Every pricing change, model release, deprecation, and incident across the AI/ML stack — detected and published by Change Radar.
Developers reported Gemini API batch jobs and Deep Research requests becoming stuck (jobs stuck in pending or in_progress) beginning in late March; multiple support/discussion threads cite jobs not progressing and timeouts for batch/deep-research endpoints. Posts indicate users experiencing blocked pipelines and long-running requests.
LangGraph released 1.1.4 which adds LangSmith integration metadata, fixes a recursion-limit default sentinel collision, and applies dependency bumps (pygments, cryptography, requests). The LangSmith metadata addition surfaces integration/observability metadata from langgraph and the recursion fix reduces potential sentinel collisions when creating agents or runtime graphs.
OpenAI posted an incident reporting elevated failures for some users joining ChatGPT Enterprise and ChatGPT Edu workspaces via SSO; affected users could see authentication errors (e.g., registration_disallowed) after completing their organization’s SSO flow. The incident was marked resolved with a recovery update on 2026-03-31.
Mistral announced 'Spaces', an opinionated CLI that scaffolds projects, generates configs, and supports both interactive human use and headless/agent-driven workflows. It provides flags/config equivalents for interactive prompts, a plugin registry for components, and generated context files to let agents operate reliably without manual intervention.
Anthropic's public status page reports an incident on 2026-03-31: elevated error rates impacting Claude Opus 4.6. The incident was investigated and later marked resolved (updates show Investigating -> Monitoring -> Resolved, with resolution noted at 09:44 UTC).
openai-agents-python published v0.13.3 on 2026-03-31. The release fixes stale hydrated input IDs in the server conversation tracker and improves acceptance of raw image_url content parts on chat-completion inputs. These are bugfixes and do not announce breaking changes.
Multiple news outlets reported that Anthropic accidentally published a Claude Code npm package that included a source map exposing original source code. Anthropic acknowledged the exposure in coverage, describing it as a packaging mistake; reports indicate no confirmed customer-data breach but the proprietary code was exposed publicly until removed.
OpenAI updated Containers billing to specify a per-20-minute-session charge. The pricing table shows container rates of 1 GB $0.03, 4 GB $0.12, 16 GB $0.48, 64 GB $1.92 and clarifies these will be billed per 20-minute session per container starting March 31, 2026. The page also notes that container pricing includes Hosted Shell and Code Interpreter and that eligible container sessions are billed at the full 20-minute session rate.
A reported account-state bug on the Google AI Developer forums describes a user who deleted a Tier 1 paid project and subsequently had all Free Tier API keys become unavailable and return Status 429 (Resource Exhausted). Cloud Billing support reportedly closed a related case and forum participants requested engineering intervention to reset account state. The issue prevents creation of new purely free projects and forces billing linkage for new keys.
Weaviate published v1.36.8 on 2026-03-30. The release is a patch that fixes a vector index compression error, improves some test hardlink behavior, adds baseURL validation support, and includes a few backward-compatibility/refactor items. The release notes list no Breaking Changes and no licensing or pricing updates.
AWS's Service Health Dashboard shows a multiple-services operational issue on March 30, 2026 (logged as a global/multiple-services event). AWS listed updates in the health dashboard for that event; depending on region and service dependencies this could have affected Bedrock or related services for some customers.
The anthropics/claude-code repository published release v2.1.88 on 2026-03-30. The release adds CLAUDE_CODE_NO_FLICKER=1 (opt-in flicker-free alt-screen rendering), a PermissionDenied hook (with retry guidance), named subagents in mention typeahead, and a broad set of reliability/UX bug fixes (including prompt-cache, streaming/error messaging, memory leaks, and tool/integration fixes).
Developers and users reported receiving 429 RESOURCE_EXHAUSTED errors from the Gemini / Google AI Studio APIs on March 29, 2026; forum threads show users experiencing rate-limit errors even under low usage and multiple reports of quota/Resource exhausted responses. Community threads and status-aggregator signals indicate Google is aware and users are troubleshooting quota/configuration or backend throttling.
Anthropic's public status page reports a Dispatch-related incident on 2026-03-29 where Dispatch sessions in the Claude Desktop app stopped responding. The incident was identified and resolved; the status message instructs affected users to update the Claude Desktop app to version 1.1.9493 and notes Cowork sessions are unaffected as a workaround.
The anthropics/claude-code repository published a new release entry (v2.1.87) on 2026-03-29. The release notes state it fixes messages in Cowork Dispatch not being delivered and includes reliability/UX fixes that align with recent SDK/agent parity work.
