in place of 1M input plus 64K outputParallel batching of read and search tool calls

Benchmarks

BenchmarkScoreSource
Terminal-Bench 4.054.04%independent
Terminal-Bench 4.0 54.04% (±3.31), measured by Vals AI at effort xHigh, rank 11/64 (page updated 2026-11-12). Vals AI is the reference harness for this field. 4.0 is its own generation and not comparable with the 2.x or 3.0 fields.
  • GPQA Diamond: not reported by the vendor — Not on Google's model card; no standalone GPQA Diamond value on Vals AI or Artificial Analysis (checked 2026-11-13).
  • Humanity’s Last Exam (no tools): not reported by the vendor — Google cites HLE-Diamond 58.8 from the Artificial Analysis leaderboard. HLE-Diamond is a refined 1,000-question subset, not the HLE tracked here, and is not written into this field.
  • SWE-bench Verified: not reported by the vendor
  • SWE-bench Pro: not reported by the vendor — Google reports its own SWE-sweep (13.5) instead; not a tracked field.
  • MMLU: not reported by the vendor
  • MMLU-Pro: not reported by the vendor
  • Terminal-Bench 2.x: not reported by the vendor — Only the 4.x generation is reported, recorded in terminal_bench_4.
  • Terminal-Bench 3.0: not reported by the vendor
  • DeepSWE v1.1: not reported by the vendor — No DeepSWE figure in any version on Google's model card or evaluation PDF (checked 2026-11-13).
  • MMMU-Pro: not reported by the vendor

API pricing

$1.5 per million input tokens, $7.5 per million output tokens (USD, provider list price, checked 2026-11-13)

Links

More from Google

Data curated by AI Model Timeline. See the methodology for admission criteria, benchmark eras and source priority.