OpenAI ships an Agents API, opening its Codex agent harness to developers
OpenAI released an Agents API that gives applications direct access to the managed agent harness behind Codex, rather than leaving orchestration to the calling app. It handles sandboxed command execution, file editing, and MCP tool connections; keeps durable session state so a run can be paused and resumed later; and supports native subagent delegation for splitting work across parallel sub-tasks. That's a step beyond the existing Assistants/Responses APIs, where an app has to manage its own execution environment and session bookkeeping. Code samples in the docs are built against gpt-6-astra. Real constraints for now: data residency is US-only, and there's no Zero Data Retention option yet, which will matter for regulated users.
Sources & depth
Cognition launches SWE-2, a coding-agent model built on Kimi K3
Cognition (maker of Devin) released SWE-2, built by post-training Kimi K3 (2.8T parameters) with new reinforcement-learning techniques aimed at engineering judgment rather than raw pass rate. By Cognition's own numbers it needs 58% fewer steps than its predecessor SWE-1.7 while scoring higher: 50.0% on FrontierCode 1.1 Main (up from 42.0%), 73.0% on DeepSWE 1.1 (up from 37.7%), and 92.8% on Terminal-Bench 2.1. Cognition places it near GPT-5.6 Sol and the Fable 5/5.1 line on FrontierCode at a fraction of their price, and roughly a quarter of GPT-6 Astra's cost for slightly lower scores. It's live now across Devin Desktop, CLI, Web, and Fusion. All benchmark numbers above are Cognition's own — no independent reproduction exists yet.
Sources & depth
Meta launches Muse, a shopping-and-errands AI agent, amid a trademark scramble
Meta unveiled Muse on 2026-09-09, a proactive AI agent meant to handle tasks like shopping, email, and trip-planning on a user's behalf. The rollout got overshadowed by a trademark dispute: Meta reassigned the @muse Instagram and X handles — held by the rock band Muse since 1999 — to the new agent, pushing the band's own accounts to @museband with no clear public explanation of how the handles changed hands. Meta executives' launch posts then briefly tagged the band's account instead of the agent's, which a New York Times reporter called "the best bug in twitter history." It's the same pattern as Meta's earlier acquisitions of the @Meta and @threads handles from unrelated prior owners.
Sources & depth
GPT-6 Astra's short reasoning traces spark a 'looped transformers' debate
A technical analysis piece digs into reports (originating with The Information) that GPT-6 Astra "hides" its reasoning using looped transformers — reusing the same transformer blocks multiple times per forward pass instead of stacking distinct layers, a technique already visible in smaller open models like Nanbeige4.2-3B (which runs 22 blocks twice), Universal Transformers, and Mixture-of-Recursions. The author argues Astra's shorter visible reasoning traces more plausibly reflect real capability gains — a more capable model needing less scratch work, the way a strong student uses less notepaper — than deliberate obscuring, and points to an OpenAI chief scientist's comment that architecture changes aren't behind current difficulties in monitoring model reasoning. There's no official OpenAI confirmation of Astra's actual architecture either way.
Sources & depth
Qwen 3.8's reasoning shifts oddly toward GPT-5.5 Pro's answers — a distillation clue
An independent test fed several open models the first ~1% of GPT-5.5 Pro's reasoning trace before their own answer, to see whose answers moved closest to GPT-5.5 Pro's. Across 45 problems, Qwen 3.8 moved far more than the others: +18.2 percentage points of answer overlap with the prefilled trace, rising to +27.0pp on STEM problems alone, while DeepSeek, Kimi, and a model called Inkling barely shifted. The author reads the size of that gap as a hint that Qwen 3.8 may have trained on outputs from GPT-5.5 Pro or a closely related OpenAI model — a provenance signal about training data, not a capability score.
Sources & depth