OpenAI updated its Sora discontinuation guidance: the Sora web and app experiences will be discontinued on 2026-04-26, and the Sora API will be discontinued on 2026-09-24. The article instructs users to export content before the web/app shutdown and describes possible limited export windows and permanent deletion after export windows close.
The OpenAI pricing page was updated to change web-search / search-tool pricing. The page now shows "Web search (all models) $10.00 / 1k calls" and adds separate "Web search preview" lines: reasoning models (including gpt-5 and o-series) at $10.00 / 1k calls and non-reasoning-model previews at $25.00 / 1k calls. The page also clarifies which search content tokens are billed at model rates versus being free.
LangChain released langchain-core==1.2.23. The release reverts a previous change that traced invocation parameters in metadata and bumps requests from 2.32.5 to 2.33.0 (plus other chores). The notes show a revert of a trace-invocation metadata change and dependency updates which may affect tracing/metadata visibility and dependency resolution in agent environments.
Anthropic publicly acknowledged the issue and said it adjusted Claude Code’s 5‑hour session limits during peak hours to manage demand. The company said weekly limits remain unchanged and estimated roughly 7% of users will hit the more aggressive session limits; affected windows are weekdays 5am–11am PT (1pm–7pm GMT).
anthropics/claude-code published release v2.1.86 (2026-03-27) which includes changes such as adding an X-Claude-Code-Session-Id header to API requests and a variety of reliability and UX fixes; the release notes indicate parity work with agent/SDK updates.
Anthropic published a sequence of TypeScript Claude Agent SDK releases in late March 2026 (notably v0.2.85 on 2026-03-26 and v0.2.86 on 2026-03-27). The releases add SDK methods (eg. reloadPlugins(), getContextUsage()), make session_id optional for SDKUserMessage, introduce parity changes with Claude Code, and include various bug fixes and type/dependency updates.
OpenAI updated the ChatGPT Release Notes on March 27, 2026 to roll out updated Box, Notion, Linear, and Dropbox apps with new app actions (including new write capabilities where supported) and the latest app experiences. The notes advise reconnecting apps to begin using the updated experiences and call out other recent product changes (e.g., model retirements and feature updates).
LangChain published langchain-exa==1.1.0 (partner package). Key changes include switching the default search type from neural to auto and bumping the minimum langchain-core dependency to 1.2.21, plus various dependency and CI updates.
Anthropic published multiple updates to the TypeScript Agent SDK (releases v0.2.76 through v0.2.84 in March 2026, with v0.2.84 published 2026-03-26) adding features such as taskBudget, enableChannel()/McpServerStatus capabilities, exported EffortLevel type, and various fixes to parity with Claude Code releases.
The docs now state that the free Grounding with Google Search prompts allowance (e.g., 1,500 RPD or 5,000 prompts per month) is “free, shared across Gemini 3” — clarifying that grounding free quotas are pooled across Gemini 3 models rather than per-model. Pricing for paid search queries remains $14 / 1,000 search queries.
Cohere released 'Transcribe' (cohere-transcribe-03-2026), an open-source automatic speech recognition (ASR) model (~2B parameters) available on Hugging Face and announced on Cohere's blog, published under an Apache 2.0-style license and intended for high-accuracy transcription use cases.
Multiple developer reports and a Forbes article (Mar 26, 2026) describe a pricing/usage bug in Claude Code where users' monthly/session usage allotments were consumed far faster than expected. Reports indicate some accounts saw their 5-hour session limits or paid-plan balances depleted in minutes to hours. There is no clear, dated Anthropic official announcement or corrected billing effective date in the sources found.
The anthropics/claude-code repository published a new release (v2.1.85) on 2026-03-26 that adds MCP environment variables (CLAUDE_CODE_MCP_SERVER_NAME and CLAUDE_CODE_MCP_SERVER_URL), conditional hook filtering, timestamp markers for scheduled tasks, deep link length changes, RFC 9728 discovery for MCP OAuth, plugin/managed-settings behavior changes, and many fixes and improvements.
Public reporting and status updates indicate a service incident on March 26–27, 2026 that caused elevated error rates and throttling for Claude (reports reference Opus 4.6 and session limits). Anthropic published a retrospective/updates on the outage via status channels.
The anthropics/claude-code GitHub releases show v2.1.84 published on 2026-03-26 with multiple changes: new PowerShell tool opt-in preview, environment variables to override model detection, streaming idle timeout configuration, new hook events, a new header for debugging, VSCode rate-limit banner, many bug fixes and improvements.