curl -LsSf https://raw.githubusercontent.com/NVIDIA/OpenShell/main/install.sh | sh
openshell sandbox create --name demo
openshell exec --sandbox demo -- claude --allowedTools Read,Write,Bash
claude-opus-4-7 → claude-opus-5, gemini-3-5-flash or gemini-3-6-flash → gemini-3-8-flash, kimi-k2-7-code → kimi-k3. If you're on the Copilot app or CLI without explicit model selection, GitHub will fall back automatically, but you may see different output behavior — validate your common workflows against the new model. A second deprecation wave hits October 19: GPT-5.5, GPT-5.4, GPT-5.4 mini, GPT-5 mini, Gemini 3.7 Flash, and Grok 4.5 are also being removed on that date — audit your Copilot API integrations for those IDs now.claude-opus-4-7; you maintain CI/CD jobs using Copilot CLI with a specific model flag.EffortQuick — swap model IDs; test key workflows against the replacement.No new model releases today. (Claude Haiku 5.5 and GPT-6.1 Sol Ultrafast remain pending — see Worth Watching.)
transport object on Tunnel responses and one-time relay token in create response; Organization API endpoints move from beta to GA. claude-sonnet-4-5 formally deprecated in SDK (matching the Python SDK v1.11.0 update from the same day).create response now returns a one-time relay token and a read-only transport object — update any code that reads the create response shape. (2) If you manage Claude Enterprise deployments: you can now pull per-user usage/cost from the Admin API at /v1/admin/analytics/usage — useful for team chargeback or quota alerting without screen-scraping the console. (3) If you call Organization API endpoints that were in beta: they're now GA — remove any beta headers you added for those routes. The Plugins and Plugin Marketplaces admin endpoints allow installing/removing/configuring org-wide plugins programmatically; useful if you're rolling out Claude Code plugins across a team. Python SDK v1.11.0 (same date) added only the spend limits endpoint — TypeScript gets the full suite.create response shape changed (Moderate if you parse it).arXiv.org was egress-blocked this scan cycle; full cs.AI/cs.CL new submissions for October 1, 2026 not directly read. See Near-misses.
n:<text> session name/task filter to the agents view, prompt_text field to OpenTelemetry user_prompt events, URL prompts from MCP servers on the 2025-11-25 protocol version, a Windows startup warning when Bash tool access is denied (which disables PowerShell), and built-in gh api (REST only) for Anthropic-managed git in self-hosted runner mode. Multiple fixes: Fast mode now stays enabled in remote sessions, Remote Control message delivery delay resolved, model switching no longer breaks earlier MCP tool announcements in the same session, various screen reader mode improvements, file upload and image handling fixes./plugin enable cc-plugin-you-should-know@builtin. This is also the first release to support URL prompts from MCP servers on the 2025-11-25 protocol — if your MCP server uses the newer protocol version and returns url prompt types, Claude Code now follows those. The prompt_text addition to OpenTelemetry events is useful if you have observability tooling on Claude Code sessions: you can now log the actual text of user prompts in your trace data. Self-hosted runner users: the bundled gh api support for Anthropic-managed git means scripts that call gh api REST endpoints no longer need the gh CLI pre-installed separately in the runner./plugin enable cc-plugin-you-should-know@builtin; review your MCP server protocol version if you see new prompt-type handling./computer on; use /computer show to check status, /computer off to disable. In the Copilot app: Settings → Computer Use → Enable Computer Use. On macOS, Copilot will prompt you to grant Accessibility and Screen Recording permissions on first use — these are required for it to see and interact with apps. Organization-managed Copilot accounts can disable the feature via policy. Copilot asks for approval before taking control of an application; you can set "always allow" for apps you use frequently. Best prompt pattern: describe the outcome you want, the app involved, and any constraints ("summarize my unread Jira tickets in Firefox, ignoring anything from before September"). Important: this is public preview in the standalone Copilot CLI and app — the VS Code Copilot extension does not yet have this capability./computer on or a settings toggle; grant macOS permissions on first run. No code changes needed.web_fetch tool is vulnerable to DoS when given attacker-controlled HTML with deeply nested elements that consume excessive CPU and memory during conversion; fixed in v2.52.0 (also backported to v1.107.7). Four new sandbox backend types added: SpritesSandbox (Fly.io), E2BSandbox, SSHWorkspace, BubblewrapSandbox. Built-in CLAI2 plugins for GitHub, Pylon, Google Workspace, Day AI, Ordinal, Notion, Slack, and Logfire now ship with the