Gemini 4 Hydrogen Is Live: Faster, Cheaper Agent Execution

Key Takeaways

• Available now: Google released Gemini 4 Hydrogen on 12 November and APIYI connected on 14 November. Model name gemini-4-hydrogen, available in the default and svip groups on both Gemini-native and OpenAI-compatible endpoints
• Priced at $1.50 / $7.50 per 1M tokens, with cached input at $0.075. The price is flat across the full 1M window — a standard rate, not a limited-time promotion, with no expiry date
• Google’s “daily execution” tier: adapted from Carbon and positioned as the upgrade path from 3.8 Flash. At the default high level it finishes tasks 2.5x faster, scores 57% on Terminal-Bench 4.1 versus 18.2%, and cuts fabricated claims by more than 5x
• One 1M window, four thinking levels: input, thinking and output share 1,048,576 tokens; a synchronous request returns up to 65,536 tokens, with long decode for more; thinkingLevel supports low, medium, high and xhigh, defaulting to high
• Tested before launch: both endpoints, streaming, function calling with thoughtSignature round-trips, JSON Schema, image, PDF, audio, video, a 326K-token needle test and six-way concurrency all passed

Background

The Gemini 4 family shipped three models in three weeks: Argon for long-horizon work (opened 13 October), Carbon for the hardest problems (opened 10 November), and Hydrogen for everyday execution (12 November). It drops the Flash branding, but Google’s upgrade advice for 3.8 Flash users is to move to Hydrogen. Hydrogen is adapted from Carbon, and the trade-off is deliberate: not the deepest single-shot reasoning, but speed, reliability and cost per step. Google pitches it at coding, tool calls, test runs and document reading — the high-frequency agent loop rather than multi-day architecture work. It also introduced Orchestrator: Carbon explores and plans, Hydrogen executes, and Google says the pair reaches near-Carbon success rates with 24% less quota and 19% faster completion on average. APIYI connected two days after launch, using the same access pattern as the rest of the Gemini 4 family: either the native endpoint or the OpenAI-compatible one.

Detailed Analysis

Official benchmarks

All figures below are published by Google and were not independently reproduced by APIYI: On safety, Google reports up to 81% less overreach and 79% less under-service in agentic tasks, 12x fewer misaligned actions across tool-use and coding evaluations, and more than 2.3x resilience to indirect prompt injection. These are vendor claims too; production deployments still need their own guardrails.

Four thinking levels

thinkingLevel is the main cost and latency dial:

• low: fewest tokens and lowest latency, for high-throughput extraction, classification and short Q&A
• medium: interactive coding, live UI prototyping and latency-sensitive tool calls
• high (default): complex coding, agent execution, tool calling and multimodal document work
• xhigh: new with Hydrogen, for multi-step problems where high falls short; Google warns it can overthink simple prompts and burn tokens for no better answer

Migration notes: 3.8 Flash defaulted to medium while Hydrogen defaults to high; minimal is no longer accepted (the native API returns an error), and thinkingBudget is deprecated — sending it together with the level returns 400. Thinking tokens are billed as output, so higher levels cost more.

The 1M window and output limits

Input, thinking and output share 1,048,576 tokens at a single flat rate, with no long-context surcharge. A synchronous request returns at most 65,536 tokens; Google says long decode (background or continuation) can push toward 1M. We ran a needle test on APIYI with a 520K-character document: promptTokenCount came back as 326,238 and the model found the buried ID in 29.7 seconds. A full 1M window was not tested; validate your own document shape at 200K–300K tokens before scaling up.

API changes

Four things change when moving from 3.8 Flash:

• Sampling parameters are ignored: temperature, top_p and top_k are accepted but have no effect. Pipelines that relied on low temperature need prompt-level constraints or structured output instead.
• No explicit caching: cachedContents is unavailable; only implicit prefix caching for prompts of 4,096+ tokens applies, and hits are not guaranteed. Model costs as if there were no cache.
• thoughtSignature must be returned: multi-turn and tool calls need the signature from the previous response passed back verbatim, or the request may fail with 400. This is a Gemini 4 requirement, not Hydrogen-specific.
• Knowledge cutoff is March 2026, so time-sensitive work still needs search or your own data.

Pre-launch testing

We ran 30 scenarios on APIYI at low concurrency, covering both endpoints, streaming, tools, structured output, multimodal input and long context. What passed: non-streaming and streaming on both endpoints; function calling produced get_weather({"city":"Beijing"}) and continued correctly after the tool result; JSON Schema returned valid JSON; receipt OCR, chart reading, PDF extraction, audio ID recall and video frame reading were all correct; the 326K-token needle test passed; six concurrent requests all succeeded. Latency: 10 interleaved short requests gave a P50 of 3.5 seconds. Streaming first token arrived in 7–10 seconds because thinking emits nothing until it finishes, then content arrives in chunks — frontends should show a “thinking” state rather than expect instant first-token output. One limit to avoid: thinking tokens count against max_tokens. With a small budget such as 300, thinking-heavy requests returned empty content with finish_reason=length; 600 and above produced text reliably. Keep max_tokens at 1024 or higher by default, and 2048+ for harder tasks.

Pricing and availability

• Model name gemini-4-hydrogen, available in the default and svip groups.
• Native endpoint: POST /v1beta/models/gemini-4-hydrogen:generateContent; OpenAI-compatible endpoint: POST /v1/chat/completions.
• Unlike Carbon and Argon’s limited-time rates, Hydrogen’s $1.50 / $7.50 is the standard price with no expiry.
• Recharge promotions can be stacked; see Recharge Promotions.

Summary and recommendations

Hydrogen is not a cheaper Carbon; it is the default execution tier. Reach for it for everyday coding, tool loops, document processing and multimodal extraction, and keep Carbon or Argon for planning, architecture and root-cause work. Suggested starting points:

• Start at the default level: run high, then evaluate whether medium or low meets your quality bar, or xhigh is needed.
• Give max_tokens room: thinking is part of the budget, and a small cap yields an empty response.
• Return signatures: pass thoughtSignature back verbatim in multi-turn and tool flows.
• Test long documents in stages: 326K tokens is verified; stress-test larger inputs first.
• Model cost without cache: implicit caching helps when it hits, but it is not guaranteed.

The model is ready to call on APIYI with the same access pattern as the rest of the Gemini family.