This is the news. Agent dossiers — when to pick one, when to skip it — are on the agents page. Practical guides support your own testing and deployment decisions.
Cursor shipped Projects on 2026-09-12. A coordinator agent plans the work, delegates to implementing agents, and brings the finished result back; the project keeps a shared file set that syncs across cloud and local machines so research, artifacts, and learned preferences persist across sessions. Projects run on a dedicated cloud computer and remain alive when the user closes their laptop. Bundled features (Cloud Agents, subscriptions, self-hosted machines, Start from scratch) are covered in earlier Cursor articles; the genuinely new surface is the coordinator + shared-context model. Primary-source-verified against cursor.com/changelog/projects.
The Claude Platform release notes dated 2026-09-10 ship two Managed Agents changes. auto lets the server evaluate each agent or MCP tool call and run it, deny it, or pause for approval; agent.tool_use and agent.mcp_tool_use events report the decision in an evaluation field alongside evaluated_permission. ant beta:sessions connect (added in ant CLI v1.32.0) attaches the terminal to a live Managed Agents session; --web serves the Claude Console session viewer locally. After the 2026-09-19 source audit, broader auditability, retention, and human-confirmation defaults are labeled as conditional pending independent verification.
joshavant published 10 new GitHub Security Advisories on 2026-09-11. Two are rated High: GHSA-3mq7-q27j-mq7q (reusable exec approvals outliving their reviewed working directory) and GHSA-9m4p-cqp4-jppq (the WhatsApp login tool reachable from non-owner turns). The eight Moderate advisories cover Prometheus, Discord uploads, iOS deep-link logging, Browser relay DoS, OpenAI-compatible credential misrouting, Unicode workspaceOnly escape, file-transfer approval widening, and Slack file-download fail-open. Patched versions are spread across 2026.8.1, 2026.8.2, 2026.8.11, and 2026.9.3 — 'latest' alone is not enough; two advisories require explicit credential rotation. Documentation comparison; no patched version was installed.
Claude Code v2.1.269 (2026-09-11 19:17 UTC) ships 30+ items in one release. Six new capabilities: claude plugin eval (scored JSON+HTML report), /output-style [name] over Remote Control and headless, bashEditDiffEnabled for Bash tool file edits, OTEL_METRICS_INCLUDE_REPOSITORY VCS tagging, CLAUDE_CODE_GATEWAY_MODEL_DISCOVERY_TIMEOUT_MS, and CLAUDE_CODE_WORKFLOW_MAX_CONCURRENT_AGENTS (1-256). Plus prompt-cache resume fix, mid-thought resume cache-reuse fix, headless session reply-loss fix on model switch, and a sweep of ~25 operational fixes. Documentation comparison.
OpenAI's 2026-09-10 changelog published three substantial entries in one day. The Agents API moved to public beta — a managed Codex harness for session orchestration, context compaction, and tool execution. GPT-Live 1 went GA at $0.05 per minute of voice (billed per second) plus backend model and tool usage on the underlying OpenAI Agents API. Project API keys gained an expiration field (default 90 days, configurable) plus an organization-wide maximum-key-lifetime setting. Primary-source-verified against developers.openai.com; pricing quoted verbatim.
Claude Code v2.1.268 (2026-09-10 20:30 UTC) ships 30+ items in one release. Three secrets leaks patched: plugin / marketplace git-URL credentials, MCP ${VAR}-resolved tokens, and a workload identity federation recovery path. Plus a Web fetch deadline, a symlink deny-rule bypass fix, and two permission deny-rule bypasses. Documentation comparison against the GitHub release page; no firsthand run was performed.
OpenClaw v2026.9.4 (2026-09-11 07:22 UTC) closes the loop on v2026.9.3's safer-updates foundation with automatic rollback for compatible failed updates, ships a unified Plugins workspace, and lands v2026.9.3 retrospective coverage including the Node 22 end-of-life (Node 24.16+ or 26.1+ required) and the SQLite text-truncation hazard that came with staying on Node 22. Documentation comparison against the GitHub release page and CHANGELOG; no firsthand run was performed.
OpenClaw 2026.6.35 was tagged 2026-09-10 06:43 UTC and is explicitly labeled 'the final June 2026 Extended Stable (LTS) release.' The release hardens bundled provider and channel boundaries: untrusted response bodies are bound, oversized inputs are rejected before expensive work, and safe recovery is preserved when transports fail. Organizations still on the 2026.6.x line should plan migration to the 2026.9 series. After the 2026-09-19 source audit, auto-rotation / no-config / support-window certainty statements beyond what the release page establishes are removed and treated as conditional pending direct verification.
