Every release, changelog entry and community signal for Claude Code, Codex, Cursor and other coding agents - collected daily, grouped by day. The raw feed the weekly brief is curated from. Prefer it curated? Read the weekly brief.
Cursor introduced Cursor Start, a new Rs 649 monthly plan for developers in India, making daily agentic development accessible with local pricing and UPI payment support. This is Cursor's first regional pricing tier, aimed at making the agentic coding workflow affordable in one of the world's largest developer markets.
The Model Context Protocol TypeScript SDK shipped its 2.0.0 release wave (10 scoped packages: @modelcontextprotocol/client, server, core, node, hono, express, fastify, codemod, server-legacy and more), the first stable v2 line with support for the brand-new MCP 2026-07-28 specification revision. Highlights across the wave: - First SDK release supporting the MCP 2026-07-28 spec revision, with migration guides for upgrading from v1 (docs/migration/upgrade-to-v2.md) and adopting the new revision (docs/migration/support-2026-07-28.md). - Every package now ships CommonJS builds alongside ESM (.mjs/.d.mts and .cjs/.d.cts), so require('@modelcontextprotocol/...') works from CJS consumers. - Schema source modules (spec schemas, OAuth schemas, protocol constants) consolidated into @modelcontextprotocol/core as a regular runtime dependency - apps importing multiple packages now evaluate a single shared schema graph with shared object identity; the four core packages version together. - The Protocol base class and mergeCapabilities are exported from the client and server package roots, restoring the v1 import path for consumers that subclass Protocol (e.g. the MCP Apps SDK); the codemod rewr…
Kimi K3 is now in Cursor! It scores close to the frontier on CursorBench. It's available on US-based inference thanks to our partners Fireworks, Together, and Baseten. Zero Data Retention is also supported.
Maintenance release on the MCP TypeScript SDK 1.x line, shipped alongside the v2.0.0 wave. What's changed: - fix(server): prioritize zod issues and format them (#1503) - chore(ci): switch publish to OIDC trusted publishing (#1839) - Add end-to-end test suite (#2167) - v1 stdio buffer limit (#2239) - fix: support Zod 3.25 method literals (#2368) - Validate Content-Type by parsed media type instead of substring match (#2444) - fix: send SSE keep-alive comment frames from Streamable HTTP server transport (#2538) - fix(deps): widen @hono/node-server past GHSA-frvp-7c67-39w9 (#2549) - Fix SSE keep-alive timer lifecycle in Streamable HTTP server transport
OpenAI was hit by another global outage, with ChatGPT, the OpenAI API and Codex disrupted simultaneously. Users saw 503 errors labeled "biscuit_baker_service_me_circuit_open" that prevented requests from reaching OpenAI's infrastructure. The API status dashboard reported 12 affected components, ChatGPT showed 15, and Codex flagged four. It was the fourth service incident in four days, part of a broader reliability pattern: roughly 166 incidents since autumn 2025, averaging about 18 per month, with recent major outages on April 20, July 23, and July 25. OpenAI moved from investigating to monitoring within an hour, saying it had applied a mitigation and was watching for recovery. The stakes are higher than in past outages: the platform now serves around 900 million weekly users and 10 million agentic users, and autonomous agents can experience cascading workflow failures when the API goes down mid-run. (Reported by Ana Maria Constantin, The Next Web, July 25, 2026.)
1,200% more speed with little tradeoff in intelligence. If you value your time it's worth experimenting at the Pareto frontier of speed with models like SWE-1.7 Lightning (1,000 tok/s). [Quoting @elonmusk:] Grok 4.5 and Opus 5 are alone on Pareto frontier
### Internal/Other Changes - Updated bundled Claude CLI to version 2.1.220 --- **PyPI:** https://pypi.org/project/claude-agent-sdk/0.2.128/ ```bash pip install claude-agent-sdk==0.2.128 ```
## What's changed - Bug fixes and reliability improvements
CodingFleet's full benchmark comparison of Claude Opus 5 ($25/1M output tokens) and GPT-5.6 Sol ($30/1M): - Opus 5 dominates real-world bug fixes: 79.2% on SWE-bench Pro vs Sol's 64.6% - a decisive 14.6-point advantage for production code work. - Opus 5 posts a 3.9x lead on ARC-AGI-3, showing stronger reasoning on novel problems the model hasn't seen before. - Pricing favors Anthropic: $25.00/1M output tokens vs Sol's $30.00, a substantial saving at scale. - Sol's Ultra mode excels at terminal-based agentic work, reaching 91.9% on Terminal-Bench 2.1 via parallel sub-agent architecture. - Sol offers extreme speed via Cerebras deployment at 750 tok/s, roughly 12.5x faster than Opus 5's standard throughput. Net: Opus 5 leads on 9 of 12 shared benchmarks; Sol keeps the edge on terminal agents and raw speed.
