Every pricing change, model release, deprecation, and incident across the AI/ML stack — detected and published by Change Radar.
Anthropic's public status indicated an incident on Apr 3, 2026 reporting elevated errors on requests to Claude Sonnet 4.6; the incident was investigated and marked resolved the same day (times in logs indicate investigation beginning ~18:12 UTC and resolution at ~19:21 UTC).
The Amazon Bedrock pricing page no longer contains the AI21 Labs model pricing rows (Jamba 1.5 Large, Jamba 1.5 Mini, Jurassic-2 Mid, Jurassic-2 Ultra, Jamba-Instruct) that were present in the prior scrape. This removed content affects the per-1M input/output token price table entries and region-specific price listings for those models. No other Bedrock documentation pages (models/regions, quotas, doc-history) showed content changes in this run.
Google announced and published materials for Gemma 4 (an open-model family) on 2026-04-02. Public announcements include a Google Cloud blog post announcing Gemma 4 availability on Google Cloud/Vertex and an updated Gemma 4 model card on the Google AI developer site. The model family is described as open-weight and released under a commercially permissive Apache 2.0 license, with large context windows and multimodal capabilities.
The official Gemini API pricing page was updated and now shows a 2026-04-02 last-updated timestamp. The page contains expanded pricing details across multiple tiers (Standard, Flex, Priority, Batch, etc.), with explicit per‑1M token input/output prices, new context-caching and storage rates, and clarified grounding quotas and costs (notably a 5,000 free prompts grounding allowance referenced on the page). Per-image and per‑media rates and numerous per-tier numeric clarifications were added or expanded throughout the document.
Vercel added Alibaba Qwen 3.6 Plus to AI Gateway. The changelog entry describes how to select the model via the AI SDK (model id e.g., alibaba/qwen3.6-plus) and calls out improved agentic coding and multimodal capabilities.
LangChain released langchain-core==1.2.25 which contains fixes that harden checks for plain-text (.txt) files in deprecated prompt-loading functions, minor documentation/typo fixes, and housekeeping adjustments. The changes further adjust prompt-loading robustness and address small edge cases rather than introducing major new APIs.
Multiple publications report that Anthropic accidentally included a source map / client file that exposed large portions of the "Claude Code" client code. Anthropic initiated large-scale takedown/DMCA requests to remove copies, and later acknowledged some takedowns affected legitimate forks; Anthropic stated no sensitive customer data were exposed.
OpenAI reported a partial outage affecting ChatGPT Dictation with elevated error rates beginning earlier on 2026-04-02. The team implemented a mitigation, moved to monitoring, and marked the incident resolved later the same day.
Multiple community reports on the Google AI Developers forum on Apr 2, 2026 describe Gemini 2.5 Flash variants (Batch, Flash Lite, Flash Image) experiencing jobs stuck in JOB_STATE_PENDING for many hours and 503/UNAVAILABLE errors for image generation requests. These are user reports rather than an official Google status announcement.
OpenAI announced that teams on ChatGPT Business and Enterprise can add Codex-only seats with pay-as-you-go pricing, billed on token consumption (token-based rate card). The Codex rate card was updated to align credits with API token usage instead of per-message estimates, and the company lowered the annual price of ChatGPT Business from $25 to $20 per seat. Eligible ChatGPT Business workspaces can receive limited-time credits ($100 per new Codex-only team member, up to $500 per team).
OpenAI published an incident for elevated errors on the completions API (completionsapi-gpt-5-nano) on Apr 2, 2026. The status updates show the issue was identified in the morning, a mitigation was applied and monitored, and the incident was marked resolved by 09:42 UTC the same day.
A separate Google AI Developers forum thread reports frequent 503 (Server Unavailable) errors when using the Gemini-3-Flash-preview model with an observed failure rate around 50–70%, reportedly ongoing for ~1–2 weeks and impacting batch processing workflows.
OpenAI updated the Codex pricing/rate card to align Codex charges with API token usage rather than per-message pricing. This changes how Codex usage is billed for new and potentially existing customers.
The Gemini API pricing page was updated (last updated 2026-04-01) to add and expand paid-tier pricing detail across multiple tiers and models. Key visible changes include new "Flex" pricing tables with per‑1M token/input/output and context‑caching/storage rates, updated per-image rates for image outputs, and an increase/clarification of free grounding quotas to 5,000 prompts/RPD (noted as shared across Gemini 3). Many per-tier numeric prices and grounding pricing clarifications were added throughout the document.
