OpenAI has finally named what comes after GPT-5.6 Sol.
It is called Astra, and OpenAI is describing it as its “next major model.” It is not publicly available yet, and there is no confirmed release date. But the first official evidence of its capability is already here.
This is not another chart showing a model scored a few points higher on a benchmark. OpenAI says an internal Astra model generated results that resolved or substantially advanced ten long-standing problems in mathematics and theoretical computer science.
That is the headline this week.
🔥 The Big One
OpenAI’s Astra is moving from answering questions to producing new knowledge
OpenAI revealed Astra in a research update on 1 August, calling it the company’s next major model.
The internal model worked across high-dimensional geometry, coding theory, complexity, group theory, cryptography and combinatorics. Its results included a construction establishing the existence of non-sofic groups, a disproof of Connes’s rigidity conjecture and progress on problems connected to post-quantum cryptography.
“The results were achieved by an internal version of Astra, our next major model.” (OpenAI)
OpenAI says Astra generated the mathematical arguments. Humans then prepared them into manuscripts with the model, before Astra formalised each argument as a Lean certificate. OpenAI has published those certificates so the work can be checked rather than accepted on the strength of a demo.
Read OpenAI’s research update →
⚡ What shipped this week
1. Vercel turned v0 into app-building infrastructure
Vercel released the v0 API, giving developers headless access to the agent behind v0.
A product can start a persistent workspace from a prompt, stream the build, run the application inside Vercel Sandbox and return an embeddable preview. Follow-up instructions continue refining the same project.
The bigger opportunity is not another general app builder. It is focused generators inside existing products: a course platform creating an interactive exercise, or a property tool turning listing data into a working page.
The winning products will hide the blank prompt, collect the right context and return an outcome shaped for one specific job.
2. Spotify built a command centre for 50+ parallel coding agents
Spotify introduced Xirp, a vendor-neutral environment for managing coding sessions across Claude Code, Gemini CLI, Codex and other harnesses.
Each session gets its own worktree, so dozens of agents can operate on the same codebase without colliding. Context stays separate from the model, allowing engineers to switch tools mid-project without rebuilding the working state.
Paired with Spotify Portal, Xirp also injects organisational context: service ownership, dependencies, architecture and prior decisions. After a session, transcripts and metadata flow back into Portal for the next engineer or agent.
Spotify says thousands of its engineers have used Xirp across more than 36,000 sessions, and the public product is currently offered as a beta.
This solves the problem most agent demos avoid: generation gets faster, but coordination, context and duplicated work become the new bottlenecks.
3. VS Code is becoming a control plane for coding agents
VS Code 1.132 expands its Agent Host, a dedicated process that can run multiple agent harnesses including Copilot, Claude and Codex.
The same session can connect across VS Code windows, while activity indicators surface file changes, browser work and subagents. The feature is progressively rolling out rather than universally available.
The IDE is accepting that developers will use several agents. The valuable layer becomes the place that coordinates sessions, permissions, changes and human attention.
4. GitHub Copilot can remember your project and use local models
GitHub added persistent Copilot memory and Ollama support to JetBrains.
Memory reduces the repeated briefing tax between sessions, while Ollama lets developers select local models through BYOK. The caution is that remembered context becomes another dependency: keep it visible, correctable and scoped to the project.
🧰 Worth your time
-
Vercel Agent Plugins 1.0.0: A common package format for shipping Agent Skills and MCP servers behind one manifest.
-
Amazon Bedrock AgentCore runtime instances: Persistent managed compute, GPU support and agent sessions lasting up to 14 days.
-
Coldtea: Visual QA, release testing and monitoring around coding agents—a reminder that faster generation increases the need for independent proof.
The practical challenge this week is simple.
Stop asking only whether a model can generate an answer. Ask how you would verify the result if the model were proposing something genuinely new.
Do you think Astra’s research results are a bigger leap than another coding or benchmark upgrade?
Hit reply and tell me.
I read every single one.
Talk soon PAPAFAM,
Sonny 👋🏼
👇🏽 Don't forget to follow me across socials!
Responses