Friday 13 November 2026
Gemini 4 Hydrogen triples 3.8 Flash's Terminal-Bench score
Five bullets
The one thing
Google released Gemini 4 Hydrogen, the lowest-cost model in the Gemini 4 family. It replaces Gemini 3.8 Flash. Google reports 57% on Terminal-Bench 4.1, up from 18.2% for 3.8 Flash. It costs $1.50 per million input tokens and $7.50 per million output, flat across a 1 million-token context.
Google says Hydrogen batches file reads and searches in parallel instead of opening files one by one. Partners saw it run 2x to 2.8x faster with 12% to 23% fewer tokens than 3.8 Flash. It offers four thinking levels: low, medium, high by default, and a new xhigh.
The pricing needs a careful read. Simon Willison notes Hydrogen is not far below its bigger siblings, Carbon and Argon. Those two cost $2.00 input and $10.00 output through November 30. Both double on December 1.
Google pitches Hydrogen as the worker that writes code and runs tests while a bigger model plans. Its new Orchestrator in Antigravity switches between Carbon and Hydrogen within one session. On Terminal-Bench 4.0 it came near Carbon's success rate on 24% less quota. The skill that grows is matching each step to the cheapest model that still passes your checks.
Test Hydrogen on tasks your agent already finished before you move any real work to it.
Try this · 15 minutes
See whether Gemini 4 Hydrogen handles your routine agent tasks
- Pick a small task your current agent finished well this week. Save the prompt and the diff.
- Run the same prompt on Gemini 4 Hydrogen through Google AI Studio or the Gemini API.
- Run it again at the low thinking level. Compare speed and output against the default high.
- Check both results against your tests. Note any files Hydrogen edited that the task never needed.
Stories
- TRYClaude Code tests a mode where Claude scripts its own tool callsClaude Code shipped Code Mode as an experimental opt-in in v2.1.323. In Code Mode, Claude writes JavaScript that calls its tools, rather than making one tool call at a time. The release notes link to docs for turning it on, or off for an organization. Daily Claude Code users can try it on one real task.
- TRYopencode and Codex make parallel agent sessions easier to manageopencode v2.3 runs subagents in the background by default . Subagents return at once and report back when idle, so you keep working in the parent session. Separately, Codex 0.169.0 adds side-chat tabs to its opt-in terminal workspace, plus search over older tasks. Try it if you already split work across subagents.
- WATCHLLM judges missed forbidden writes that a simple rule caughtA tau2-bench study scored 2,160 agent runs four ways . The LLM judge passed 10 to 11 of the 11 runs where the agent made a forbidden write. A rule-based check of expected versus actual writes caught all 11 with no model call. Anyone grading agent output with another model should read it.
- TRYGitHub explains how to write agent skills that work across hostsA skill is a versioned directory built around SKILL.md. One packaging script generates Claude Code, Cursor and Copilot files from the same source.
- TRYMartin Fowler suggests generating a glossary of your domain's terms for agentsAn LLM drafts terms from internal docs, then domain experts correct it. Keep it in version control, with one glossary per bounded context.
- WATCHAmp replaces its effort Dial with a per-thread model selectorYou can now switch models mid-thread. The first message after a switch costs more, because prompt caches belong to one model.
Safe to ignore
57 sources read. 129 items. 12 shortlisted. 5 bullets.