# Google Says Gemini 4 Hydrogen Runs Tasks 2.5× Faster

Canonical URL: https://superpowerdaily.com/videos/a4359a9a-eaeb-4128-b423-4dea0913917f
Published: 2026-11-13T10:00:26.000Z
Episode date: 2026-11-13
Language: English
Duration: 8:41
Publisher: Superpower Daily
YouTube: https://www.youtube.com/watch?v=nOaVlTv-sKs

Google says Gemini 4 Hydrogen completes tasks faster in company-published tests, while Anthropic says Mythos ported one model’s prompt-processing workload to AMD GPUs in three days. OpenAI reports stronger resistance to evaluated jailbreaks after adversarial training, and WIRED says some Nvidia employees are urging Jensen Huang’s return—though an internal comment count is not a companywide vote.

## Key takeaways

- Gemini Hydrogen targets fast agentic workflows and competit­
- Anthropic reports a rapid, narrowly scoped AMD port; broader
- OpenAI reports stronger jailbreak resistance, but evaluated
- WIRED reports internal pressure over Huang’s exit; comments

## Chapters

- 0:23 Gemini gets faster: https://superpowerdaily.com/videos/a4359a9a-eaeb-4128-b423-4dea0913917f?t=23
- 2:35 AI ports chip software: https://superpowerdaily.com/videos/a4359a9a-eaeb-4128-b423-4dea0913917f?t=155
- 4:26 Training stronger defenses: https://superpowerdaily.com/videos/a4359a9a-eaeb-4128-b423-4dea0913917f?t=266
- 5:44 Nvidia leadership pressure: https://superpowerdaily.com/videos/a4359a9a-eaeb-4128-b423-4dea0913917f?t=344

## Stories covered

- Google releases Gemini 4 Hydrogen for coding, Search and the Gemini app: https://superpowerdaily.com/posts/google-releases-gemini-4-hydrogen-for-coding-search-and-the-gemini-app
  API rates start at $1.50 per million input tokens. Google’s performance pitch combines adjustable reasoning with automatic model switching in Antigravity.
- Anthropic says AI cut an AMD chip port from months to three days: https://superpowerdaily.com/posts/anthropic-says-ai-cut-an-amd-chip-port-from-months-to-three-days
  The investor disclosure points to lower software barriers between chip suppliers. Its headline result covers prompt processing for one model—not every part of running Claude.
- OpenAI finds training on AI-generated attacks improves jailbreak resistance: https://superpowerdaily.com/posts/openai-finds-training-on-ai-generated-attacks-improves-jailbreak-resistance
  Controlled experiments showed gains against both fixed prompts and attackers that adapt mid-conversation. The results measure the model layer, without production safeguards.
- Nvidia employees call for Jensen Huang’s return after his resignation, WIRED reports: https://superpowerdaily.com/posts/nvidia-employees-call-for-jensen-huang-s-return-after-his-resignation-wired-reports
  An internal campaign is challenging the leadership transition, with employees seeking transparency about Huang’s departure and a voice in the board’s decision.

## Transcript

Automatically transcribed from the narration.

[0:00] Welcome to your Superpower Daily AI Briefing I'm your host and today we are unpacking four massive market stories We've got a lot to cover so let's get right into it Today we're looking at Google's super fast new Gemini 4 Hydrogen model anthropic porting chips in absolute record time OpenAI's latest jailbreak resistance data and finally a major internal push from NVIDIA employees who want Jensen Wang back All right kicking things off with chapter one Google Gemini 4

[0:27] Hydrogen bringing high speed and adjustable reasoning to the table So Google is officially dividing the labor within the Gemini 4 family which rolled out globally for coding search and the Gemini app on November 12th 2026 On one end you've got carbon and argon These are your heavy lifters built for deep extended reasoning But hydrogen Hydrogen is your speed demon It's specifically designed for rapid execution and what we call agentic workflows That basically means it can act

[0:56] autonomously to write code manage tasks or debug right alongside those heavier models without slowing you down at all It's like having a lightning fast assistant that actually knocks out the busy work for you And here's the really cool part These models aren't just stuck operating in silos Google introduced something called the anti gravity orchestrator This system dynamically switches between the models mid task so carbon handles the complex coordination and then instantly hands it off to hydrogen

[1:24] for quick execution The result of this tag team effort A massive 24% drop in quota usage and it finishes 19% faster all without sacrificing accuracy So what does that actually mean for us It means getting more done in way less time for less money Speaking of money the crucial point for developers here is definitely the cost Hydrogen API rates start at just $1.50 per million input tokens and $7.50 per million output tokens And that flat rate

[1:54] applies across its entire 1 million token context window That makes it incredibly cheap if you're processing massive documents giant spreadsheets or even feeding audio and video into structured reports Now the performance jump here is literally wild According to Google's own benchmarks hydrogen hit a 57% score on Terminal Bench 4.1 When you compare that to the 18.2% we saw from the older Gemini 3.8 Flash you can see just what a massive leap in capability this is You