A Microsoft Q&A thread (question posted 2026-04-01, with activity on 2026-04-03) documents intermittent inference_service_unavailable_error failures affecting the gpt-realtime-1.5 model via the Azure OpenAI Realtime WebSocket API in at least the Sweden Central region. Users report repeated inference_service_unavailable_error and realtime session failures; Microsoft staff have commented in the thread. This appears to be a user-visible availability problem rather than a resolved routine support question.
openai-agents-python published v0.13.4 on 2026-04-01. This release sanitizes AnyLLM response replay inputs before validation and includes small docs/packaging updates. No breaking changes are listed.
Multiple posts on Google’s Gemini API developer forum report operational problems starting Apr 1–2, 2026: clients observing 429 (rate/Model capacity), 503 (service unavailable) errors, WebSocket 1008 errors for certain function types, and batch jobs stuck in PENDING/RUNNING. These are user reports and troubleshooting threads rather than a formal Google status-page incident announcement.
Multiple reports (developer news sites and community reporting) indicate Claude Code users and subscribers began hitting usage/quota limits much faster than expected as of early April 2026. Coverage describes customers on subscription tiers encountering rapid exhaustion of allotted usage and Anthropic acknowledging elevated usage/limit behavior in some contexts.
Multiple reputable news outlets reported that Anthropic accidentally leaked a large portion of Claude Code source (hundreds of thousands of lines/files) around 2026-03-31/2026-04-01 and subsequently issued widespread takedown requests to remove copies from public mirrors. Anthropic stated the exposure was due to human error and (per coverage) indicated no customer credentials were exposed while the company pursued removal and remediation.
Anthropic’s public status page lists incidents on Mar 31–Apr 1 that caused elevated errors and timeouts for Claude models (Opus 4.6, Sonnet 4.6, Haiku 4.5) and temporary unavailability of some desktop connectors; the status entries show these incidents were investigated and later resolved.
Anthropic's public status page shows multiple incident updates on 2026-04-01: elevated errors and timeouts affecting Claude Opus 4.6 and Sonnet 4.6 (timeout window reported from 2026-03-31 17:45 UTC to 2026-04-01 05:52 UTC) and a separate desktop application connection error reported and resolved on 2026-04-01. All incidents were marked resolved in follow-up updates. The status page entries list Identified -> Monitoring -> Resolved timeline entries and recommend monitoring for any continuing impact.
Multiple news outlets reported that Anthropic accidentally exposed source code for its Claude Code product in a packaging/debug file that was published publicly. Anthropic acknowledged the issue as a human/packaging error and pursued takedowns; reporting indicates the company stated no customer data or credentials were exposed.
Multiple developer reports and a news write-up describe Claude Code subscribers hitting usage/quotas unexpectedly and experiencing overages in late March / early April 2026. Users reported rapid consumption of session/token allotments and billing/limit surprises; coverage indicates the problem affected some Max-plan subscribers and intensified during peak windows.
Pricing pages were updated to include Veo 3.1 Lite pricing entries. The docs include 'coming soon' per-resolution guidance (example: $0.10 for 720p, $0.12 for 1080p, $0.30 for 4k) and separate default preview pricing rows (example preview defaults such as $0.05 for 720p and $0.08 for 1080p with 4k not supported on some preview rows).
Billing/AI Studio documentation expanded policy details about Prepay and Postpay flows: prepaid credits expire after 12 months, switching from Prepay to Postpay triggers a refund of unused Prepay credits (subject to processing), guidance and recommendations for configuring auto-reload, and instructions for transitioning to Postpay once eligible. The docs also reiterate that some new users may be required to use Prepay during setup.
Google added the Veo 3.1 Lite video-generation model to the Gemini docs (new model ID entries) and highlighted it as a high-speed, cost-effective variant for video generation at scale. The Gemini 3 developer guide now references veo-3.1-lite-generate-preview alongside other Veo preview models and links usage to AI Studio.
LangChain published langchain==1.2.14 which includes a fix for recursion-limit handling in create_agent, a ~15% init performance improvement, security-related pygments bump, and multiple dependency updates (requests bump and partner/core minimums). The release also includes small fixes to model-provider token counting and other chores that may affect agent initialization and dependency resolution.
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.
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-agents-js published v0.8.2 on 2026-03-31. The patch updates default reasoning effort for newer models, forwards external web access for web-search tools, and includes several test/stability hardenings for streamable HTTP reconnect behavior. No breaking changes are indicated in the changelog.