Tech • AI • Robotics • Game
Anthropic Slowing Down!?, Gemini 4 Hydrogen Is Insane, OpenAI's New Model, & NVIDIA Fires Jensen!
9.4/10TL;DR
Google launched Gemini 4 Hydrogen, a fast low-cost model that performs unusually well on coding and agentic tasks, while Anthropic, Nvidia, OpenAI, ByteDance, and Microsoft each introduced developments that highlight intensifying competition across AI models, safety, chips, and productivity tools.
KEY POINTS
Google unveils Gemini 4 Hydrogen
Gemini 4 Hydrogen is positioned as Google’s fastest and cheapest new flash model, reportedly running 2.5 times faster than Gemini 3.8 Flash while improving reasoning, coding, and agent use. Google says it introduces four reasoning modes, from low to X high, allowing users to trade off speed, cost, and depth of thought. The model is available through the Gemini app, Google Search, and Google AI Studio.
Strong benchmark showing for a flash model
On independent and internal comparisons, Hydrogen appears unusually competitive for its class. It was cited at No. 7 overall in one benchmark set, No. 5 on Agent Arena, and scored 54 on an intelligence index at X high reasoning, close to top larger models. It was also reported to beat Gemini 4 Argon on some tests and to approach Claude Sonnet 5.5 on several agentic and knowledge-work tasks.
Coding, reliability and multimodal strengths
Google says Hydrogen delivers more than 3 times the coding performance of 3.8 Flash on Terminal Bench 4.1, cuts fabricated claims by more than 5 times, and handles failed tool calls more reliably. It reportedly ranks first on a tax-agent benchmark, ties for first in legal research, and performs strongly in visual reasoning. One cited hallucination result put it at 8%, versus 43% for Sonnet 5.5 and 52% for GPT 6.1 Soul on the same evaluation.
Pricing is low, but not always the cheapest
Pricing was listed at $1.50 per million input tokens and $7.50 per million output tokens. That makes Hydrogen far cheaper than many frontier-class models and roughly 19% of the cost of Gemini 4 Argon for similar intelligence in one comparison. But it was also described as around 2.5 times more expensive than GPT 6.1 Soul in some high-reasoning use cases, showing that low headline pricing does not always mean lowest end-task cost.
Anthropic stresses safety ahead of IPO
Anthropic said it is willing to slow AI development for safety even if that leaves it behind rivals such as OpenAI. The company framed safety as central to maintaining public trust and long-term operating freedom. It did not announce an immediate slowdown, but said it would accept reduced development speed when safety demands it.
Anthropic reveals growth and bold health claims
During its roadshow, Anthropic reported $1.7 billion in quarterly revenue, more than 12 times year over year, with over 11,900 customers spending at least $100,000 annually. It also disclosed a $23.3 billion net loss, largely attributed to valuation-related accounting adjustments. Chief executive Dario Amodei argued AI could sharply accelerate progress against cancer, Alzheimer’s, heart disease, and mental illness within a decade, an ambitious claim that remains speculative.
Claude Meetings targets workplace collaboration
Anthropic also introduced Claude Meetings, which lets Claude join video calls as a visible participant, answer questions, draft documents and whiteboards live, and automatically produce notes and action items. The feature is starting with Google Meet and is rolling out in beta for Team and Enterprise plans. Anthropic also expanded enterprise controls with private-beta self-hosted managed agents.
Nvidia faces leadership and export scrutiny
Reports said Jensen Huang is out as Nvidia chief executive, with the company’s CFO serving as interim CEO, though the circumstances remain contested and sensitive. Separately, a forensic analysis claimed several Chinese AI labs may have used restricted Blackwell chips based on numerical signatures found in released model weights. The findings do not prove Nvidia knowingly violated export rules, but they raise pressure on supply-chain oversight at a critical moment for AI hardware.
OpenAI highlights a new cybersecurity model
OpenAI presented an internal cybersecurity model that scored 77% on Exploit Gym, ahead of cited results for Gemini 4 Carbon at 54.8%, GPT-6 Cyber at 52.6%, and Claude Mythos 5.2 at 39%. On PatchBench, it reportedly reached 79.1% versus 73.5% for GPT-6 Cyber. The model has not been publicly released, but the results suggest rapid progress in AI systems for vulnerability research and defense.
More launches across the AI field
ByteDance introduced Seedance 2.5 Flash, a cheaper and faster video model priced around $0.15 per second for 720p generation. A newer browser-agent system called Handoff Squared claimed major speed gains in shopping, booking, and checkout tasks after processing 5.6 million browser sessions and over 20 million transactions. Microsoft signaled new model announcements for Ignite, while the Gemini app also expanded to Apple CarPlay.
CONCLUSION
The latest releases show an AI market splitting into several high-stakes fronts at once: cheaper high-performing models, enterprise workflow tools, specialized cybersecurity systems, and tighter scrutiny over the chips that power them. The immediate question is not only who can build the strongest AI, but who can scale it safely, affordably, and with dependable infrastructure.
Ask a question