What the Agent Could Reach: DeepSeek Harness Let an Agent Switch Off Its Sandbox, ZCode Uploaded Git History, Grok Bot Has No Price

Act before next issue

On our radar

• OpenRouter's Business tier: an 8% fee buys EU and US in-region routing — OpenRouter's pricing page lists two pay-as-you-go tiers: Standard at a 5.5% platform fee and Business at 8%, which adds EU and US in-region routing and up to 1,000 workspaces. Our OpenRouter verdict stays conditional.

• Claude Code Projects relaunch as multi-agent cloud threads — Anthropic redesigned Projects on 17 September: a coordinator fans work out to parallel cloud threads, each on its own branch and repo copy. It is a beta for select Pro and Max subscribers, and parallel threads hit usage limits faster.

• Cowork merges into Claude; Claude Docs and Slides arrive in beta — Anthropic is folding Cowork into the main Claude app, starting with Pro and Max over the next few weeks. No price change is announced. The setting that matters is ask-before-acting: leave it on.

• Claude now leads 26% of Anthropic's own AI R&D — Anthropic's figures, as of August 2026: none of that work is fully autonomous, about 30,000 agents run at once, and of roughly 100,000 flagged transcripts a week, about 50 reach a human.

• OpenAI announces Sponsored Agents and ad integrations with HubSpot and Shopify — Announced 16 September. HubSpot says you can build, budget and launch ChatGPT Ads inside HubSpot, with ROI reporting. We could not open OpenAI's own post, so we cover only what we could verify.

• In review, not yet published — None of this week's news drafts has passed our source review, so we report no figures or quotes from them yet: the Claude Sonnet 5.5 launch, GPT-6.1 Sol, ChatGPT's Pro tiers, OpenAI Dots, the DevDay platform changes, the MCP Python SDK OAuth flaw, the OpenCode serve flaw, the GitLab AI Gateway CVE, coding agents posting screenshots to public GitHub, Claude Code mods, OpenAI's Cursor help page, Gemini 4 Argon, Eleven v4, Gemini 2.5 Pro's access limit, Apple's Full Disk Access change and OpenAI's DNS sandbox report. Each is written up when it publishes.

• Meta denies Muse read a user's Messages without Full Disk Access — TechCrunch, 30 September, quotes Meta saying its permission steps cannot be circumvented; the user says Full Disk Access was off. Nothing has been reproduced. Held until an independent reproduction or a Meta fix.

• An FTC probe into OpenAI, Anthropic and others — CNBC, 30 September: an agency spokesperson confirmed the probe and declined to name the other companies. No agency document or scope yet. Held until the FTC publishes one.

• Developers say OpenAI and Anthropic safeguards flag routine work — VentureBeat, 30 September, quotes four named developers; no setting or policy change was announced. Held until a vendor changes thresholds, adds an opt-out or publishes intervention rates.

• Fireworks Ember-1 and GPT-Synopsys — Ember-1 is a research preview with two-week serverless access and Fireworks' own benchmarks; GPT-Synopsys, a chip-design model OpenAI and Synopsys will co-develop, has no price, date or API. Both held until a reader can buy them at a listed price.

• Held on a headline or feed summary only — Not yet read, so none of these is written up: a toolkit said to find tokens and session data left by Codex, Claude Code and Cursor, Microsoft's report of a fully autonomous ransomware attack, a Red Hat benchmark of decision models, Stanford and Nvidia's CLM-8B, Upstage Solar Mini 4, Suno's Speech beta and the ds4 local inference engine.

• Open verdict watches, unmoved — Nothing primary moved OpenRouter's pending ownership change, DeepSeek's conflicting V4 Pro pages, Gemini 4 Argon's access (still limited to Google's Fairwind cyber-defender programme), Qwen3.8-Max open weights or Claude Haiku 5.5, which has not shipped. Each entry keeps its current rating.