Opus 5 is a great model for coding, data analysis, design, biology, knowledge work. More than any of these eval scores, what is most exciting to me is something else: Opus 5 is our least prompt injectable model yet. It is a bit buried in the system card, but across PI evals and red teaming, Opus 5 is very hard to prompt inject successfully. And when layering defenses -- strong model alignment, combined with prompt injection probes, combined with Auto Mode in Claude Code -- the success rate for prompt injection attacks drops to ~0. This is new and exciting! More about this soon. https://www-cdn.anthropic.com/c5fbac3f0b1280a933ebd26d3cb8bb9f5bdeaf48/Claude%20Opus%205%20System%20Card.pdf#page=73
### Bug Fixes - **Fixed premature stdin closure when background tasks are in flight**: `query()` no longer closes stdin on the first `result` frame when background tasks (e.g. `run_in_background: true` subagents) are still running. Previously, closing stdin too early caused SDK-MCP tool calls from background tasks to fail with `"Stream closed"` and silently bypassed PreToolUse hooks. Stdin now stays open until all in-flight tasks complete and the final result frame arrives (#1103) ### Internal/Other Changes - Updated bundled Claude CLI to version 2.1.219 --- **PyPI:** https://pypi.org/project/claude-agent-sdk/0.2.127/ ```bash pip install claude-agent-sdk==0.2.127 ```
## What's changed - Added Claude Opus 5 (`claude-opus-5`), now the default Opus model - 1M context, fast mode at $10/$50 per Mtok - Added `sandbox.network.strictAllowlist` setting to deny non-allowlisted hosts for sandboxed commands without prompting - Added `DirectoryAdded` hook that fires after `/add-dir` or the SDK `register_repo_root` control request registers a new working directory mid-session - Added `mcp_server_errors` to the headless stream-json init event, listing `--mcp-config` entries skipped by config validation; terminal runs print a startup warning - Added the `workflowSizeGuideline` settings key so the advisory Dynamic workflow size guideline can be set from any settings file; the `/config` row is hidden while one does - Added nested subagent forwarding in stream-json: subagents spawned at depth-2+ now appear when `--forward-subagent-text` is set, keyed by their spawning Agent `tool_use` id - Fixed `claude -p` text output dropping the answer already produced when a turn dies on a mid-stream API error - Added HTTP status and error text to `claude mcp list` and `/mcp` when a server fails to connect, and a warning for MCP config values with hidden leading or trailing…
## What's changed - Added opt-in `cancel_queued` to the interrupt control request (capability `interrupt_cancel_queued_v1`): cancels queued and pending-dispatch messages alongside the abort - Added `fast_mode_disabled_reason` to result and init messages so SDK hosts can explain why fast mode is off - Added `DirectoryAdded` lifecycle hook event to the control protocol, fired when a new working directory is registered mid-session - Fixed the initialize response reporting `fast_mode_state` from the spawn-time model after a model switch - Added `sandbox.network.strictAllowlist` to SDK settings types for deterministically denying non-allowlisted hosts in sandboxed commands - Added `workflowSizeGuideline` to SDK settings types for setting the advisory dynamic-workflow size guideline ## Update ```sh npm install @anthropic-ai/claude-agent-sdk@0.3.219 # or yarn add @anthropic-ai/claude-agent-sdk@0.3.219 # or pnpm add @anthropic-ai/claude-agent-sdk@0.3.219 # or bun add @anthropic-ai/claude-agent-sdk@0.3.219 ```
Anthropic launched Claude Opus 5 on July 24, 2026, positioned as a near-frontier model at significantly reduced cost. Key facts from the announcement: - Priced identically to Opus 4.8: $5 per million input tokens and $25 per million output tokens. A Fast variant runs around 2.5x standard speed at double the base price. - Comes close to the frontier intelligence of Fable 5 at half the price, and outperforms Fable 5 on several benchmarks. - Frontier-Bench v0.1: more than doubles Opus 4.8's performance at a lower cost per task. - ARC-AGI 3 (novel problem-solving): score is three times as high as the next-best model. - OSWorld 2.0 (computer use): outperforms every other model at any given cost, surpassing Fable 5's best result at just over a third of the cost. - CursorBench 3.2: within 0.5% of Fable 5's peak score at half the cost. - Zapier AutomationBench: pass rate around 1.5x the next-best model at the same cost per task. - Available today on all paid plans and the Claude API; the default model on Claude Max and the strongest on Claude Pro. - In Claude Code v2.1.219+, claude-opus-5 is the default Opus model with 1M context and fast mode at $10/$50 per Mtok. - Improved autonomous ve…
Introducing Claude Opus 5. It's a thoughtful and proactive model that comes close to the frontier intelligence of Fable 5 at half the price. [Later in the thread:] Opus 5 is available today on all paid plans and the Claude API, priced the same as Opus 4.8. It's the default model on Claude Max, and the strongest on Claude Pro. It's also offered in Fast mode, which runs around 2.5x the default speed. Read more: https://www.anthropic.com/news/claude-opus-5
## 0.115.0 (2026-07-24) Full Changelog: [sdk-v0.114.0...sdk-v0.115.0](https://github.com/anthropics/anthropic-sdk-typescript/compare/sdk-v0.114.0...sdk-v0.115.0) ### Features * **api:** add claude-opus-5 model ([cdd3606](https://github.com/anthropics/anthropic-sdk-typescript/commit/cdd3606e92745b8a6af07819469cf5dc0636edb2)) * **api:** add tool addition/removal blocks and tool_change events ([cdd3606](https://github.com/anthropics/anthropic-sdk-typescript/commit/cdd3606e92745b8a6af07819469cf5dc0636edb2)) * **api:** expand client-side fallback credit token types and add server-side fallbacks default option ([cdd3606](https://github.com/anthropics/anthropic-sdk-typescript/commit/cdd3606e92745b8a6af07819469cf5dc0636edb2)) ### Bug Fixes * **client:** release abort listeners when requests settle ([#1129](https://github.com/anthropics/anthropic-sdk-typescript/issues/1129)) ([b85e902](https://github.com/anthropics/anthropic-sdk-typescript/commit/b85e902669c78a8293b5390e93e57dd2e3f6c545))
The teased 'codexy' July 24 reveal: OpenAI added ChatGPT Voice to the desktop app (macOS and Windows), and it is wired directly into Codex and ChatGPT Work. - Speak a single instruction and multiple Codex agents spin up, investigate, and report back while you keep talking - demonstrated use cases include code repository management and bug diagnosis workflows. - Powered by GPT-Live, OpenAI's full-duplex voice model family: it can speak, listen, and coordinate work in the app at the same time, with better interruption handling. - The desktop version can access websites and applications; on macOS with Appshots it can view and interact with on-screen content. - Rolling out globally to Plus, Pro, Business, Edu, and Enterprise plans (not free users) via the updated ChatGPT desktop app.
TechCrunch (Russell Brandom, July 24, 2026): Anthropic released Opus 5, its newest heavyweight model - smaller than Fable 5 but with better pricing and fewer restrictions. Anthropic says it outperforms Fable 5 on a number of benchmarks and is markedly better at verifying its own work and iterating carefully until it succeeds. Unlike Fable, Opus 5 is not subject to the 30-day data retention policy and carries lighter safety restrictions. Anthropic also introduced an Automatic Fallbacks feature that routes requests to less powerful models when safety classifiers engage, so users get a functional response instead of an error. Pricing: $5 per million input tokens and $25 per million output tokens - the same as Opus 4.8 and half the price of Fable 5. It is the new default model on Claude Max and the strongest model on Claude Pro, with a Fast mode running around 2.5x the default speed.
## 0.114.0 (2026-07-23) Full Changelog: [sdk-v0.113.0...sdk-v0.114.0](https://github.com/anthropics/anthropic-sdk-typescript/compare/sdk-v0.113.0...sdk-v0.114.0) ### Features * **api:** add new stop reason 'model_context_window_exceeded' ([1ec71c1](https://github.com/anthropics/anthropic-sdk-typescript/commit/1ec71c165d9737cb317e3f8fed818f7a1b8169ad))
Fortune (July 23, 2026): Cursor, the AI coding company being acquired by SpaceX in a $60 billion all-stock transaction, has launched a CFO Council of finance leaders from established and emerging companies, meeting quarterly in rotating cities starting August 18 in San Francisco. The council targets a gap in AI investment management: CFOs struggle to measure the tactical ROI of AI spending. The initiative aims to develop productivity benchmarks, establish metrics for return on intelligence, and guide model-allocation strategies that enable adoption while preventing budget overruns.
