Welcome back!

Sam Altman just confirmed GPT-5 is coming this summer. OpenAI may move toward continuous incremental upgrades, where breakthroughs will come more from new applications than from new model versions. Let’s unpack…

Today’s Summary:

• 🔥 GPT-5 confirmed for summer release

• 📖 MiniMax launches 1M-token open model

• 📘 Anthropic multi-agent blueprint

• ⚠️ OpenAI-Microsoft tension increases

• 💼 McKinsey says AI needs agent workflows

• 🌍 Google rolls out Gemini 2.5 in production

• 🛠️ 2 new tools

TOP STORY

Sam Altman confirms GPT-5 release this summer

The Summary: Sam Altman says the GPT-5 release is coming this summer. On a new OpenAI podcast, he questioned whether model versioning even makes sense anymore, and suggested that the AGI definitions are already outdated as model development becomes more continuous. The conversation reveals how OpenAI is thinking about the next phase of AI.

Key details:

• OpenAI is debating whether to abandon model versioning and turn updates into a continuous stream of a single system

• Memory is where ChatGPT has really leveled up, Altman says it now feels like it “knows your life” and gives better answers with less input

• The latest update of Operator using o3, able to watch users’ screens and complete tasks on its own, stunned even OpenAI insiders, many privately say it was their personal “AGI moment”

• Altman defines superintelligence not as general chat ability, but as the power to discover new science

• OpenAI’s stance on ads is cautious: Altman says inserting paid influence into the model’s output would break trust instantly

Why it matters: AI starts to feel nonlinear: frontier models bring more incremental improvements, but agentic tools are doing things no one expected. Users are experiencing AGI-like moments not because a new model has arrived, but because AI is applied in novel ways to real-world workflows.

FROM OUR PARTNERS

No-Fluff Finance for the Curious

Smarter Investing Starts with Smarter News

The Daily Upside helps 1M+ investors cut through the noise with expert insights. Get clear, concise, actually useful financial news. Smarter investing starts in your inbox—subscribe free.

Join 1M+ readers today.

MINIMAX

MiniMax M1 pushes open models into the 1M token era

The Summary: MiniMax has released M1, the first open-source AI model with a one million-token context window. Built with a hybrid-attention design and a novel reinforcement learning method, M1 is capable of deep reasoning tasks and challenges Gemini 2.5 Pro in benchmarks while being fully open-source and API-accessible at low prices.

Key details:

• Supports 1M token input and 80K output, more than proprietary models like Claude 4 and OpenAI o3

• Uses Lightning Attention and a new RL algorithm requiring only 30% of DeepSeek R1’s compute for deep reasoning

• Beats Gemini 2.5 Pro on TAU-Bench, which tests agentic tool use

• Early testers are impressed by the model hardware efficiency and low training cost

• MiniMax, backed by Alibaba and Tencent, is exploring a $3B IPO in Hong Kong to fuel its expansion beyond M1

Why it matters: MiniMax lowers the barrier for long-context reasoning AI, putting high memory agents within reach of developers and researchers. It demonstrates that frontier performance for long-context reasoning can be now achieved efficiently in local models.

Try it on HuggingFace

FROM OUR PARTNERS

A Smarter Way to Read the News

Seeking impartial news? Meet 1440.

Every day, 3.5 million readers turn to 1440 for their factual news. We sift through 100+ sources to bring you a complete summary of politics, global events, business, and culture, all in a brief 5-minute email. Enjoy an impartial news experience.

Join for free today!

ANTHROPIC

Anthropic shares how Claude uses multiple AI agents

The Summary: Anthropic has shared the blueprint behind Claude’s Research feature, which uses a swarm of Claude agents to run web searches and data analysis in parallel. A lead agent plans the research, while subagents gather information independently. Tests show this multi-agent setup beats single-agent by over 90%. The writeup offers deep insights into the architecture and tradeoffs of multi-agent systems.

Key details:

• The system uses Claude Opus 4 to plan tasks, and many Claude Sonnet 4 subagents running the searches and analysis in parallel

• Multi-agent runs consume 15x more tokens, driving quality gains

• Claude can revise flawed prompts, acting as its own prompt engineer

• Builds on ideas from Anthropic’s Dec 2024 classic write-up Building Effective Agents, a guide widely referenced by developers for building real-world agents

Why it matters: Multi-agent design is becoming the default for complex tasks, especially research, strategy, and analysis. Anthropic describes AI agents as structured systems that plan, delegate, and adapt autonomously. Their findings suggest that coordination and prompt clarity may matter more than model size.

QUICK NEWS

Quick news

• Tensions grow between OpenAI and Microsoft

• McKinsey says AI won’t pay off unless companies build their workflows around agents

• Google pushes Gemini 2.5 into general availability

TOOLS

🥇 New tools

• Flowstep - Create flowcharts, wireframes and designs

• Copilot Vision - AI that sees your screen and guides you through apps