releases.sh

Frontier AI Labs

Frontier AI research labs and model providers, side by side.

This week: Price cuts, Fast mode, and the end of GPT-5.4 in Codex →
WedAug 5, 20261 release
Anthropic
Daily summary

Claude Code fixes Bash permission bypass

Claude Code patched a Bash permission bypass that let crafted commands evade permission checks, plus similar gaps in prompts and sandboxes, while adding a warning for restricted subagent models and making /review an alias of /code-review.

  • Claude Code fixed a Bash permission bypass where crafted commands could hide from permission checks, along with related gaps in permission prompts and workflow sandboxes.
  • Claude Code now warns when a requested subagent model is restricted and the parent model runs instead, and added a /teleport hint for continuing cloud sessions locally.
  • Claude Code added owner wildcard entries to marketplace management settings and made /review an alias of /code-review.
AnthropicAnthropic1
Claude Code

Bash permission bypass fixed; sandbox escapes closed

v2.1.223

Fixed a Bash permission bypass where crafted commands could hide from permission checks, plus similar gaps in permission prompts and workflow sandboxes. 1M-context models are now held to 200K via auto-compaction, and unknown model IDs keep sessions within the assumed window unless explicitly disabled. Also added a warning when a requested subagent model is restricted, a /teleport hint in cloud sessions, and made /review an alias of /code-review.

Read post
TueAug 4, 20262 releases
OpenAIAnthropic
Daily summary

Claude Code patches worktree and hook-bypass flaws

Claude Code's latest release closes two security holes — destructive git commands from worktree-isolated sessions and a PreToolUse hook bypass — while also fixing an HTTPS proxy hang and a usage-credits block for dismissed Team/Enterprise requests.

  • Claude Code now enforces worktree isolation on file edits and Bash across all session types, preventing destructive git commands against the main checkout.
  • PreToolUse auto-allow hooks can no longer be bypassed in background agent tasks like summaries, compaction, and renames.
OpenAIOpenAI1
AnthropicAnthropic1
Claude Code

Worktree isolation strengthened; hook bypass fixed

v2.1.222

Fixed a security issue where worktree-isolated sessions could run destructive git commands against the main checkout, and another where PreToolUse auto-allow hooks could be bypassed in background agent tasks; isolation now applies to file edits and Bash in every session type. Also fixed a startup connectivity check that hung behind HTTPS proxies and a usage-credits block for Team/Enterprise members with dismissed requests, plus a range of other fixes across MCP attribution, Pull Request linking, and screen-reader mode.

Read post
MonAug 3, 202611 releases
AnthropicOpenAI
Daily summary

Claude Code fixes a bash permission bypass; OpenAI SDKs add gpt-5.5

Claude Code patched a critical Bash tool permission-check bypass via zsh hidden commands, while the OpenAI Node and Python SDKs added gpt-5.5 model support and new tool metadata fields.

  • Claude Code fixed a Bash tool permission bypass where zsh regex conditionals could execute hidden commands, and added a Focus view for collapsing tool activity and Linux/WSL sandbox credential masking
  • OpenAI's Node and Python SDKs added the gpt-5.5 model, and the Node SDK also introduced tool name and namespace metadata fields to the API
AnthropicAnthropic1
Claude Code

Bash tool permission bypass fixed; sandbox credential masking shipped

v2.1.221

Fixed a Bash tool permission-check bypass where zsh could execute hidden commands in regex conditionals. Added Focus view for collapsing tool activity, and sandbox credential file masking on Linux and WSL. Plus dozens of fixes across MCP, Vim mode, PowerShell, and session management.

Read post
OpenAIOpenAI10

ChatGPT Enterprise & Edu

3 updates
Large pastes now auto-convert to attachments for Enterprise & Edu

Pasting more than 10k characters into the composer now automatically converts the content into an attachment instead of inserting it directly into the text field, keeping the composer cleaner and preserving context window space.

Read post
College Educator, K–12 Educator, and College Student plugins launch

ChatGPT Edu workspaces now have access to three guided plugins — College Educator, K–12 Educator, and College Student — for creating lessons, assessments, and study plans grounded in course materials. ChatGPT for Teachers workspaces have access to K–12 Educator. Plugins are not installed by default; admins manage availability.

Read post
Weekly usage limits migrate to monthly; auto-migration August 15

Workspace owners can now migrate weekly role-based limits to monthly limits from Permissions & roles. Remaining workspaces on weekly limits will be automatically moved on August 15.

Read post

ChatGPT

2 updates
Sign in with ChatGPT rolls out to six partner sites

Sign in with ChatGPT is rolling out across select plugins and partner sites, starting with Airtable, GitLab, HubSpot, Notion, Supabase, and Vercel.

Read post
Large pastes auto-convert to attachments for Enterprise and Education

Pasting over 10k characters in ChatGPT for Enterprise and Education now automatically converts the content to an attachment, keeping the composer cleaner and preserving context window space. Users can revert to inline text via "Show in text field".

Read post

K–12 and college plugins for ChatGPT Work and Codex

New plugins for K–12 educators, college educators, and college students bring agentic capabilities to ChatGPT Work and Codex, enabling tasks like lesson planning, syllabus updates, and personalized study guides using course materials and approved tools.

Read post
K–12 educator plugin screenshot 1
Node SDKv7.4.0
Go SDKv3.50.0
FriJul 31, 20263 releases
OpenAI
Daily summary

OpenAI SDKs add content provenance verification

All three OpenAI SDKs gained content provenance checks for verifying AI-generated content authenticity, while the Python SDK also improved retry handling for rate limits.

  • OpenAI's Python, Node, and Go SDKs all added API support for content provenance checks, letting developers verify the origin and authenticity of AI-generated content
  • The Python SDK now honors Retry-After delays up to two minutes, improving rate-limit resilience
OpenAIOpenAI3
Python SDKv2.52.0
Node SDKv7.3.0
Go SDKv3.49.0
ThuJul 30, 20263 releases
OpenAI
Daily summary

OpenAI slashes Luna pricing 80%, adds Sol Fast mode

OpenAI dramatically cut GPT-5.6 Luna pricing by 80% and Terra by 20%, while launching Fast mode for GPT-5.6 Sol in the API at 2.5× speed for twice the price, replacing the Priority Processing tier.

  • OpenAI cut GPT-5.6 Luna pricing 80% and Terra pricing 20%, passing on efficiency gains from the model's self-optimization
  • GPT-5.6 Sol in the API gained Fast mode with up to 2.5× faster speeds, replacing Priority Processing as a backward-compatible tier at double the base price
  • The Python, Go, and Node SDKs all added support for the new Fast tier in API helper methods
OpenAIOpenAI3
ChatGPT

DALL·E GPT retired August 30; migrate to ChatGPT Images

The official DALL·E GPT in ChatGPT will be retired on August 30, 2026. Users should download images they want to keep before then and use ChatGPT Images for continued image creation and editing. User-created GPTs with image generation are not affected.

Read post
Python SDKv2.51.0