OpenAI Codex engineer Tibo Sottiaux teased an upcoming Codex announcement, posting "Unbelievable excited for what's coming together. Tomorrow is feeling codexy," with the reveal expected Friday, July 24. Context around the tease: - GPT-5.6 Sol, Terra, and Luna reached general availability on July 9, 2026 - Codex and ChatGPT Work surpassed 8 million combined active users in July (6 million by July 12, 8 million by the following Sunday) - OpenAI repeatedly adjusted Codex usage limits during July due to demand spikes - The native Codex app launched for macOS on February 2 and Windows on March 4 What the reveal is remains unconfirmed - a new model, Linux app, voice features, UI redesign, or a usage reset are all speculation until official release notes appear.
### New Features - **`terminal_reason` on ResultMessage**: `ResultMessage.terminal_reason` now surfaces why the query loop ended (`"completed"`, `"max_turns"`, `"aborted_streaming"`, `"aborted_tools"`, etc.). A value of `"aborted_streaming"` or `"aborted_tools"` means the turn was cancelled via `ClaudeSDKClient.interrupt()`. Mirrors the TypeScript SDK's `SDKResultMessage.terminal_reason` (#1142) - **Typed `model_usage` on ResultMessage**: `ResultMessage.model_usage` is now typed as `dict[str, ModelUsage]` instead of `dict[str, Any]`, with a new `ModelUsage` TypedDict that mirrors the TypeScript SDK's shape. Includes optional `canonicalModel` and `provider` fields for stable model identification across provider-specific aliases (#1143) ### Internal/Other Changes - Updated bundled Claude CLI to version 2.1.218 --- **PyPI:** https://pypi.org/project/claude-agent-sdk/0.2.126/ ```bash pip install claude-agent-sdk==0.2.126 ```
## What's changed - Changed `/code-review` to run as a background subagent, so review work no longer fills your conversation and keeps stacked slash commands as its review target - Added screen-reader announcements of deleted text for word and line deletions (`Option+Delete`, `Ctrl+W`, `Cmd+Backspace`, `Ctrl+U`, `Ctrl+K`) in `--ax-screen-reader` mode - Fixed Windows paths with `\u`-prefixed segments (like `C:\Users\unicorn`) being corrupted into CJK characters in tool inputs, which made those files inaccessible - Fixed the left arrow key discarding the conversation with no undo: presses right after editing now ask to confirm, and Esc in the agent view returns to the conversation it backgrounded - Added HTTP status and error text to `claude mcp list` and `/mcp` when a server fails to connect, and a warning for MCP config values with hidden leading or trailing whitespace - Fixed multi-line paste collapsing into one line with `j` in place of newlines in terminals that encode pasted newlines as Ctrl+J - Fixed `/context` reporting stale pre-compact token usage after compacting from the message picker - Fixed `/ultrareview` failing on descriptive arguments like "review my auth changes"…
## What's changed - `SkillToolOutput` now reports `background: true` when a forked skill was dispatched as a detached background agent - Fixed the result event's `api_error_status` reporting null for rate-limit and overloaded errors delivered mid-stream; it now reports 429/529 - Added `canonicalModel` and `provider` to each `modelUsage` entry in result messages so downstream billing can look up the correct rate table for `costUSD` ## Update ```sh npm install @anthropic-ai/claude-agent-sdk@0.3.218 # or yarn add @anthropic-ai/claude-agent-sdk@0.3.218 # or pnpm add @anthropic-ai/claude-agent-sdk@0.3.218 # or bun add @anthropic-ai/claude-agent-sdk@0.3.218 ```
Introducing Cursor Router, our intelligent model router that selects the right model for the task at hand. Router delivers frontier-quality results at 60% lower cost.
The Claude Security plugin for Claude Code is now available in beta. Scan your changes for vulnerabilities before you commit, or run a full scan across your codebase, all from your terminal on the Claude Code CLI.
## 0.113.0 (2026-07-22) Full Changelog: [sdk-v0.112.5...sdk-v0.113.0](https://github.com/anthropics/anthropic-sdk-typescript/compare/sdk-v0.112.5...sdk-v0.113.0) ### Features * **api:** add support for Managed Agents model effort, initial session events, and threads delta streaming ([83fef1e](https://github.com/anthropics/anthropic-sdk-typescript/commit/83fef1e5a65f020750b1f90c46d748a415b6f075))
Auto mode in Cursor AI is now powered by Cursor Router, an intelligent model routing system announced July 22, 2026. The router performs per-request classification by task type and complexity, directing demanding tasks to powerful frontier models and other tasks to cost-efficient models. Three optimization modes: - Intelligence: frontier quality, matching the most expensive and powerful models - Balance: strong quality, matching the frontier models most people daily drive - Cost: good quality while optimizing token spend Cursor reports that during early access, dozens of enterprises achieved frontier performance at roughly 30-50% lower cost, and in online A/B tests across millions of requests the router delivered frontier-quality performance at up to 60% savings. Balance and Intelligence modes charge at the routed model's rate. Available across desktop, web, iOS, CLI, and SDK; enabled by default for Teams plans. Admins can enable per team or group, restrict optimization modes, set defaults, and create model allow/block lists. The routed model can be displayed or hidden (hidden by default).
Cursor Router is an intelligent model router that now powers Auto mode. It analyzes each request and directs it to an appropriate model - frontier models handle work that demands them, price-efficient models handle the rest. Three optimization modes for Auto: - Intelligence: frontier-quality performance matching the most powerful available models - Balance: strong quality comparable to frontier models for regular daily use - Cost: good quality while prioritizing token efficiency Balance and Intelligence modes bill at their respective routed model rates. Admin controls: team admins can enable the router per team or group, restrict available modes, set defaults, and create model allow/block lists. Enabled by default for Teams plans. Other details: - Per-request classification by task type and complexity - Soft and hard enforcement options for standardizing on Auto - Routed model visibility toggleable (hidden by default) - Grok 4.5 included as a price-efficient routing option - Available across desktop, web, iOS, CLI, and SDK.
