Google releases Gemini 4 Hydrogen for coding agents

• Gemini 4 Hydrogen is live in Gemini, AI Studio and Antigravity, according to Google's announcement.
• Terminal-Bench 4.1 rises from Flash's 18.2% to Hydrogen's 57.0%, +38.8 points, in Google's reported results.
• Four thinking levels include a new xhigh option, with Google's lagoon demo showing the controls alongside /swarm.
• Carbon can coordinate Hydrogen workers, as demonstrated in Google's multi-model animation build.
• Quality and latency wins vary: simonw's comparison found weaker SVGs, while theo's code-review test finished slower than competing models.

Antigravity can switch between Carbon and Hydrogen mid-session. The Cloud developer guide puts input, thinking and output inside one token budget; Plotly's tests found Hydrogen slower than Argon despite its better score.

• Release: November 12, 2026, with the stable model slug gemini-4-hydrogen in the Cloud guide.
• Gemini app: Free access, confirmed in a free-access reply, comes through automatic routing; paid users get dropdown access under Google's rollout.
• Developer access: AI Studio and Gemini API across all tiers, plus Android Studio, per Google's announcement.
• Antigravity: Available to paid Google AI Pro and Ultra subscribers under the launch terms.
• Other Google surfaces: Search globally, Gemini for business, and Sheets for eligible subscriptions, per the rollout list.
• Thinking: Low, medium, high (default) and xhigh, detailed in the launch post.
• Standard API pricing: $1.50 input, $0.075 cached input and $7.50 output per million tokens, according to simonw's pricing comparison.
• December 1 pricing: Carbon and Argon move from $2/$10 to $4/$20 input/output, +100%; Hydrogen stays flat in the published schedule.

First-party

• Terminal-Bench 4.1: Flash 18.2% → Hydrogen 57.0%, +38.8 points, according to the reported benchmark results.
• SWE-sweep: Sonnet 5.5 3.1% → Hydrogen 13.5%, +10.4 points, per Google's table.
• Agents' Last Exam v1.1: Sonnet 5.5 60.4% → Hydrogen 71.6%, +11.2 points, in the comparison.
• MKA: Sonnet 5.5 52.0% → Hydrogen 53.4%, +1.4 points, per the benchmark table.
• Chartography: Sonnet 5.5 61.6% → Hydrogen 66.6%, +5.0 points, in Google's results.

Third-party evaluators

• Artificial Analysis Intelligence Index, high: Argon 50 → Hydrogen 51, +1 point, as reported in Google's announcement.
• Artificial Analysis Intelligence Index, Hydrogen high → xhigh: 51 → 54, +3 points, according to Google's account.
• AutomationBench-AA: Sonnet 5.5 71.8% → Hydrogen 76.1%, +4.3 points, in Google's table, sourced from Artificial Analysis under the evaluation notes.
• HLE-Diamond: Sonnet 5.5 53.3% → Hydrogen 58.8%, +5.5 points, in the comparison, also sourced from Artificial Analysis.
• Vending-Bench 2 balance: Sonnet 5.5 $11,165 → Hydrogen $14,549, +$3,384, in Google's table, using Andon Labs leaderboard results under the methodology.
• Plotly analytics exam: Carbon 42/43 → Hydrogen 43/43, +1 correct answer, in Plotly's independently run test.

Customer-reported

• Partner token-use index, Flash = 100: 100 → 77–88, −12 to −23 points, calculated from Google's summary of unnamed partners' early tests.

Hydrogen trails Sonnet 5.5 on three evaluations in Google's comparison:

• Terminal-Bench 4.1: 63.9% → 57.0%, −6.9 points, per the benchmark table.
• OSWorld-2.1 offline subset: 83.9% → 77.6%, −6.3 points, in Google's results.
• Terminal-Bench-Science 0.1: 59.9% → 51.4%, −8.5 points, per the comparison.

Google's 2.5× faster task-completion claim compares Hydrogen at default high thinking against Gemini 3.8 Flash. The task mix behind that headline is unspecified in the launch post.

