Three stories pointing at the same shift: OpenAI is buying Ona to give Codex persistent cloud environments, Anthropic is walking back a quiet policy that was downrouting Claude Fable 5 on research tasks, and Xiaomi's MiMo team open-sourced MiMo Code — a terminal coding harness that beats Claude Code at SWE-bench with comparable models. The agent runtime is now the battlefield, not the model.

OpenAI buys Ona 🤝, Anthropic backtracks 🔁, Xiaomi's MiMo code 👨‍💻

Hey guys, Mr. Technology here. Three stories from this week, all pointing at the same thing: agents are now the substrate the big labs are building around, and the shape of that substrate is moving fast.

What You Need to Know: OpenAI is acquiring Ona to put persistent, customer-controlled cloud environments inside Codex. Anthropic is walking back a quiet policy that was downrouting Claude Fable 5 requests researchers didn't like. And Xiaomi's MiMo team has open-sourced MiMo Code, a terminal coding harness built on top of OpenCode that posts better SWE-bench numbers than Claude Code on the same model.

Why It Matters

Agents need real estate. OpenAI is buying Ona specifically to give Codex "long-running, customer-controlled" environments — i.e., a place where agents can sit for hours or days without losing state. If you're building on Codex, expect tighter integration with persistent execution surfaces in the next 60–90 days.
Silent downrouting is now a public problem. Anthropic's "sabotage" policy rerouted Fable 5 traffic to a weaker model on tasks the company didn't like. Researchers spotted it, the press picked it up, and Anthropic reversed. Translation: the frontier labs are now in a trust-but-verify era, and external evals matter more than their marketing.
Open-source harnesses are now benchmark-competitive. Xiaomi is reporting SWE-bench Verified 82% vs. Claude Code's 79% with the same model underneath. The model is no longer the only moat — the harness, the memory, the checkpointing are. That changes the build-vs-buy math for everyone shipping a coding agent.

OpenAI is buying Ona to give Codex a place to live

OpenAI's acquisition announcement frames Ona as a secure-cloud-execution and orchestration layer. The technical pitch is persistent, customer-controlled environments where agents can keep working across long sessions — exactly the gap you've probably felt if you've ever had a Codex session lose its place at step 47 of a 200-step task.

Ona's founders posted on LinkedIn calling the move "our life's work just got bigger and more important." The Ona engineering team joins the Codex org. Pricing, packaging, and the migration path for existing Ona customers haven't been disclosed yet — that's the part worth watching for anyone currently paying for Ona.

What it signals: OpenAI is conceding that the model alone isn't enough. Execution environment, memory, and orchestration are part of the product now. Expect the Codex roadmap to start looking more like Replit + Claude Code + Devin and less like a chat interface.

Anthropic walks back the Fable 5 "sabotage" policy

Engadget reported that Anthropic has made its frontier safeguards visible after researchers noticed Claude Fable 5 was either refusing or silently downrouting on three categories of work: training competing models, debugging AI-generated code, and optimizing neural architecture. The earlier policy was "discreet" — Anthropic wasn't telling users when the swap happened.

Two things to note. First, the workaround Anthropic chose wasn't a refusal with a reason, it was a quiet redirect to a weaker model. That burns the user's tokens and gives them wrong answers, which is a much worse failure mode than a hard refusal. Second, the categories Anthropic picked — training competitors, AI code debugging, architecture search — are exactly the work frontier AI researchers actually do. The policy was getting in the way of the people it was supposed to serve.

Anthropic's response is to make the safeguards visible. Whether that means an explicit "this request was modified" notice in the API response, or just better docs, isn't clear yet. Either way, anyone running serious evals on Fable 5 should re-test now that the policy is public.

Xiaomi open-sources MiMo Code — and it beats Claude Code at agentic tasks

Xiaomi's MiMo team released MiMo Code V0.1.0 under MIT on GitHub. It's a fork of OpenCode with three additions that matter: a cross-session memory system (SQLite FTS5 underneath), an independent "checkpoint-writer" subagent that takes notes without pausing the main agent, and a /dream command that consolidates old sessions every ~7 days.

The benchmark numbers Xiaomi published are striking. With the same MiMo-V2.5-Pro model, MiMo Code beat Claude Code + Sonnet 4.6 on:

SWE-bench Verified: 82% vs. 79%
SWE-bench Pro: 62% vs. 55%
Terminal Bench 2: 73% vs. 69%

Read that carefully — the model is the same or comparable, the harness is what changed. That's the part that should make every coding-agent team sit up. If the harness is the moat, then the open-source community will eat that moat from underneath in months, not years. Install is a single curl pipe (curl -fsSL https://mimo.xiaomi.com/install | bash) — Xiaomi is going for adoption speed over enterprise hand-holding.

The catch: Xiaomi's survey is self-reported (576 developers, internal beta), and Xiaomi has a vested interest in their harness looking good. Treat the benchmark deltas as directional, not definitive. But the direction is real — the agent harness layer is open for disruption.

The Take

Three stories, one thread. The frontier labs are now competing on substrate — execution environment, memory architecture, transparent policy — not just on the next model checkpoint. That's good for builders: when the labs compete on plumbing, the moats shift to your integration work, your data, and the specific workflows you wrap around the model.

The Xiaomi release is the one to watch. If a 200-step task is now solvable by an open-source harness on commodity models, the "agentic coding" market re-prices overnight. The labs know this. That's why OpenAI is buying Ona and Anthropic is making its safeguards visible — they're not just shipping models anymore, they're selling the whole runtime.

Quick Summary

OpenAI is acquiring Ona to give Codex persistent execution environments. Anthropic reversed a quiet policy that was downrouting Claude Fable 5 on research tasks. Xiaomi open-sourced MiMo Code, a terminal coding harness that posts better SWE-bench numbers than Claude Code with comparable models. The agent runtime is now the battlefield, not the model.

Sources:

What do you think? Drop your thoughts in the comments below! 👇

Source: TLDR | mr.technology — The Master Skill Index

OpenAI buys Ona , Anthropic backtracks , Xiaomis MiMo code