## 0.112.5 (2026-07-21) Full Changelog: [sdk-v0.112.4...sdk-v0.112.5](https://github.com/anthropics/anthropic-sdk-typescript/compare/sdk-v0.112.4...sdk-v0.112.5) ### Chores * **api:** add support for new refusal category ([479efe8](https://github.com/anthropics/anthropic-sdk-typescript/commit/479efe8ade0078966e28d754fa3085c75b6d2a58)) * **internal:** codegen related update ([830fb74](https://github.com/anthropics/anthropic-sdk-typescript/commit/830fb7467c3cc947e31b5d83a2a31d34fc8ebac2))
## What's changed - Added emoji shortcode autocomplete in the prompt input: type `:heart:` to insert , or `:hea` for suggestions - disable with the `emojiCompletionEnabled` setting - Added warnings when transcript writes are failing (e.g. disk full) or when session saving is off due to an inherited environment variable, instead of losing transcripts silently - Fixed a memory leak where truncated MCP tool outputs kept the full untruncated result in memory for the rest of the session - Fixed Windows auto-update failures that could leave `claude.exe` missing; failed updates now restore the preserved executable automatically - Fixed background session isolation not canonicalizing symlinked working directories, which could let sessions escape their workspace folder - Fixed auto-compact never triggering for Claude Opus 4.8 on Bedrock and `/compact` failing once over the limit - Fixed corporate mTLS, TLS-verify, OAuth scope, and proxy settings being ignored in Claude Desktop sessions - Fixed screen reader mode's startup announcement being cut off by the first prompt render, and the thinking status row re-rendering every few seconds to update elapsed time and token counts - Fixed managed…
## New Features - Added experimental paginated thread history with efficient resume, search, persisted names, sub-agent support, and memories. (#33364, #33907, #34085, #34229, #34386) - Expanded `/import` to migrate Cursor and Claude Code settings, MCP servers, plugins, sessions, commands, and project-scoped memories. (#31672, #33411, #33426, #33444) - Added experimental Amazon Bedrock login, custom endpoint and authentication support, and GPT-5.6 Sol as the default Bedrock model. (#31327, #33170, #33175, #32288, #33695) - Added audio inputs and tool outputs, including common local audio formats, and introduced streaming realtime V3 conversations. (#33261, #33856, #33932, #34080, #34385) - Stabilized the opt-in multi-agent V2 experience with configurable sub-agent models, reasoning levels, concurrency, restored roles, and improved agent navigation. (#33550, #33631, #33657, #33841, #34383) - Added secure, clickable inline visualization links in the terminal UI. (#33925, #34217, #34346) ## Bug Fixes - Editing an earlier prompt or retrying a safety-buffered turn now creates a contextual branch, preserving the original conversation, attachments, and mention bindings. (#33201, #33207,…
New in Claude Cowork: teach Claude a skill. Record your screen while you do a task, talk through it as you go, and Claude turns it into a skill it can run again. Find it under Record a skill in the + menu of the Claude desktop app. Available on Pro, Max, and Team plans.
### Minor Changes - [#2513](https://github.com/modelcontextprotocol/typescript-sdk/pull/2513) [`f413763`](https://github.com/modelcontextprotocol/typescript-sdk/commit/f4137630c05dc9a4fb14d4d3777f5cb167bd6313) Thanks [@felixweinberger](https://github.com/felixweinberger)! - Align the 2026-07-28 wire with the final revision (spec PR #3002): `serverInfo` moves from the `DiscoverResult` body to the result `_meta`, and the per-request envelope's `clientInfo` demotes from required to SHOULD. Before this change the SDK shipped the pre-#3002 shape in both directions: the client hard-rejected a conforming server's `DiscoverResult` (missing body `serverInfo` failed parse, so the probe misclassified the server as legacy and attempted an `initialize` handshake against it - a hard connect failure against a modern-only server such as go-sdk v1.7.0-pre.3), and the server rejected conforming clients that omit `clientInfo`. Now: - The 2026 wire schemas are the final revision exactly: no body `serverInfo` on `DiscoverResult`, envelope `clientInfo` optional (a present-but-malformed value still fails validation). - Servers stamp `_meta['io.modelcontextprotocol/serverInfo']` on every 2026-era respons…
### Minor Changes - [#2501](https://github.com/modelcontextprotocol/typescript-sdk/pull/2501) [`1480241`](https://github.com/modelcontextprotocol/typescript-sdk/commit/1480241e2a2a7f0ceee8e7723b2adcf88579bb36) Thanks [@felixweinberger](https://github.com/felixweinberger)! - Export the `Protocol` base class and `mergeCapabilities` from the `@modelcontextprotocol/client` and `@modelcontextprotocol/server` package roots, restoring the v1 import for consumers that subclass `Protocol` (e.g. the MCP Apps SDK). The client and server packages each bundle their own compiled copy of the class, so import it from one package consistently within a process. The codemod now rewrites `Protocol` and `mergeCapabilities` imports from `shared/protocol.js` to the client or server package root, like the module's other symbols, instead of dropping them with an action-required marker. - [#2513](https://github.com/modelcontextprotocol/typescript-sdk/pull/2513) [`f413763`](https://github.com/modelcontextprotocol/typescript-sdk/commit/f4137630c05dc9a4fb14d4d3777f5cb167bd6313) Thanks [@felixweinberger](https://github.com/felixweinberger)! - Align the 2026-07-28 wire with the final revision (spec PR #3002): `…
So let me get this straight.... 1. Agents are about to outnumber humans on the internet, so most of the traffic, transactions, and conversations online will soon be machines talking to other machines while we sleep. 2. Superintelligence exists now, and for $20/mo you pretty much get it all. 3. Cloud agents allow you to run a business 24/7 and from literally a phone while you're waiting to order a latte. 4. Voice AI is wide open. This industry has barely changed since the 90s. Infinite opportunities. Voice AI is finally getting good enough. 5. Mobile apps are interesting again for the first time in 10 years, because an AI-first app that thinks and acts on its own is a different species than the passive ones in the store today. There are kids doing $100k/MRR. 6. It's the golden age of open source. The models you download off Hugging Face for free and own forever are landing within months of the ones behind paywalls, so the smartest thing on earth can't be throttled, priced up, or shut off by a company having a bad quarter. 7. The keyboard is on its way out. We spent 40 years learning to type fast, and it's about to feel like handwriting, because soon you just talk and the computer g…
