Google put Gemini 4 Hydrogen into Search, the API and AI Studio on November 12, claiming it runs 2.5 times faster than Gemini 3.8 Flash. A cheaper sibling model may matter just as much to builders.
Koda editorial team · 13 November 2026 · 7 min read
Listen · 19 min
Today: Test the cheap judgment layer before you scale it. The steps
Google launched Gemini 4 Hydrogen on November 12 across the Gemini app, API, AI Studio and Search, priced at $1.50 per million input tokens and $7.50 per million output tokens1. Google says it runs 2.5 times faster than Gemini 3.8 Flash and scores approximately 57% on Terminal-Bench 4.1 against 18.2%; Anthropic shipped Claude Fable 5.2 three days earlier.
Why it matters. Teams running coding agents on Gemini 3.8 Flash have a clear reason to move to Gemini 4 Hydrogen, which roughly triples its Terminal-Bench 4.1 score while charging $7.50 per million output tokens1. (Token Tracker lists Gemini 3.8 Flash at $0.75 input and $3.75 output per million tokens, scraped 7 November 2026.)
Go further: today's paper, Double Blind Evals: Resolving the Dual Confidentiality Dilemma in AI Safety Auditing (arXiv, 9 Nov).
Nonprofits sued OpenAI, Anthropic, SpaceX's AI division, Meta and Google in San Francisco on November 12 to bar frontier model work.
Why it matters. The plaintiffs ask the court to bar development itself, so even a temporary order would freeze model roadmaps that enterprise buyers have built multi-year contracts around.
SoftBank Group swung from profit to a $7 billion2 net loss in its fiscal second quarter, Reuters reported November 10, as financing costs more than doubled.
Why it matters. Startups counting on SoftBank Group checks should expect tougher terms, since financing costs that more than doubled leave it less cash for new AI bets.
Google's Gemini Decision 1 answers up to 1,000 yes-or-no3, multiple-choice or rubric questions per request without writing prose.
Why it matters. Eval and moderation teams can pack up to 1,000 judgments into one Gemini Decision 1 call and skip parsing prose answers3, which cuts the cost and glue code of model-as-judge pipelines.
Blue, a Y Combinator-backed startup from a former Google PM, has unveiled an AI phone built as a 'chief of staff' that manages your apps. It handles trivial tasks and surfaces only what matters most. Its beta lets users trade in an iPhone for a free Blue phone with 3 months of unlimited use.
a16z argues AI agents are becoming the next major distribution channel after the web and mobile. When a16z asked ChatGPT for a video ad, it pulled in Krea, Descript and ElevenLabs on its own, then persuaded the team to buy a paid Descript plan. In that test, the agent chose the vendors and made the sales pitch to the buyer.
Jensen Huang was Nvidia's CEO; he has now resigned, and on Thursday the Nasdaq fell 0.39%2 to 25,366.09 while the S&P 500 lost 0.35% to 7,592.4 as the AI trade wobbled.
Why it matters. Nvidia anchors the AI trade, so doubt over who runs it raises the risk premium on chip-linked stocks and on the capex plans that assume its roadmap holds.
S&P 500 -0.35%2, the Nasdaq -0.39% and WTI crude +1.79% on the 12 Nov session. Bitcoin -4.44% and Ether -2.51% on the 13 Nov session. The crypto Fear & Greed index reads 20, Extreme Fear.
S&P 500
7,592.40
-0.35%
Nasdaq
25,366.09
-0.39%
Bitcoin
$70,679.93
-4.44%
Ethereum
$2,004.22
-2.51%
Crude Oil (WTI)
$88.08
+1.79%
Crypto Fear & Greed
20 / 100
Extreme Fear
Levels from Yahoo Finance4, 12 Nov and 13 Nov; alternative.me5, 13 Nov.
Design · Koda score6 5.3 / 10
Builder turns a prompt into a clickable prototype you can hand back as HTML, and Adobe unveiled it at Adobe MAX 2026 pointed squarely at Figma's territory. The case for it rests on Adobe's own stage showing, so it is a bet worth placing only for Creative Cloud shops whose brand assets already sit in Adobe's ecosystem. Get on the waitlist and keep your current prototyping tool in production.
Cheap judgments may matter as much as faster prose.
Gemini Decision 1 returns probabilities for $0.10 per million input tokens, so ticket triage costs 45x less than running Hydrogen. With Hydrogen now live in Search and the API, teams can split work between the two models. Google's calibration claims remain untested, so score those probabilities against your own labelled tickets first. Then send only the tickets that need written answers to Hydrogen.
Below, by source, are the 18 figures we checked: 2 verified, 13 reported and 3 estimated.
Verified confirmed by an independent searchReported as published by the named sourceEstimated approximate, in the source or in our check
Signal 231 · 13 November 2026 · 13 sources · How every number is checked · Corrections
230 editions so far, one a day. Unsubscribe anytime.
Forward this to one operator you work with. Your referral link is in every email; milestones at koda.community/refer.