[2:23] can even dial your reasoning settings up or down based on your need for speed But as always remember these are company reported figures The real test is going to be seeing how these gains hold up in everyday developer environments Speaking of massive leaps in efficiency let's move to our second story and look at Anthropic's AMD chip port where they are actively cutting the software bottleneck For years Nvidia has maintained this massive advantage in the hardware market

[2:49] And it wasn't just because of the physical chips right It was their software ecosystem CUDA Switching to a different chip provider usually means moving absolute mountains of custom code Traditionally adapting to a new hardware platform takes dozens of engineers working for six to 12 months But Anthropic claims their Frontier AI model Mythos took that exact same workload and got it running on AMD GPUs in about three days Three days To pull off a turnaround like that

[3:16] with almost no human intervention the AI actually had to write or adapt nearly all the kernels In plain English kernels are the fundamental highly specialized mathematical instructions that tell a specific microchip exactly how to process data Because this code is so painstakingly specific to the hardware it's the exact reason Nvidia's software ecosystem has been such a powerful mode By having an AI write these kernels Anthropic is proving that this massive software barrier to switching chip suppliers

[3:45] might finally be crumbling But we've got to keep a key limitation in mind here Disclosures from their investor materials show this specific three day port was highly targeted It only covered prompt processing for one single Anthropic model running on AMD MI355X GPUs So no it wasn't some magical overnight migration of every single clod workload Anthropic's ultimate goal is to treat all their chips whether from Amazon Google Nvidia or AMD as one big flexible interchangeable pool But

[4:19] making distinct platforms genuinely interchangeable across the board that remains a much larger engineering challenge All right from overcoming hardware bottlenecks we are turning to overcoming security vulnerabilities In chapter three OpenAI boosts jailbreak resistance fighting AI generated attacks OpenAI recently released a fascinating disclosure showing how they are fighting back against adversarial attacks by essentially turning AI against itself They're using reinforcement learning to pit models against each other By generating AI attackers to try and jailbreak or trick

[4:53] the system the defending model learns to resist those exact tactics Think of it like a digital immune system constantly stress testing itself And the cybersecurity improvements are stark GPT 6 Astra successfully answered safely on 91.5% of fixed jailbreaks That absolutely crushes the older GPT 5.6 sole models 59% This automated red teaming having AI play the role of the bad guy is proving to be incredibly effective The most important limitation to note however is that these results

[5:25] only measured the base model layer's performance against specific evaluated prompts So this data doesn't account for the full suite of safety systems filters and guardrails that you actually experience when using the product in the real world Still strengthening the base model's inherent resistance right out of the gate is a massive step forward for AI safety Finally chapter four We're shifting gears to a major corporate story unfolding at NVIDIA Employees want Huang back as an internal push

[5:51] challenges the transition Following Jensen Huang's sudden resignation amid unresolved allegations and a board led transition the internal debate is definitely far from over Wired reports that an internal campaign is actively challenging the board's timeline and decision See prior to Huang stepping down an internal suggestion box post had asked the board to share investigation findings and just gauge employee sentiment But after he resigned it turned into a full blown campaign for his return Wired noted anonymous comments

[6:21] like this one If you believe in the mission bring back the boss They're directly urging directors to give Huang a second chance arguing that his past contributions and the company's overall mission completely justify it To put a number on it more than 1,600 comments landed on this internal suggestion box post Now we must carefully note that this counts individual comments not 1,600 unique employees voting in some official company wide poll Furthermore while Wired noted about two

[6:50] dozen employees publicly backed Huang on LinkedIn shortly after the resignation the actual scale of this sentiment across NVIDIA's massive global workforce remains a bit unclear But what is clear The workforce is demanding transparency They want to know exactly why Huang stepped down right as the board's investigation was just getting started So while some visible support for Huang is obvious the current reality for the company is pretty severe The board appointed finance chief Colette Kress as interim

[7:18] CEO and former Synopsys CEO Harvey Jones as board chair They are now navigating a company that has lost a staggering $1 trillion in market value since the allegations serviced all while dealing with ongoing inquiries from US lawmakers and the FBI The big unresolved question now is whether the board will formally explain its transition timeline to the staff or if they'll just hold firm on their decision amid all this internal pressure To wrap up here are your

[7:44] four concise takeaways for today's briefing One Google's Gemini 4 hydrogen is prioritizing high speed agentic workflows with highly competitive API pricing at $1.50 per million input tokens Two Anthropic successfully used its Methos AI to compress an AMD chip port workload from months down to just three days lowering major software barriers Three OpenAI's training against AI generated attacks pushed GPT 6 Astra's safety resistance past 91% And four NVIDIA staff are pressuring the board over Jensen Huang's exit

[8:18] amidst a $1 trillion market loss I will leave you with this final provocative thought to ponder As frontier models learn to write their own highly specialized kernels will AI assisted chip porting finally break the software mode of legacy hardware ecosystems Read more at superpowerdaily com Stay curious

## Editorial standards

https://superpowerdaily.com/editorial-standards

This episode reports information available at publication time. Consult the linked stories and their reporting sources for context and subsequent updates.