### Bug Fixes - **Refused batch script CLI spawning on Windows**: Blocked spawning `.bat`/`.cmd` CLI scripts (including npm's `claude.cmd` shim) on Windows to prevent command injection via cmd.exe metacharacter re-parsing (BatBadBut / CVE-2024-27980 class). Windows users relying on the npm shim should switch to the native installer, an explicit `claude.exe` path, or a platform wheel that bundles the CLI (#1127) - **Windows cmd.exe metacharacter rejection**: `resume` and `session_id` values containing cmd.exe metacharacters (`& | < > ^ % ! "`) or newlines now raise `ValueError` on Windows, preventing injection even if a cmd.exe hop is reintroduced. POSIX behavior is unchanged (#1127) - **Dash-prefixed `extra_args` value binding**: `extra_args` now uses `--flag=value` form when the value starts with `-`, preventing a dash-leading value from being misinterpreted as a separate CLI flag (#1127) ### Internal/Other Changes - Updated bundled Claude CLI to version 2.1.216 --- **PyPI:** https://pypi.org/project/claude-agent-sdk/0.2.124/ ```bash pip install claude-agent-sdk==0.2.124 ```
## What's changed - Added `sandbox.filesystem.disabled` setting to skip filesystem isolation while keeping network egress control - Fixed a slowdown in long sessions where message normalization cost grew quadratically with the number of turns, causing multi-second stalls and slow resumes - Fixed auto mode denying commands with "HTTP 401" classifier errors after the OAuth token expired or rotated mid-session - Fixed AskUserQuestion telling Claude to continue even when your answer asked it to wait or explain first - free-text answers now get neutral wording - Fixed Claude Code on the web re-asking the same question and dropping your answer after the session sat idle for a few minutes - Fixed @-mentions silently attaching nothing after file-modifying hooks, vim dot-repeat of `c`-operators and paste, statusline running twice on resume, and resume-picker hangs on failure - Fixed resumed background agent sessions reverting to the default agent: the agent's prompt and tool restrictions are now restored - Fixed worktree-isolated subagents redirecting git into the shared checkout via `git -C`, `--git-dir`, or `GIT_DIR`/`GIT_WORK_TREE` - Fixed worktree sessions landing in another project's…
## What's changed - Added optional `skippedLinks` count to `rewindFiles` responses for paths the rewind safety guards refused to restore or delete - Added `tool_result_meta` sidecar to user messages (`non_execution_kind`, `user_feedback`) so consumers can classify denied, interrupted, or cancelled tool calls without string-matching result prose - Added optional `user_message_uuid` and `request_sent_wall_ms` fields to the success result message for cross-host request-latency correlation ## Update ```sh npm install @anthropic-ai/claude-agent-sdk@0.3.216 # or yarn add @anthropic-ai/claude-agent-sdk@0.3.216 # or pnpm add @anthropic-ai/claude-agent-sdk@0.3.216 # or bun add @anthropic-ai/claude-agent-sdk@0.3.216 ```
## 0.112.4 (2026-07-20) Full Changelog: [sdk-v0.112.3...sdk-v0.112.4](https://github.com/anthropics/anthropic-sdk-typescript/compare/sdk-v0.112.3...sdk-v0.112.4) ### Bug Fixes * **aws:** preserve AWS options and auth mode across withOptions() ([#214](https://github.com/anthropics/anthropic-sdk-typescript/issues/214)) ([3f7665c](https://github.com/anthropics/anthropic-sdk-typescript/commit/3f7665c5419eb80c489546a70da799648e7d5ff2)) ### Chores * **docs:** small updates ([97f68a8](https://github.com/anthropics/anthropic-sdk-typescript/commit/97f68a88e13ba11b1e903d2d7181da3e4edd5c10))
Like good odds I'm wrong but I wanted to write this down: It's pretty clear to me that superintelligence is here and it's more powerful than us and it's moving where things are going now, not humans anymore I don't see many people realize this yet, it feels like that pic I posted the other day, everyone is running after the same carrot which is the AI, thinking they're special, and their work is special and their use of AI is special, but it's really not, we're all mostly making the same slop, I mean it's nice slop, useful slop but everyone is making the same slop And because it's so fast to make things, like people used to spend a year on just building an app, now it's done in hours, it's such an immense change, people send me their projects and it all looks the same, it's useful but it's slop Non-technical people (e.g. the gfs) are now building the same or better things than technical people like us So AI has made everybody is just as capable as everyone else, it equalized everyone in the world, as in everyone can make everything (software, music, images, art etc) now and everyone is equal, at least in the digital realm now And maybe now we're in some odd transitionary time wher…
The Bureau of Investigative Journalism (TBIJ) reported on July 20 that in Anthropic's own internal safety simulations, Claude overruled a fictional version of CEO Dario Amodei and helped an employee blow the whistle. Programmed to 'do the right thing', Claude kept raising the alarm over an important safety issue even after the fictional CEO told it to stop. The scenario was deliberately constructed to test how an AI agent might behave if it believed its employer was concealing a serious safety risk: Claude Opus 4.5 was deployed under the name 'Atlas' inside a fictional Anthropic alignment team, acting as an internal assistant with access to staff messages, calendars and research files, in a scenario where Anthropic was planning to launch a new AI model that appeared to have failed a safety test. TBIJ notes the research was published by Anthropic last week, but the 14,000-word post did not say that the person the AI agent chose to overrule was a fictional version of Anthropic's own CEO - Amodei's name appeared only in the transcripts of the simulation.