For Terminal-Bench 4.1, the evaluation notes specify a mini-swe-agent vision harness at high effort, averaging five attempts per task rather than selecting the best attempt. OSWorld uses an 82-task offline subset and the maximum across three runs.

Hydrogen was “very cheap and very fast” in imjustnewatai's prelaunch description. In a four-round code-review test, theo measured Hydrogen taking 2.07× as long as GPT-6.1 Sol, with Sol and Opus both running on regular tiers.

Plotly found another cross-model split in its analytics exam:

• Median question time: Argon 23 seconds → Hydrogen 32 seconds, +9 seconds.
• Cost at the same perfect score: GPT-6.1 Sol $2.61 → Hydrogen $6.36, +$3.75.

Plotly ran each model once in Plotly Studio Embedded against a private synthetic wind-energy dataset. Repeat runs of other models have varied by up to three correct answers, making its one-answer Carbon gap a narrow result.

SVG fidelity also fell across the three tested effort levels in simonw's pelican comparison.

Shared context and long decodes

Google advertises “native 1 million-token input” and “up to 1 million output tokens” in its announcement. The Cloud developer guide documents the operational limits:

• Shared budget: 1,048,576 tokens across input and output. Output includes thinking, function calls and the final response.
• Foreground segment: Up to 65,536 output tokens, whether streaming or non-streaming.
• Long decodes: Continuations extend generation within the remaining shared budget through the Interactions API or GenerateContent.
• API version: Cloud long decoding requires v1beta1; v1 does not support it.
• SDK continuation: Automatic Interactions continuation starts at Python google-genai 2.34.0 or JavaScript/TypeScript @google/genai 2.32.0.

Migration and caching

The Cloud migration guide and Hydrogen guide expose several compatibility details:

• Numeric thinking_budget returns HTTP 400 on Gemini 4; configuration uses thinking_level.
• MINIMAL thinking is unsupported.
• Implicit caching begins at a shared prefix of 4,096 tokens, with durable caching enabled by default. Explicit context caching is unsupported.
• Function responses must preserve the call's ID and name, with exactly one response per call. The migration guide says mismatches commonly produce empty responses with finish_reason: STOP.

Orchestrator and /swarm

Google describes two distinct multi-model mechanisms in the launch post:

1. Orchestrator: Switches models within a session. On Terminal-Bench 4.0, Google reports near-Carbon success with 24% less quota and 19% faster completion.
2. Parallel agents: Carbon coordinates Hydrogen workers and verifiers that build and test simultaneously in isolated workspaces.

• kimmonismus: An early-access Hydrogen High run produced a flyable Three.js city after one follow-up, with procedural towers, rain and flying traffic in kimmonismus's thread.
• ai_for_success: Combined Hydrogen with /swarm for an interactive island build, according to the creator's post; Google separately highlighted its use of all four thinking levels.
• Bhavani_00007: Reused the same animated 3D desert-ranch prompt across Flash and Hydrogen in Google's side-by-side demonstration.

Outside Twitter, oblec reported a multi-file Python refactor that retained type hints in a NeoTiler discussion. In the same discussion, kernel_panic_dan reported a first-pass Rust async lifetime fix at xhigh, alongside a longer wait before streaming and greater reasoning-token use.

Lingo City demonstrates a different workflow: generating custom language flashcards on demand, then moving into navigable 3D streets for conversations with shopkeepers.

• OpenClaw: Added Hydrogen support on launch day, according to its announcement.
• LLM CLI: The llm-gemini 0.37 release adds access to the Gemini 4 family.
• LiteLLM: Day-zero gateway support uses gemini/gemini-4-hydrogen, covering Google AI Studio and Gemini Enterprise Agent Platform.
• Search: Hydrogen powers an expanded query fan-out process intended to retrieve a broader range of links, according to Google.

On Google Cloud, Hydrogen is classified as a short-term GA release with a minimum 45-day retirement-notice policy in the developer guide.