OpenAI Readies GPT-4.1 with 1M-token Context and Live Memory

OpenAI is preparing a multi-pronged rollout led by GPT-4.1, an enhanced multimodal model constructing on GPT-4o, with scaled-down variants like o4-mini, o4-mini-high, and nano versions. Also surfacing in ChatGPT infrastructure are o3 and a compact o4-mini model. Though CEO Sam Altman cautioned that these models aren’t launching immediately, the infrastructure suggests release is imminent—pending capability reliefs. Alongside the models, Altman might need also hinted at a brand new development about “quasar alpha,” a context window upgrade reportedly supporting as much as 1 million tokens.

In tandem, OpenAI has expanded ChatGPT’s memory to incorporate entire conversation histories, enabling it to recall and adapt across past chats without prompt engineering. Noam Brown described memory in language models as greater than a functional upgrade, framing it as a fundamental shift in user interaction.

OpenAI has also introduced the Pioneers Program, a bid to revamp AI benchmarking by co-creating domain-specific evaluations with startups in fields like law, finance, and healthcare. These benchmarks—intended to guide reinforcement fine-tuning and model improvements—might be released publicly. Critics, nevertheless, warn that the corporate’s role as each model maker and evaluator risks blurring the road between progress and self-interest.

The backdrop to those launches is OpenAI’s countersuit against Elon Musk, alleging a campaign to discredit the corporate because it seeks to finalize a $40 billion funding round and transition right into a capped-profit entity. Read more.

Google has expanded its AI developer ecosystem with a set of recent tools and models focused on agentic computing, cost efficiency, and full-stack app development—highlighted at Cloud Next 2025.

Agent Development Kit (ADK) is now open source, providing a framework for constructing hierarchical, multi-agent systems. ADK supports modular design, dynamic routing, multimodal inputs, and integrations with Vertex AI and LiteLLM. It also offers built-in evaluation tools and simplifies deployment via Vertex AI Agent Builder or containers.

Gemini Code Assist enters preview with agentic capabilities. These agents can now autonomously translate code, generate apps from specs, manage Kanban-style workflows, run tests, conduct reviews, and handle migrations—moving toward more self-directed software engineering.

Firebase Studio, powered by Gemini and built on Code OSS, enables in-browser, no-setup app development. It supports multilingual frameworks, imports from Git repos, and natural language prototyping, with deployment via Firebase App Hosting and Cloud Run.

Gemini 2.5 Flash, a reasoning-optimized model, balances performance with low latency and price. Available soon in Vertex AI, it’s built for real-time, high-volume use cases like customer support. On-prem deployment via Google Distributed Cloud and Nvidia Blackwell will follow in Q3.

Google introduced its seventh-generation Ironwood TPUoptimized for inference and designed to run AI models. Available in 256-chip and 9,216-chip clusters, Ironwood delivers 4,614 TFLOPs peak performance. Each chip is provided with 192GB of RAM and seven.4 Tbps bandwidth, making it Google’s strongest and energy-efficient TPU. Ironwood will integrate with Google’s AI Hypercomputer for high-scale workloads. Read more.

Humanoid robots are finding recent ground beyond labs and logistics. Boston Dynamics’ Atlas has entered film production, assisting with camera operations alongside WPP and Canon. Trained using synthetic data generated through Nvidia Cosmos simulations, Atlas can carry 20 kg and maintain stability in awkward positions—ideal for long, repeatable shots or filming in hard-to-reach environments.

Meanwhile, Unitree’s $16K G1 robot—built with 43 actuated joints and trained via imitation learning—is branching from flips and kung fu to boxing, with a livestreamed robot fight within the works. The G1 also demonstrates agility on uneven terrain and resilience under physical disturbances. Priced at $16,000, the G1 undercuts high-end models like Boston Dynamics’ Atlas, offering a low-cost entry point. Read more.