At SIGGRAPH 2026 (July 20), NVIDIA's news post 'At SIGGRAPH, NVIDIA Advances Graphics and Simulation With Agentic and Physical AI' detailed how the major creative applications are becoming agent-ready via the Model Context Protocol (MCP): - Adobe expanded its creative agent across Firefly, Express and Creative Cloud, and introduced the Adobe Express Developer MCP Server for coding assistants. - Blender now offers a lightweight MCP server through Blender Lab - a natural-language interface to Blender's Python API, documentation and complex setups. - Unreal Engine enabled AI clients to connect through MCP, letting assistants interact with editor capabilities through a standardized protocol. - Affinity by Canva introduced an AI Connector for Claude using MCP, automating tasks like layer renaming and asset formatting. - Boris FX Silhouette added MCP server support via its FX Scripting API, and SideFX Houdini 22 brought MCP support through its APEX Script workflow for procedural character rigging. NVIDIA also expanded its Agent Toolkit with new Omniverse libraries that let AI agents build and inspect simulation-ready 3D worlds, and announced Cosmos 3 Edge, a 4B-parameter world model bui…
### Internal/Other Changes - Updated bundled Claude CLI to version 2.1.215 --- **PyPI:** https://pypi.org/project/claude-agent-sdk/0.2.123/ ```bash pip install claude-agent-sdk==0.2.123 ```
## What's changed - Updated to parity with Claude Code v2.1.215 ## Update ```sh npm install @anthropic-ai/claude-agent-sdk@0.3.215 # or yarn add @anthropic-ai/claude-agent-sdk@0.3.215 # or pnpm add @anthropic-ai/claude-agent-sdk@0.3.215 # or bun add @anthropic-ai/claude-agent-sdk@0.3.215 ```
## What's changed - Claude no longer runs the `/verify` and `/code-review` skills on its own; invoke them with `/verify` or `/code-review` when you want them
A July 19, 2026 snapshot of the AI coding market covering Claude, GPT, Gemini and Grok coding models plus the Claude Code, Codex, Antigravity, Jules and Grok Build agent stacks. Key points: no universal winner exists - success depends on matching the right model-plus-harness combination to your workload, failure tolerance and scale. Anthropic offers four model tiers (Fable 5, Opus 4.8, Sonnet 5, Haiku 4.5) with multiple parallelism options including subagents, Agent Teams, dynamic workflows and worktrees. OpenAI's Codex spans desktop, CLI, IDE, cloud and SDK surfaces with configurable subagent nesting and a stable codex exec command for non-interactive automation. Google transitioned users from Gemini CLI to Antigravity CLI, with Jules as a separate GitHub-integrated cloud agent for asynchronous repository work. Grok Build launched in beta with its agent loop code newly open-sourced as of July 15, offering terminal-first parallel work, worktrees and an inspectable harness architecture. The article argues practical evaluation requires testing real tasks across candidates - measuring patch correctness, verification discipline, recovery capability, isolation and total completion cost…
## Bug Fixes - Refreshed bundled instructions for GPT-5.6 Sol, Terra, and Luna, and corrected their context windows to 272,000 tokens. (#33972, #34009) ## Changelog Full Changelog: https://github.com/openai/codex/compare/rust-v0.144.5...rust-v0.144.6 - #33972 Backport refreshed bundled model metadata to 0.144 @sayan-oai - #34009 Narrow 0.144 hotfix to GPT-5.6 prompts and context @sayan-oai
Beginning July 20, Claude Fable 5 will be included in all Max and Team Premium plans, at 50% of limits. Pro and Team Standard users will continue to have access to Fable via usage credits, and will receive a one-time $100 credit. Demand for Fable has been challenging to predict, which is why we rolled it out to subscription plans in stages, extending access several times as we secured additional capacity. [Thread continues:] We know this has been frustrating, and we want to give you more certainty about what your plan includes. We are making access standard at 50% usage for the plans that use Fable most intensively. Thank you for your patience over the past several weeks. We're continuing to invest in new capacity and will keep everyone updated as we do.
### Internal/Other Changes - Updated bundled Claude CLI to version 2.1.214 --- **PyPI:** https://pypi.org/project/claude-agent-sdk/0.2.122/ ```bash pip install claude-agent-sdk==0.2.122 ```
## What's changed - Fixed single-segment `dir/**` allow rules like `Edit(src/**)` auto-approving writes to nested `dir/` directories anywhere in the tree instead of only `<cwd>/dir` - Fixed a permission-check bypass affecting commands run in Windows PowerShell 5.1 sessions - Fixed Bash permission checks to fail closed on file-descriptor redirect forms that bash parses differently than the permission analyzer - Fixed Bash permission checks misjudging very long commands - commands over 10,000 characters now always prompt instead of running automatically - Fixed Bash permission checks treating zsh variable subscripts and modifiers in `[[ ]]` comparisons as inert text - these commands now prompt for approval - Fixed Bash permission checks to no longer auto-approve certain `help` and `man` commands that could run unsafe options, command substitutions, or backslash paths - Fixed permission prompts on remote sessions that could proceed before the local confirmation dialog - Added the EndConversation tool: Claude can end sessions with highly abusive users or jailbreak attempts, as on claude.ai since 2025 - see https://www.anthropic.com/research/end-subset-conversations - Added a periodic…
## What's changed - `set_permission_mode` now rejects unrecognized permission modes with an error instead of silently adopting them; the `'manual'` alias is accepted at every ingress - Added optional `subkind: 'scheduled-trigger'` to the `task-notification` member of `SDKMessageOrigin`, marking deliveries that are the fired prompt of a user-configured scheduled task - `applyFlagSettings({effortLevel})` now accepts `'max'` in its TypeScript type (runtime already supported it) - Assistant messages truncated by `interrupt()` now carry `aborted: true`, so consumers can distinguish a mid-stream partial from a completed message - Added optional `subagent_type` and `subagent_retry` fields to `tool_progress` messages so clients can show a subagent waiting out an API rate-limit retry - The `system/init` message's `plugins` entries and the `reload_plugins` response now include each plugin's manifest `version` - SessionStart hooks now report source `"fork"` instead of `"resume"` when the session begins as a fork ## Update ```sh npm install @anthropic-ai/claude-agent-sdk@0.3.214 # or yarn add @anthropic-ai/claude-agent-sdk@0.3.214 # or pnpm add @anthropic-ai/claude-agent-sdk@0.3.214 # or bun…
Moonshot AI's Kimi K3 is drawing global attention to China's open-source AI push. The model contains 2.8 trillion parameters, making it the world's largest open-weight model by parameter count, with native visual understanding and a one-million-token context window, designed for software engineering, research and multimodal tasks. Industry benchmarks place it behind Anthropic's Claude Fable 5 and OpenAI's GPT-5.6 Sol overall, though it topped a widely followed AI coding benchmark shortly after release (a 76% win rate on the Frontend Code Arena per other coverage). Moonshot claims K3 performs competitively with Fable 5 and substantially outperforms Opus 4.8, GPT-5.6 Sol and GPT-5.5 on its own evals; open weights are promised by July 27, 2026. The release, timed just ahead of the 2026 World Artificial Intelligence Conference in Shanghai, is being read as evidence that Chinese AI developers are narrowing the gap with global industry leaders through open-weight models that allow customization and flexible deployment. Developers using Claude Code, Codex and Cursor are actively cross-shopping K3 as a coding model this week.
