<h2>OpenAI Codex becomes a broader agent platform</h2>
<p>OpenAI has pushed Codex well past its original role as a coding assistant. The latest update adds background computer use (Mac only for now), an embedded browser, image generation via gpt-image-1.5, 90 new plugins, and persistent “heartbeat” automations that run on a schedule without losing thread context.</p>
<p>On the developer side, Codex now handles GitHub review comments, supports multiple terminal tabs, and offers SSH connections to remote dev boxes in alpha. A new Chronicle feature — currently a research preview — periodically captures screenshots, sends them to OpenAI’s servers, and stores text summaries as local Markdown files so Codex can passively build context about what you’ve been working on without you having to re-explain it each session.</p>
<p>Chronicle is worth pausing on. Unlike Microsoft Recall’s local, encrypted approach, Chronicle processes data in the cloud and stores memories as plain text. It’s blocked in the EU, UK, and Switzerland — a fairly clear signal of GDPR incompatibility. OpenAI acknowledges risks including prompt injection and sensitive data exposure, and recommends manually pausing the feature during sensitive work. Whether a cloud-processing, trust-dependent model survives regulatory scrutiny is an open question.</p>
<p>Codex now has over 4 million weekly users, up from 3 million recently, growing roughly 1 million per month. OpenAI says more than 80% of its own employees use Codex for tasks beyond coding — drafting briefs, reviewing contracts, triaging Slack. The framing internally is less “coding tool” and more “command center for AI agents.”</p>
<p>Pricing has expanded too. Codex is temporarily available on the free ChatGPT tier, with higher rate limits for paid plans and a pay-as-you-go seat option for business and enterprise.</p>
<p>Practically, developers are finding distinct use cases across tools. Codex works well for 0-to-1 prototyping and running parallel workstreams. Claude Code tends to win for PR reviews and correctness-critical work. Cursor still has a place when you want more direct control over a larger codebase.</p>
<hr>
<h2>Kimi K2.6: Moonshot AI’s open-weight agentic model</h2>
<p>Moonshot AI has open-sourced Kimi K2.6, a model built specifically for long-horizon agentic coding. The headline benchmarks are 80.2% on SWE-Bench Verified and 58.6% on SWE-Bench Pro. It supports Agent Swarm architectures scaling to 300 sub-agents executing 4,000 coordinated steps simultaneously.</p>
<p>The launch blog includes some striking demos: a 13-hour continuous run with over 1,000 tool calls and 4,000+ lines of code modified, resulting in a 185% throughput improvement on a financial matching engine. In another run, the model deployed a Qwen3.5-0.8B model locally on a Mac, then coded and optimized a Zig-based inference engine to 193 tokens/sec. There’s also a “Claw Groups” research preview for heterogeneous human-agent collaboration, and a proactive background agent called OpenClaw capable of 5-day autonomous operations.</p>
<p>The model is available through Augment Code (their first open-weights model, served via Fireworks AI), Vercel’s AI Gateway, the AI/ML API, and Kilo Code’s VS Code and JetBrains extensions.</p>
<h3>How it actually performs</h3>
<p>A head-to-head comparison against Claude Opus 4.7 on a complex workflow orchestration spec — DAG validation, atomic worker claims, lease expiry recovery, pause/resume/cancel, SSE event streaming — put the gap in perspective. Claude Opus 4.7 scored 91/100; Kimi K2.6 scored 68/100, but at roughly 19% of the cost.</p>
<p>Both models passed their own test suites, but targeted code review found one real bug in Claude Opus 4.7 and six in Kimi K2.6. The specific gaps: non-global claim ordering, replay-only SSE with no live streaming, expired lease acceptance on complete/fail endpoints, wrong HTTP status codes, overly narrow validation, and a broken build path.</p>
<p>The practical takeaway is that Kimi K2.6 is viable for scaffolding and prototyping where cost matters more than correctness guarantees. For state-machine logic where bugs are expensive, Claude Opus 4.7 is the safer bet.</p>
<p>Early user reactions are mixed in an interesting way — some report noticeably better personality and context retention when switching to K2.6 for agent workflows, while the benchmark comparison suggests it still trails on correctness for complex specs. Both things can be true.</p>


Collections

OpenAI has expanded Codex into a general-purpose agent platform with background computer use (Mac only), an embedded browser, image generation, 90 new plugins, and persistent scheduled automations. A new Chronicle feature passively captures screenshots and builds context in the cloud — raising GDPR concerns and privacy risks. Codex now has 4 million weekly users, with 80%+ of OpenAI employees using it for non-coding tasks. Moonshot AI has open-sourced Kimi K2.6, an agentic coding model scoring 80.2% on SWE-Bench Verified and supporting Agent Swarm architectures with up to 300 sub-agents. A head-to-head comparison against Claude Opus 4.7 shows K2.6 scoring 68/100 vs 91/100 on a complex workflow spec, but at roughly 19% of the cost. K2.6 is practical for scaffolding and prototyping where cost matters; Claude Opus 4.7 remains stronger for correctness-critical work.

OpenAI's Codex expands beyond coding with computer use, image generation, and scheduled automations