Google Retires Pro/Flash, Ships Gemini 4 Hydrogen: 2.5x Faster, Four Thinking Gears, and a Hidden API Pitfall
Developers who run cheap, fast agent loops can cut latency and token costs by switching from Flash-era models to Hydrogen, but must account for the API ignoring sampling parameters and the shared 1M context window that eats thinking tokens.
Google DeepMind has retired its Pro/Flash naming scheme and shipped the second model in its chemical-element series: Gemini 4 Hydrogen. The model takes over the Flash role as the fast, lightweight worker for high-frequency coding, quick tool calls, and multi-model routing, while Gemini 4 Carbon remains the heavy flagship for complex architecture and long-horizon agent tasks.
Hands-on testing shows Hydrogen completes tasks about 2.5x faster than Gemini 3.8 Flash, scoring 51 on the default high thinking level and 54 on xhigh in Artificial Analysis Intelligence Index v5.0, with 57.0% on Terminal-Bench 4.1. The model natively supports four thinking levels—low, medium, high, and xhigh—but ignores user-supplied temperature, top_p, and top_k; all sampling is governed by thinking_level, and thinking tokens count toward the shared 1M context window. Pricing drops to $1.50 per million input tokens and $7.50 per million output tokens with no doubling beyond 200k, and a Free Tier remains available in AI Studio.
The author also shares a 15-line Python snippet using the google-genai SDK to invoke Hydrogen with thinking_level and to log thoughts_token_count, and recommends a three-step orchestration pattern: Carbon defines contracts and splits tasks, Hydrogen handles concurrent edits and tests, and escalation to xhigh or Carbon covers dead ends.
Chemical-element naming makes the split between heavy reasoning and fast iteration clearer than Pro/Flash ever did, so developers can pick a model based on its role in a pipeline rather than its prestige.
The API quietly dropping temperature and top_k signals a shift away from user-controlled sampling toward fixed thinking-level presets; teams that rely on those knobs need to audit their prompts.
Counting thinking tokens inside the shared 1M context changes cost math: verbose reasoning can consume budget without producing visible output, so logging thoughts_token_count becomes a necessary cost-control practice.
The 2.5x speedup with comparable intelligence scores suggests lightweight models have reached production usability for terminal-heavy agent work, which lowers the barrier to fully autonomous coding loops.
The hydrogen model's real selling point isn't a single benchmark score, but cost-effectiveness at the execution layer. Its terminal benchmark of 57.0% far exceeds the Flash model's 18.2%, but when the thinking level is set to ultra-high, token consumption and latency also rise. At the $1.50 input / $7.50 output pricing, the cost advantage shrinks considerably for output-heavy tasks. When routing in tools like Cursor, use the low tier for simple tasks and switch to ultra-high only for complex ones.