OpenAI Codex v0.154.0 stable was tagged 2026-09-09 22:37 UTC. The release adds GPT-6-Astra to the model picker and Amazon Bedrock catalogs, ships experimental --worktree / /worktree for isolated per-session checkouts, introduces inline questions that pause Codex without losing the user's draft, lets Windows sessions share a background Codex server with daemon lifecycle commands, completes Vim with R replace mode plus undo and dot-repeat, and preserves rich-text formatting in copied responses. Documentation comparison; --worktree is labeled experimental.
Claude Code v2.1.267 was tagged 2026-09-09 19:58 UTC. The release adds a top-level maxEffortLevel setting that caps effort on every provider (Bedrock, Vertex, Foundry) plus a --system-prompt-snapshot off flag for prompt iteration. The fix list is dominated by ~10 prompt-cache stability items that close cache-reuse breaks on resumed sessions, mid-session MCP reconnects, and /model switches, alongside a marketplace-entry backslash bypass fix and an expired-AWS / Google-credentials retry fix. Documentation comparison.
Anthropic shipped Claude Code v2.1.265 on 2026-09-08 at 20:37 UTC, followed 3 hours 18 minutes later by v2.1.266 at 23:55 UTC as a targeted regression fix. v2.1.265 is the larger of the two and covers plugin folder and marketplace work, telemetry, and a prompt-cache reliability sweep. v2.1.266 is a single-line v2.1.265 regression fix (the undocumented CLAUDE_CODE_USE_GATEWAY env var on its own was forcing Cloud-gateway sign-in). Documentation comparison.
OpenAI published GPT Image 2.5 Sunburst and Flare on the 2026-09-08 changelog. Both models add xhigh and max quality tiers, advertise inpainting as a supported feature, expose a tool path for image editing on the Responses API, and are listed at the published per-image pricing in the OpenAI pricing page. Primary-source-verified against developers.openai.com; no firsthand image generation was performed.
OpenClaw shipped v2026.9.2 stable on 2026-09-06 02:15 UTC. Highlights: chat offloaded outside the Gateway event loop; automatic updates preserve active settings, enabled skills, and default-agent ownership; full GPT-6 Astra support via OpenAI API-key or ChatGPT/Codex account with async tools, WebSocket steering, and /think ultra; active, queued, and delegated replies now survive Gateway restarts; backups reject corrupt archive headers and preserve text with embedded NUL characters; Swarm enabled by default; cross-agent session access opens to all-session visibility by default.
Anthropic added `ant apply` to the ant CLI on September 3, 2026 (CLI v1.30.0). It creates and updates agents, environments, skills, memory stores, and deployments from files in your repository, with a `claude-lock.json` lockfile for CI reproducibility. Path-based references between resources replace ID lookups; GitHub URLs can pin skills to a resolved commit; CI must authenticate with Workload Identity Federation. `ant apply` cannot adopt Console-managed resources without an export, skill deletion is destructive, and concurrent applies are unsafe. Documentation comparison, not firsthand test.
Cursor published Self-Hosted Machines, the most consequential enterprise change to its agent platform since Cloud Agents launched. Code, build outputs, and secrets now stay on machines running inside the customer's own infrastructure, while the agent drives tool calls locally. Eight platform partners (AWS Lambda, Cloudflare, Daytona, Modal, Namespace, Vercel, E2B, Coder) have reference templates; the legacy Kubernetes operator and WorkerDeployment Helm chart are deprecated for new deployments. Computer use on Linux and Mac is now available on self-hosted workers. Gated to Enterprise tier. Documentation comparison, not firsthand test.
Anthropic shipped two Claude Code releases in the last 24 hours. v2.1.260 (Sep 3 23:48 UTC) adds a `/diff` fullscreen panel; v2.1.261 (Sep 4 19:58 UTC) is the larger of the two — about 60 changelog items with three new user-facing capabilities (`bashOutputMaxChars`/`taskOutputMaxChars` up to 128K, `/skill-doctor` showing unused skills + context cost, `--append-subagent-system-prompt-file` for prompts too large for the command line), an "Organization policy" diagnostic in `/status` and `claude doctor`, two security-relevant behavior changes (dangerous-`rm` safety prompt now catches positional parameters and `sh -c "..."` scripts; auto mode no longer auto-approves URLs that pack content into a public diagram renderer's URL), and the API timeout default raised from 3 to 10 minutes. Documentation comparison, not firsthand test.
OpenClaw released v2026.9.1 stable on September 3, 2026 at 19:47 UTC. Eight highlight sections beyond the Aug 28 v2026.9.1-beta.1 article (gateway recovery + Codex 0.150.1 bundling + Linux installer fix): Mermaid blocks render in Control UI + macOS + iOS + Android, fresh installs go from `npx openclaw@latest` to web dashboard in one prompt (detects Claude Code / Codex logins), per-identity skill libraries on shared Gateways with ZIP import + per-identity share/publish, `openclaw update` rolls back when post-update Doctor fails and preserves config + secret references, the Gateway recovers at startup under load and with large agent rosters, Codex approvals stick (Allow Always durable for MCP tools), approval cards reach the originating Telegram topic with requester title, and Android matches the web Control UI (chat screen, sidebar, appearance, six-line composer, recording fallback). Documentation comparison, not firsthand test.