framework. pydantic-ai-harness ships alongside every release (previously separate). Model support expanded to include Claude Sonnet 5.5 and OpenAI gpt-6.1-sol. Default max_tokens for Claude Sonnet 4.5+ models bumped from 4096 to 16384.max_tokens from 4096 to 16384 — breaking behavior for code relying on the 4096 default.web_fetch and accepts URLs from user input or external sources — update to v2.52.0 (or v1.107.7 if on the v1 track) immediately. The attack vector is any HTML page the agent fetches with deeply nested DOM elements; it doesn't require a malicious prompt injection, just a crafted URL. (2) max_tokens change: If you have code that relied on the implicit 4096 default for Claude Sonnet 4.5 or later (and built prompt logic around that limit), the new default is 16384. Review any length-sensitive output handling in your agents — this could silently produce longer-than-expected outputs. (3) New sandbox backends: E2BSandbox is now a first-class named type rather than requiring custom config. BubblewrapSandbox is a Linux-native isolation option without requiring Docker or a cloud sandbox service. If you're building agents that execute code locally, BubblewrapSandbox is now worth evaluating as a lightweight isolation layer.web_fetch in pydantic-ai agents with user-provided or externally-sourced URLs; you have Claude Sonnet 4.5+ integrations with pydantic-ai that depend on the 4096 max_tokens default; you're building pydantic-ai agents that need sandbox code execution.EffortModerate — security fix is a version bump (Quick), but the max_tokens change may affect output handling logic (review required); sandbox integration changes are additive.openshell exec -- <command> pattern works with any agent binary. Policy files control what file paths, network destinations, and syscalls are permitted. Pairs naturally with Claude Code's --allowedTools flag — OpenShell enforces the OS-level sandbox while Claude Code enforces the tool-call sandbox; together they cover both application-level and kernel-level limits.Nothing cleared the quality bar this period. arXiv.org, HuggingFace, and TLDR egress-blocked; HN threads not yet at score >200 for today's items.
Anthropic Python SDK v1.11.0 (Sep 30) — Adds list spend limits endpoint; claude-sonnet-4-5 formally deprecated. github.com/anthropics/anthropic-sdk-python/releases/tag/v1.11.0
E2B v2.52.0 (Oct 1) — Sandbox fork count now capped 1–20 (client validation); .dockerignore patterns applied to file uploads; HTTP version pinning for /1.1 on control-plane requests; improved exponential-backoff retry (100ms–10s). github.com/e2b-dev/E2B/releases
Cline v4.1.22 + SDK 0.0.89 (Sep 30) — New providers Bee and Pareto Inference; Anthropic requests now use server-side refusal fallback; default models updated to GPT-6.1 Sol and Claude Sonnet 5.5 across affected providers; oversized tool results recovered via in-session memory cache; reasoning token double-count bug fixed. github.com/cline/cline/releases/tag/v4.1.22
OpenAI Codex rust-v0.160.0 (Oct 1) — Agent command center gains keyboard-accessible pagination for older tasks; X11 terminal middle-click paste; Guardian review now retrieves earlier user instructions and handoff context; sessions can start outside a project using workspace defaults; Windows sandbox ACL auto-repair; plugin manifest caching. github.com/openai/codex/releases/tag/rust-v0.160.0
GPT-6.1 Sol now GA on Amazon Bedrock (Oct 1) — Available at $2/MTok in / $10/MTok out, same pricing as the direct OpenAI API. Grok 4.7 added to Bedrock with US Geo and Global cross-region inference options. aws.amazon.com/blogs/machine-learning
Gemini CLI nightly 20261001 (Oct 1) — Fixes CPU hang and quote handling in code blocks; adds serialization and atomic writes for file tool operations. github.com/google-gemini/gemini-cli/releases
Announced at DevDay (Sep 29) but not in the API as of October 1. Up to 8× faster than standard GPT-6.1 Sol in Codex, 300 tokens/sec, at 6× the standard rate. Watch platform.openai.com/docs/changelog.
Limited preview only as of October 1; ~150ms latency, ~10× faster than GPT-6 Luna standard. No public endpoint, model ID, or price published. [openai.com/index/devday-2026-recap/]
No model card, pricing, or release date as of October 1. [Anthropic Opus 5.5 announcement, September 22, 2026]
Free billing period ends October 4. Standard pricing applies from October 5. [platform.openai.com/docs/changelog]
Migrate to antigravity-preview-09-2026; rename snake_case built-in tool parameters to PascalCase. [ai.google.dev/gemini-api/docs/changelog]
GPT-5.5, GPT-5.4, GPT-5.4 mini, GPT-5 mini, Gemini 3.7 Flash, and Grok 4.5 removed from all Copilot experiences. Audit your Copilot API integrations now. [github.blog/changelog/2026-09-18-upcoming-deprecation-of-selected-github-copilot-models-in-mid-october/]
Filtered from 30+ primary sources against a published quality rubric. No press releases, no fluff — only what changes what you build.