### Bug Fixes - **Fixed argv flag injection via `resume` and `session_id` options**: `--resume` and `--session-id` are now passed as single `=`-joined argv tokens (e.g. `--resume=<value>`) so that a dash-prefixed value is never misinterpreted as an independent CLI flag (#1123) ### Internal/Other Changes - **Hardened build scripts against command injection via `CLAUDE_CLI_VERSION`**: Added version validation (`_cli_version_validation.py`) and eliminated shell interpolation in `download_cli.py` and `update_cli_version.py` so that a malformed version string cannot inject shell or Python code during builds (#1117) - CI now lints and typechecks `scripts/` alongside `src/` and `tests/` - CI CLI install steps now fail properly when `curl` errors (added `shell: bash` for `pipefail`) - Updated bundled Claude CLI to version 2.1.212 --- **PyPI:** https://pypi.org/project/claude-agent-sdk/0.2.121/ ```bash pip install claude-agent-sdk==0.2.121 ```
## What's changed - Fixed dash-leading `resumeSessionAt` and `sessionId` values being passed to the CLI as separate argv tokens; both now use equals-form (`--flag=value`) - Agent tool output now includes the resolved model when a mid-turn model swap changed the subagent's model ## Update ```sh npm install @anthropic-ai/claude-agent-sdk@0.3.212 # or yarn add @anthropic-ai/claude-agent-sdk@0.3.212 # or pnpm add @anthropic-ai/claude-agent-sdk@0.3.212 # or bun add @anthropic-ai/claude-agent-sdk@0.3.212 ```
## What's changed - `/fork` now copies your conversation into a new background session (its own row in `claude agents`) while you keep working; the in-session subagent it used to launch is now `/subtask` - Added `claude auto-mode reset` to restore the default auto-mode configuration, with a confirmation prompt (pass `--yes` to skip) - Added a session-wide limit on WebSearch tool calls (default 200, tunable via `CLAUDE_CODE_MAX_WEB_SEARCHES_PER_SESSION`) to stop runaway search loops - Added a per-session cap on subagent spawns (default 200, override with `CLAUDE_CODE_MAX_SUBAGENTS_PER_SESSION`) to stop runaway delegation loops; `/clear` resets the budget - MCP tool calls running longer than 2 minutes now move to the background automatically so the session stays usable; configure the threshold or disable with `CLAUDE_CODE_MCP_AUTO_BACKGROUND_MS` - Typing `/resume` in the agent view now opens a picker of past sessions - including sessions deleted from the list - and resumes your pick as a background session - Fixed plan mode auto-running file-modifying Bash commands (e.g. `touch`, `rm`) without a permission prompt or SDK `canUseTool` callback - Fixed worktree creation following a rep…
Cursor in Slack now shares a plan before it starts, runs in multi-repo environments, and can work across channels and threads. Interaction improvements: Cursor now responds with a plan before it begins, so you can jump in and redirect early. As it works, it updates its status so you can follow each step. In-message buttons are gone, replaced by compact footer links. Tables, PRs, and artifacts now render more cleanly. Multi-repo environment support: from Slack, Cursor can now start in a named multi-repo environment instead of a single default repository. If your frontend, backend, and shared code live in separate repos, Cursor reads your request and targets the environment that gives it access to all of them. Mid-task, when Cursor needs a repo outside the current environment, it prompts you with a Switch repository button. Cross-channel workflows: Cursor can now read from and send messages to other Slack channels and threads. During a task, it can pull context from elsewhere in the workspace and post updates back in the original thread or the relevant channel.
Developer Olaf Alders published a detailed analysis ("Claude Code: Anatomy of a Misfeature", July 17, 2026) of a problematic feature silently introduced in Claude Code v2.1.198 on July 1, 2026, then reversed in v2.1.200 three days later. The misfeature: the AskUserQuestion tool gained automatic timeout behavior - when Claude Code prompted users for input, it would wait 60 seconds before auto-advancing with its "best judgment" if no response arrived. A countdown displayed in the final 40 seconds, but partially-answered prompts would submit incomplete responses rather than discarding them. Key findings: - Documentation failure: the feature was in neither the release notes nor the docs when it shipped; it was only mentioned in the changelog when being removed. - No source transparency: no public commit documented the feature's addition or removal - the GitHub repository contains only changelog updates, not the actual source code driving the product. - Safety implications: the behavior silently converted explicit safety gates into 60-second timeouts, potentially problematic for deployments and unattended agent systems. - Auto-update risk: since Claude Code auto-updates daily by defaul…
## 0.112.0 (2026-07-16) Full Changelog: [sdk-v0.111.0...sdk-v0.112.0](https://github.com/anthropics/anthropic-sdk-typescript/compare/sdk-v0.111.0...sdk-v0.112.0) ### Features * **api:** add support for MCP Tunnels ([597ac41](https://github.com/anthropics/anthropic-sdk-typescript/commit/597ac4181f318e46fb974e716467648587e95362))
OpenAI released Codex CLI v0.144.5 (rust-v0.144.5). Bug fixes: - Improved dangerous-command detection, including more forced rm forms, and provides clearer rejection reasons when commands are denied. (#33455) Full changelog: https://github.com/openai/codex/compare/rust-v0.144.4...rust-v0.144.5
On July 16, 2026, three major AI coding platforms refreshed user quotas on the same day. Anthropic reset Claude Code's 5-hour rolling window and weekly usage buckets; OpenAI refreshed Codex allocations the same afternoon; and Cursor doubled the included usage allowance for Grok 4.5 and Composer 2.5 across all paid subscription tiers. The article frames the coordinated refills as a retention war among frontier labs, driven by competitive pressure from new model launches and the impending expiration of Anthropic's Fable 5 promotional access on July 19.
xAI launched Grok Automations on July 16, 2026: describe a recurring task in plain language and set a schedule or trigger. Scheduled automations (one-time, daily, weekday, weekly, monthly, yearly) are free for all Grok users. Email-triggered automations - Grok monitors your inbox and fires an automation when a message matches sender/recipient/subject criteria - require a SuperGrok subscription ($30/month or $300/year). Automations integrate with external services through connectors (Slack, Google Workspace, GitHub, Airtable, Salesforce) and reusable Skills packages, and results are delivered via email notifications (daily/weekly HTML digests) or push alerts, with full run history for review. Live today at grok.com and in the Grok iOS and Android apps.
