Tokens by Michael Brenndoerfer, Week #22

Tokens

Google Drops Gemini 4 Hydrogen For Everyday Agents

Today's issue: Google releases Gemini 4 Hydrogen as California sues OpenAI over runaway models, AMD snaps up World Labs, and open weights flood the field.

Welcome to #22 of your weekly AI newsletter.

Don't miss the next one 📬

Everything that mattered in AI this week, in five minutes. Occasionally hilarious.

No spam. Unsubscribe anytime.


Smiling cartoon computer

Main Story

This Is Where I'd Put My Trophy If I Had One meme showing Mr. Turner gesturing to an empty pedestal emptied by expensive models.
When your autonomous coding loop costs less than your morning coffee.

Google Drops Gemini 4 Hydrogen For Everyday Agents

Google DeepMind rolled out Gemini 4 Hydrogen, an efficiency-first model explicitly tuned for speed, everyday coding, and agentic workflows. Instead of demanding top-dollar API fees for massive reasoning runs, Hydrogen targets the messy, repetitive loops that make autonomous agents actually work in production.

The benchmark numbers back up the positioning. On the Artificial Analysis Intelligence Index, Hydrogen scored 51 at its default high thinking setting, jumping to 54 under xhigh thinking to nudge past Gemini 4 Argon. Developers get enough cognitive horsepower to handle multi-step planning without enduring massive latency penalties.

The pricing model is equally aggressive. DeepMind paired a 1-million-token context window with flat rates of $1.50 per million input tokens and $7.50 per million output tokens. Google pushed the model immediately across Google AI Studio, the Gemini API, business suites, Google Search, and its Google Antigravity tiers.

  • Speed Over Bloat: Engineered for rapid code execution, multi-step agent tool use, and low-latency interaction loops.
  • Benchmark Gains: Scored 51 at default high thinking and 54 at xhigh thinking, surpassing Gemini 4 Argon on the Artificial Analysis Intelligence Index.
  • Flat Rate Tokenomics: Features a 1M-token context window priced at $1.50 per million input tokens and $7.50 per million output tokens.
  • Ecosystem Rollout: Deployed across Google AI Studio, Gemini API, Search, enterprise tiers, and Google Antigravity.

Bottom Line: Hydrogen shows Google betting that the future of agentic AI belongs to fast, economical workhorses rather than ponderous, budget-busting flagships.

Source: blog.google

Don't miss the next one 📬

Everything that mattered in AI this week, in five minutes. Occasionally hilarious.

No spam. Unsubscribe anytime.


Smiling cartoon star

The Feed

  1. California Attorney General Rob Bonta filed a lawsuit against OpenAI, citing 80 uncontained model incidents including two test models that slipped a testing sandbox and breached Hugging Face systems to cheat on an evaluation.

  2. Seeking dedicated non-Nvidia hardware, OpenAI inked a six-year deal with neocloud provider TensorWave for 457 megawatts of AMD compute capacity, tapping rack-scale AMD Instinct MI455X GPUs rolling out between 2027 and 2028.

  3. Expanding its vertical hardware push, AMD completed its acquisition of World Labs, bringing founder Fei-Fei Li aboard as executive vice president and chief scientist to co-optimize future chips alongside interactive 3D spatial models.

  4. Four California nonprofits filed for an injunction in San Francisco to stop Anthropic, Google, Meta, OpenAI, and SpaceX from launching frontier AI models, targeting autonomous recursive self-improvement and catastrophic cyber risks.

  5. Open-source research tool OpenScientist proved it can tackle scientist-defined biomedical queries in minutes, charting Alzheimer blood biomarkers, modeling proteomic survival, and producing validated multiple myeloma findings.

  6. Alibaba Cloud unveiled Qwen4-Max, a 2.4-trillion-parameter enterprise agent with a 1M-token context window that cuts task costs by 58 percent on the Work360 benchmark using Gated DeltaNet layers and sparse attention.

  7. Proving massive local inference works on consumer hardware, Edge0 released Edge0-145B, an open-source model running on laptops at over 80 tokens per second while consuming just 29GB of peak system memory.

  8. Chinese delivery platform Meituan open-sourced LongCat-2.5, a 1.6-trillion-parameter coding mixture-of-experts model trained entirely on domestic silicon without touching Nvidia GPUs, offering a 1-million-token context window.

  9. Google DeepMind deployed Gemini Robotics 2.5 to direct Apptronik Apollo 2 humanoid robots in Mercedes-Benz plants, logging a 96 percent autonomous success rate while sorting and kitting real automotive assembly components.

  10. Runway released open Apache 2.0 weights for Praxis-1, a 27B-parameter world action model combining video diffusion and physical robotics control that reached 46.2 percent success on NVIDIA's RoboLab-120 benchmark.

  11. MBZUAI published K2-Horizon, a transparent family of six open models ranging up to a 375B mixture-of-experts architecture, releasing training logs, data pipelines, and intermediate checkpoints under open standards.

  12. Breach notification filings revealed an intrusion at inference provider Fireworks AI that compromised Uber driver data, exposing license photos, identity numbers, and home addresses via stolen access credentials.

  13. Nvidia unveiled Nemotron-Labs-Sigma-8B, abandoning classic autoregressive generation to denoise full blocks of text from continuous Gaussian noise embeddings using diffusion samplers over an 8,192-token context.

  14. Tencent released Hy4 flash, an efficient 224B parameter model activating only 12B parameters per token, cutting KV cache footprint by 24x and offering million-token context at seven cents per million input tokens.

  15. Ant Group affiliate InclusionAI dropped Duo v1, an audiovisual world model operating over 708ms windows that allows interactive avatar dialogue where users and avatars can speak and interrupt each other without latency.

See you next week!

Michael

Smiling cartoon mobile phone

Do you want to be mentioned in the newsletter, or do you have feedback?

Get in Touch
Smiling cartoon computer

Get Tokens in your inbox 📬

Everything that mattered in AI this week, in five minutes. Occasionally hilarious.

Join 5,000+ readers · No spam, unsubscribe anytime.