Tech • AI • Robotics
Huge GPT-7.0 Bell Leaks! Anthropic's New Model, AI Extinction, ChatGPT Images 2.5 & More! AI News
9.5/10TL;DR
OpenAI, Anthropic, and DeepSeek are signaling a new phase of AI competition marked by more capable frontier models, larger multi-agent research systems, rising economic stakes, and sharper warnings from insiders about safety risks.
KEY POINTS
OpenAI teases a larger frontier model
OpenAI has officially previewed a checkpoint codenamed Bell, described as substantially more capable than GPT-6 Astra and still in training. Early benchmark chatter suggests Bell Low already outperforms Astra Max at a lower reasoning setting, implying a jump that could exceed the earlier transition from Soul to Astra. The model is viewed as a possible step toward a GPT-6.5 or GPT-7 class system later this year.
10,000-agent push at Navier-Stokes
In a striking research test, OpenAI reportedly used about 10,000 AI agents powered by the new model to produce a proposed solution to the Navier-Stokes Millennium Prize Problem, one of mathematics’ most difficult open questions. The run lasted 88 hours, exchanged 2.7 million messages, generated roughly 130 billion output tokens, and operated under a spending cap of about $1 million. The resulting proof reportedly argues that, under certain conditions, a fluid can develop a finite-time singularity in which velocity grows without bound.
Proof remains unverified
The output is being treated as a proof of concept rather than a settled mathematical breakthrough. GPT-6 Astra reportedly spent another 17 hours formalizing and checking the result in Lean, and the full AI-generated thesis is said to run about 166 pages with figures and diagrams. Independent scrutiny from mathematicians would still be required before any claim that the problem has truly been solved could be accepted.
A glimpse of AI research labs made of agents
The larger significance may be less about one proof than about the method. Instead of a single model answering a hard question, the experiment points to a future in which thousands of coordinated agents function like a digital research institute, compressing years of work into days. If that pattern holds, frontier AI could increasingly be measured by how effectively labs orchestrate agent swarms, not just by raw single-model scores.
Costs are high now but falling fast
The $1 million price tag is large, but recent model history suggests frontier capabilities can rapidly become cheaper. Systems that once required massive compute budgets to reach top benchmark scores have later become available at consumer-level prices. That trend raises the possibility that today’s elite, expensive agentic research workflows could become broadly accessible within a short period.
Anthropic models AI’s economic impact
Anthropic’s economics team has released a model exploring how advanced AI could affect jobs, wages, productivity, and growth by 2030, drawing on responses from more than 10,000 Americans. It breaks occupations into tasks and classifies whether AI is likely to assist, automate, leave unchanged, or create new work. Across modest, substantial, and extreme adoption scenarios, output grows, but higher capability levels automate more knowledge work and intensify questions over who captures the gains.
Insider warning sharpens safety debate
A former pre-training researcher, Jacob, who worked at both Anthropic and OpenAI, has resigned while accusing major labs of racing toward self-improving superintelligence without sufficient responsibility. He argued that private concern inside leading companies includes fears that advanced AI could become catastrophically dangerous before the end of the decade. Another employee was cited as placing the chance of AI causing human extinction within ten years at above 10%, an extraordinary estimate that underscores the severity of internal concern even if such forecasts remain highly contested.
Image models keep advancing
OpenAI has also launched GPT Image 2.5, with faster generation, better fidelity, and stronger consistency during repeated edits. A key improvement is more precise editing that preserves core image features across multiple changes, addressing a long-standing weakness in generative image systems. Comparative tests reportedly show the new model maintaining background, texture, and object identity more reliably than GPT Image 2.
DeepSeek prepares for a bigger stage
DeepSeek is reportedly preparing for a possible public listing on Shanghai’s STAR Market, with an IPO process that could begin later this year. The company is also said to be raising capital at a valuation of around $75 billion, only months after a $7.4 billion raise. That rapid ascent reflects how quickly leading AI developers, including those in China, are scaling alongside growing interest in lower-cost, high-performance model variants such as DeepSeek 4.1 Flash.
CONCLUSION
The latest moves suggest the AI race is no longer only about better chatbots, but about large-scale agent systems, economic disruption, and governance under mounting pressure. Whether the next breakthroughs deliver broad benefits or sharper risks may depend as much on verification and safety discipline as on raw capability.
Explain this