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.
v2.1.282 ships three security fixes that change the threat model for managed Claude Code deployments: a macOS repository-symlink path traversal that lets CLAUDE.md read macOS /Network through ../.vol-style paths, a managed-settings boolean-lock-key typo bypass that silently disabled disableClaudeAiConnectors / allowManagedPermissionRulesOnly, and a plugin-manifest self-approval bypass under allowManagedPermissionRulesOnly. Plus store.readiness_grace_seconds so /readyz can ride through a Postgres failover.
On September 22, 2026 Anthropic shipped Claude Opus 5.5 at $4/$20 per MTok, 1M context, always-on adaptive thinking, and $0.20/MTok cache reads. Claude Code v2.1.280 made it the default Opus model the same day, and pushed the Pro and Team Standard plans over to Opus as their default. Four breaking API changes ship with the model — thinking cannot be disabled, forced tool use returns 400, thinking blocks are tied to the model and conversation, and the older computer_20251124 tool is rejected on Claude API and Google Cloud. Here is what an agent stack needs to know before the next request.
On September 22, 2026 at 00:50 UTC, Anthropic returned elevated error rates simultaneously on Claude Opus 5, Claude Fable 5 and 5.1, and Claude Mythos 5 and 5.1 for one hour and twenty minutes. Agent stacks that routed everything through Anthropic's top tier stopped. Independent user-impact confirmation from Downdetector, Reddit, and two news outlets. Recommended action: audit your routing table, wire Status-page alerts, and test the fallback before you need it.
OpenClaw v2026.9.5 stable was tagged 2026-09-19T08:40:59Z. The release ships Atomic Updates that check the next version before switching over, plugin installation that no longer requires a Gateway restart, read-only conversation sharing and durable conversation archives, and GPT Live in meetings and phone calls. Fresh OpenAI / ChatGPT / Codex setups default to Astra as the primary model. Documentation comparison against the v2026.9.5 release notes, changelog, and atom feed; no firsthand Gateway run was performed against this release.
Claude Code v2.1.278 (2026-09-19 03:10 UTC) ships two changelog items: the auto-mode safety classifier now runs on Anthropic's side on Claude API, Enterprise, Bedrock, Vertex, Foundry, and Claude apps gateway sessions — and Anthropic does not bill for those classifier requests. The classifier stays client-side (and billed) on Pro / Max / Team plans and when an LLM gateway in the path strips or rewrites request fields. A new /status 'Auto mode server' row shows Enabled / Disabled. Documentation comparison against the GitHub release page and the new code.claude.com/docs/en/auto-mode-classifier-billing page.
Correction notice. The previously published v2026.9.3 brief attributed 22 reliability, security, and channel fixes to the v2026.9.3 release. A source-audit cross-check against the GitHub release bodies found that 20 of those 22 cited PRs resolve to the v2026.9.2 release body, not v2026.9.3. The 1,247-PR commit-range context and 22-of-22 count also belong to v2026.9.2. The previous article has been replaced with this correction notice; the v2026.9.3 verified deltas (safer updates, performance, Skill Workshop, browser tabs, account priority, conversation sharing, searchable meeting library, optional team activity reports) are listed from the GitHub release body.
Sixteen published items in seven days. Three themes carried the week: a security sweep across Claude Code and OpenClaw, OpenAI's quietest governance week of the year, and a Cursor rebrand of multi-agent coordination. Plus Anthropic Chrome session coverage in the Compliance API, the Node 22 death at the OpenClaw door, and Claude Code's seven-versions-in-seven-days cadence. Prices, model IDs, version numbers, and changelog text are surfaced verbatim; inferred claims are labeled.
Claude Code v2.1.277 (2026-09-18T18:06:32Z) ships roughly four additions — AGENTS.md fallback when CLAUDE.md is absent, CLAUDE_GATEWAY_PROXY_IS_EGRESS_BOUNDARY=1 for forward-proxy-only deployments, an optional headers: map on Claude apps gateway upstreams, and a background-task wait indicator — plus ~26 bug fixes. Documentation comparison; no firsthand production run. The 'matters for production' framing in the body is a documentation-based inference pending independent verification.