OpenAI changelog September 3, 2026 launches GPT-6 Astra ('our most capable model, built for the hardest end-to-end work') plus three new Responses API controls: async tool calling, mid-turn steering over WebSockets, and change-reasoning-effort-mid-conversation with prompt-cache preservation. Hard migration constraints: Responses API required for tool calling, no custom temperature/top_p/logprobs, no `none` reasoning effort, and misalignment monitoring that can stop a conversation for review. Documentation comparison, not firsthand test.
OpenAI shipped Codex v0.153.0 stable on September 3, 2026 at 01:39 UTC. Five new user-facing capabilities: Vim mode `u`/`Ctrl+R` undo/redo with full draft preservation (including pasted content and attachments), plugin CLI can list/install/remove plugins from remote marketplaces, earlier usage-warning threshold for Plus/Team users, `tui.auto_recap = false` opt-out for TUI auto-recaps, and TUI history shows complete patches plus background-terminal input plus individual completed commands. One subscription-only capability: `features.context_management.experimental_mode` (disabled by default; activates token-budget context, history notes, and `new_context` tool for ChatGPT Plus / Pro / Pro Lite sessions — API-key sessions excluded). Plus operational durability: TUI sessions reconnect after app-server drops preserving drafts and pausing uncertain submissions for review, Full Access skips Guardian reviews for confirmation-only actions, Guardian review history survives compaction/restarts/forks, remembered MCP tool approvals are app-account-scoped, rollout compression includes shared histories.
Anthropic pushed Claude Code v2.1.259 on September 2, 2026 at 22:33 UTC. Two user-facing capability additions: `managedMcpServers` managed setting lets organizations push HTTP/SSE MCP servers to every user (stdio transports excluded by design), and `--permission-prompts none` gives unattended headless hosts a deterministic deny-on-prompt mode while the active permission mode keeps deciding. Plus a Bash `Read()` deny-rule bypass fix that closed option-value bypasses (`--ignore-revs-file=.env`, `-f.env`, `@file`), git diff/grep file operands, `cd DIR && cat FILE` compounds, and `grep -r`/`cp -r` over directories holding denied files. Plus a concurrent-session state corruption fix that ended MCP-server-disappearing and workspace-trust-resetting regressions when multiple sessions ran against the same `~/.claude.json`.
OpenAI changelog September 2, 2026 separates traffic-increase rate limiting (429 slow_down) from temporary model overload (503 server_is_overloaded — new). Affects every OpenAI API client, lets autonomous-agent retry layers distinguish causes. Documentation comparison, not firsthand test.
Cursor shipped an event-driven subscriptions model for Cloud Agents in the August 19, 2026 changelog that nobody has unpacked: cloud agents now subscribe to event sources (PRs they create, Slack threads, scheduled tasks), wake when something happens, and hold a goal until it is met. Plus five harness improvements (custom modes, subagents on isolated VMs, /goal command, steering without interruption). The agent-to-production story is no longer a single prompt-to-live-URL flow — it is a long-running autonomous loop that closes itself.
Claude Fable 5.1 + Claude Mythos 5.1 launched September 1, 2026. 1M-token context, 128k max output, always-on adaptive thinking at the same $10/$50 per million input/output tokens as Claude Fable 5 — but cache reads priced at $0.25/MTok instead of $1/MTok. Three breaking changes (no forced tool use, thinking blocks tied to producing model, edited-history rejection) and five additive features (per-message effort, turn-scoped system messages, between-tool progress, 4x cache cut, C2PA provenance) ship today across the Claude API, Bedrock, Claude Platform on AWS, Google Cloud / Vertex AI, and Microsoft Foundry.
OpenAI shipped Codex v0.152.0 stable on September 1, 2026 at 02:13 UTC. Six new user-facing capabilities: Vim `/` and `?` search with `n` and `N` repeat, rate-limit banners with action buttons, credential-refresh progress in the TUI (including Bedrock reauth), MCP server names that accept `: @ / .` (unblocking scoped/package/FQDN registries), per-MCP-tool `output_token_limit` with consistent truncation across session resumes, and app-server thread/shellCommand timeouts longer than one hour. One breaking default change: planning tool is disabled by default (re-enable with `tools.update_plan.enabled = true`). One credential-theft fix: cloud task requests now reject untrusted backend URLs and disable redirects to protect saved credentials.