Spend limits applied to individual members; group-restricted upstream routing and pooled group budgets were not available through these settings. -> Administrators can restrict upstream access by group and impose a shared monthly group budget alongside each member's individual limit.
Failed SDK hook callbacks could let actions run, and auto mode could fall back to its built-in classifier when server checks were unavailable. -> Builders can make hook failures block execution, and affected auto-mode sessions require approval or deny tool calls that cannot receive server checks.
The Agent tool lacked an explicit effort parameter, and subagent definitions lacked a separate auto-compaction window. -> Builders can request subagent effort levels, compact subagent context earlier than the main conversation, and override workflow-agent models independently of other subagents.
Idle terminal conversations did not have the newly added automatic compaction timed to precede prompt-cache expiry. -> Long terminal conversations can be compacted while idle before cache expiry, so the next message does not reprocess the entire conversation.
Builders could configure hooks and tool-call controls, but lacked the newly added skipped-hook transcript notices and per-intervention safety telemetry. -> Builders can identify skipped hooks in transcripts and collect events when permission rules, auto mode, safety checks, hooks or the sandbox stop, prompt or send back a tool call.
Plugins did not have the newly introduced mod capabilities or the $.ui.selection() interface for reading fullscreen selections. -> Plugin builders can modify deeper Claude Code behavior and give mods access to user-selected transcript context.
The terminal workspace supported a side conversation beside the main chat. -> Users can switch among multiple side chats with the keyboard or mouse and reopen saved local side chats across restarts.
Codex CLI did not have a built-in Google Cloud Vertex AI provider. -> Builders can configure model_provider = "gcp" and use GPT-6.1 Terra through Codex's built-in Vertex AI integration.
Both Deep Research Preview and Deep Research Max Preview were listed with 131,072-token context windows. -> The listed limits accommodate research inputs exceeding the previous 131,072-token ceiling, subject to runtime confirmation.
Builders could grant agents access, but these items do not describe spend caps tied to agent identity or additional authorization scopes requested during a session. -> Builders can attribute and limit agent spend by principal and request additional permissions for specific actions during a session.
Computer interaction was not established here as a generally available, priced model feature builders could select through public APIs and products. -> Builders can use generally available API computer use or choose a fast, cheaper computer-use model in OpenAI's products and API.
Builders could evaluate agent responses, but the items do not describe final database state checks serving as both benchmarks and training rewards. -> Builders can score whether an agent achieved a task’s final state and use that result as an evaluation or training signal.
Builders had no cited public composite index for comparing models that select among fixed options. -> Builders can consult a documented composite benchmark to compare decision models across four evaluation areas.
Builders could use hosted gateways for model calls, but these items do not establish access there to the newly reported decision-model options and Decisions API. -> Builders can access decision-oriented models or APIs through multiple hosted AI gateway offerings.
Builders could not use these newly launched MAI audio models for real-time transcription or voice generation. -> Builders can access the MAI audio models, including through Vercel AI Gateway, for audio applications.
Agents needed manually authored JSON schemas or code to specify how to parse and extract fields from documents. -> Agents can interactively create and run extraction pipelines by describing the desired fields in natural language, enabling on‑the‑fly schema generation.
Operators can increasingly inspect what an agent stored, recalled, merged, and spent instead of treating memory as hidden prompt state.
BeforeMemory is hidden inside session context; operators infer its effects from an answer or failure.
Enabling changeMemory operations and long-running trajectories are emitted into observable runtime surfaces.
AfterAn operator can review memory actions, cost, confidence, and their role in a run.
Debugging and governance shift from reconstructing hidden context to inspecting evidence, while privacy and retention risks become visible obligations.