OpenClaw v2026.7.33 (2026-09-18T05:33:23Z) is the final release in the July 2026 extended-stable branch, moving only the extended-stable* aliases. The tag bundles 126 merged PRs across security and credential safety, message and session integrity, Gateway reliability, channel delivery (UTF-16 truncation fixes across Discord / Matrix / Telegram / Slack / WhatsApp / LINE / Feishu / Zalo / Mattermost / Google Meet / ACP), and provider robustness. Documentation comparison against the GitHub release page and atom feed; no firsthand Gateway run.
OpenAI Codex v0.155.0 stable (2026-09-17 23:16 UTC) ships six capability additions. Headline items: Touch ID verification for MCP requests, real Amazon Bedrock credential rotation, and an experimental /voice with live WebRTC transcripts. Plus TUI reasoning summaries, AppleScript notarization defaults for macOS, and completion-cache eviction fixes. Documentation comparison; no firsthand Codex run.
Claude Code v2.1.275 (2026-09-17 22:33 UTC) ships plugin supply-chain hardening: npm-hosted plugins are now fetched with `npm pack --ignore-scripts` and integrity-verified so postinstall scripts do not run at install. The release also patches two plugin-URL credential leaks (Git-URL password and HTTP basic auth), the same-day v2.1.276 ANTHROPIC_BASE_URL regression fix, and a Claude.ai skills / plugins sync fix. v2.1.275 is a substantive release; v2.1.276 is a single-line regression fix. Documentation comparison.
The OpenAI API changelog entry dated 2026-09-15 (page updated 2026-09-17) ships API key creation governance controls at the organization and project levels. Administrators can allow only service-account keys, allow only user-owned project keys, or disable all new API key creation. Organization restrictions take precedence over project settings; existing API keys are unaffected. Documentation comparison; no admin console was exercised and no creation mode was set.
The Claude Platform release notes entry dated 2026-09-18 (page updated 2026-09-17) extends the Compliance API local session endpoints with a new product_surface value, claude_in_chrome, that returns transcripts of Claude in Chrome sessions. Beta for Claude Enterprise organizations; uses the existing Compliance Access Key and the read:compliance_user_data scope. The article was narrowed after a 2026-09-19 source audit removed earlier claims of changed response shape, new permission grants beyond the documented scope, default enablement for non-Enterprise organizations, and pricing changes. Documentation comparison.
Claude Code v2.1.274 (2026-09-17 00:12 UTC) ships an MCP-and-Gateway release Anthropic shipped quietly overnight. Headline fix: Streamable HTTP MCP tool calls were timing out at a hard 5-minute limit that ignored the per-server setting, now honored via per-request timeouts. Plus mcp startup wait behavior changes, an MCP-503 disconnect fix, a self-hosted Gateway SIGTERM drain fix, and a Plugin MCP request-timeout bug. Documentation comparison; no firsthand Gateway run.
Claude Code v2.1.273 (2026-09-15 20:23 UTC) is the substantive release Anthropic skipped the named notes on for v2.1.272. Headline additions: five x-claude-code-* request headers wired for LLM gateways, MCP disconnect notifications streamed as JSON, and a reverted claude plugin eval deny rule that was over-broad in v2.1.272. Plus OAuth callback reliability and ~25 smaller fixes. Documentation comparison; no firsthand Gateway run.
Claude Code v2.1.271 (2026-09-14 22:12 UTC) ships eight capabilities and ~40 fixes. Headline capability: fast mode now applies in Claude Code Remote sessions (cloud and self-hosted runners), gated by org policy. Headline security: per-command allowed_domains for sandboxed Bash / PowerShell / Monitor in auto mode, and managed-mcp.json now keeps exclusive MCP control when unparseable (fail-closed). Headline safety net: --host-config-snapshot disk|memory prevents self-hosted runners from silently losing host config when the directory exceeds 64 MiB. Documentation comparison.