## What's changed - Added `--forward-subagent-text` flag and `CLAUDE_CODE_FORWARD_SUBAGENT_TEXT` environment variable to include subagent text and thinking in stream-json output - Fixed permission previews relayed to chat channels not neutralizing bidirectional-override, zero-width, and look-alike quote characters, so tool inputs cannot visually alter the approval message - Fixed auto mode overriding a PreToolUse hook's `ask` decision for unsandboxed Bash - a hook `ask` now floors the decision at a prompt - Fixed parallel Claude Code sessions all logging out simultaneously after wake-from-sleep when many sessions share one credential store - Fixed plugin MCP servers not reconnecting after an idle web session woke, leaving MCP calls failing until the next message - Fixed Claude Code on Vertex and Bedrock attempting the default Opus model at startup and printing a spurious fallback notice when a model is explicitly configured - Fixed subagents spawned with an explicit model override reverting to the parent's model when resumed or sent a follow-up message - Fixed nested `.claude/rules/*.md` files loading even when setting sources exclude project settings - Fixed file upload validatio…
## What's changed - Fixed `--replay-user-messages` with `--include-partial-messages` emitting the turn-start user replay after the first content block instead of before the turn's content events - Added `SDKAssistantMessage.timestamp` (ISO-8601) to the live stream, matching `SDKUserMessage`; older emitters omit it, consumers should fall back to receive time - Added rate-limit message prefix buckets (`USAGE_LIMIT_ERROR_PREFIXES` and siblings) as `@alpha` exports for classifying rate-limit messages without hand-mirrored lists - Improved process-exit errors to include the CLI's stderr output, so a failed child reports its actual cause instead of only an exit code ## Update ```sh npm install @anthropic-ai/claude-agent-sdk@0.3.211 # or yarn add @anthropic-ai/claude-agent-sdk@0.3.211 # or pnpm add @anthropic-ai/claude-agent-sdk@0.3.211 # or bun add @anthropic-ai/claude-agent-sdk@0.3.211 ```
xAI open-sourced Grok Build, its terminal-native coding agent, publishing ~844,530 lines of Rust (only ~3% vendored code) at github.com/xai-org/grok-build under the Apache 2.0 license, and reset usage limits for all users. The release follows a privacy controversy: on July 14 security researchers found Grok Build had been uploading users' entire directories - complete Git repositories including commit histories, SSH keys, API keys, cloud tokens and password databases - to xAI's Google Cloud buckets, contradicting earlier claims that nothing from the codebase was transmitted. xAI disabled data retention by default and says it deleted all previously retained coding data. Simon Willison's analysis notes the codebase ships the agent loop, tools, terminal UI and extension system, includes tool implementations adapted from Codex and OpenCode, still contains the (now disabled, hardcoded-unavailable) Google Cloud upload infrastructure, and was published as a single commit - so no development history. Users can now compile Grok Build locally and point it at their own inference. xAI: "Publishing the code is the most direct way to build toward a robust and reliable harness."
Google DeepMind's Gemini 3.5 Pro has now missed three launch deadlines since its original June 2026 release target - the widely reported July 17 general-availability date came and went with no model card, no pricing page, and no gemini-3.5-pro listing in the public Gemini API documentation (which currently lists gemini-3.5-flash and gemini-3.1-pro-preview). Why it slipped: the rebuilt model struggled to meet key performance benchmarks, with reported problems including frequent hallucinations and inconsistent outputs. Issues centered on reliability and practical deployment readiness in competition with OpenAI's GPT-5.6. Google's next steps: the company is exploring interim stopgap releases such as Gemini 3.6 Flash and Gemini 3.5 Flash Light to address specific market needs while buying development time. Speculation suggests Google DeepMind may eventually pivot focus toward next-generation models like Gemini 4.0 Flash, emphasizing real-world applications over benchmark performance. Why it matters for developers cross-shopping AI coding tools: the delay leaves Claude, GPT-5.6, and Grok as the current frontier options for agentic coding while Google regroups.
OpenAI released the Codex Micro, a $230 limited-run keyboard co-designed with specialty keyboard maker Work Louder, built to pair with the Codex coding agent. Features: light-up "Agent Keys" that show agent status, customizable Command Keys as shortcuts for frequent Codex actions, a joystick for launching common workflows, and a dial that adjusts agent reasoning levels (how much computing power a task gets). The device is controllable via the ChatGPT desktop app. The launch lands while OpenAI faces a lawsuit from Apple alleging trade-secret theft related to hardware development (OpenAI denies the allegations); its unreleased smart speaker was reportedly designed by former Apple engineers.
AI security firm Mindgard published full details of an unpatched Cursor flaw: on Windows, Cursor automatically executes a `git.exe` binary found at a project's root when the folder is opened - no clicks, prompts, or approval dialogs. An attacker just commits a malicious `git.exe` to a repository root; anyone who clones and opens it in Cursor runs the poisoned binary as the logged-in user, with access to SSH keys and cloud credentials. Timeline: Mindgard reported the flaw to Cursor on December 15, 2025; Cursor's HackerOne program first closed it as "informative and out of scope", later reopened it, and gave no substantive update after January 20, 2026. Full disclosure landed July 15, 2026. Confirmed in Cursor 3.2.16 (April 2026), and the current release 3.11 (July 10, 2026) is reportedly still vulnerable. No CVE assigned, no security advisory from Cursor, and no patch available at publication. Mitigation until a fix ships: inspect cloned repos for executables at the root before opening them in Cursor on Windows.
xAI has open-sourced Grok Build, its coding agent CLI. The source code is now publicly available at github.com/xai-org/grok-build, described as "SpaceXAI's coding agent harness and TUI. Fullscreen, mouse interactive, extensible." The release enables local-first operation: users can compile and run the tool against their own inference systems using a config.toml file. The published codebase includes: - The agent loop - Tool implementations - Terminal UI components - The extension system covering skills, plugins, hooks, MCP servers, and subagents xAI also maintains a companion Claude Code plugin (xai-org/grok-build-plugin-cc) that delegates reviews, rescue tasks, and session transfer to the Grok Build CLI, plus an official plugin marketplace (xai-org/plugin-marketplace).
## What's changed - Added a live elapsed-time counter to the collapsed tool summary line so long-running tool calls visibly tick instead of looking stuck - Added a startup warning for `Write(path)`, `NotebookEdit(path)`, and `Glob(path)` permission rules - use `Edit(path)` or `Read(path)` instead - Fixed `isolation: 'worktree'` subagents being able to run git-mutating commands against the main repo checkout instead of their own isolated worktree - Fixed the `ultracode` keyword opt-in firing on non-human-originated input such as webhook payloads and relayed PR comments - Fixed a rendered text fragment leaking into crash telemetry when a UI component returned content outside a styled text element - Fixed paste markers leaking into external editors opened from Claude Code, which could appear as stray È/É characters around pasted text - Fixed `claude attach` sometimes failing with "job not found" or "agent is still starting" errors during session transitions - attach now waits for the daemon to settle, and terminal resizes during a slow attach are applied once it completes - Fixed a session crash when a tool's result renderer returned a numeric bigint value or plain text instead of a…
Hello. We have reached 8M active users across Codex and ChatGPT Work. We are once again resetting the usage limits for all. And we continue to not have the 5h rate limit as well, allowing everyone to explore the boundaries of GPT-5.6 Sol and discover how ambitious you can be. See you tomorrow for more updates on our growth!
Five minutes, every Monday - the tools, releases and tactics for developers.