Weaviate v1.36.6 and v1.35.15 (2026-03-19): audio + Gemini/multi2vec-google updates, backup and replication fixes; no breaking changes
Weaviate published two patch/minor releases on 2026-03-19 (v1.36.6 and v1.35.15). The updates include audio support in the multi2vec-google module (Google AI Studio / Gemini-related modules), async replication binary-encoding improvements, backup chunking / split-file handling fixes, IPv6 clustering support, various race-condition/test fixes, and a security bump of google.golang.org/grpc. All release notes list "none" under Breaking Changes and no license or pricing changes were announced.
Azure Foundry model retirements page updated (page updated 2026-03-19): new GA entries and revised retirement dates
Microsoft updated the Azure OpenAI / Microsoft Foundry model deprecations & retirements page on 2026-03-19. The page now lists new GA model entries (example: gpt-5.4 and gpt-5.4-pro marked GA with launch dates in early March 2026, and gpt-5.4-mini/nano GA entries dated 2026-03-17), plus added or clarified retirement dates for multiple models (examples: several gpt-4o standard deployment retirements remain 2026-03-31 with auto-upgrades from 2026-03-09; audio preview models such as gpt-4o-audio-preview and related realtime/mini variants show retirements on 2026-03-24; tts and whisper retirement windows updated to mid-June 2026). The update also included metadata and page-version updates (updated_at 2026-03-19).
The openai-agents-python repository published several follow-up releases after 2026-03-14: v0.12.3 (2026-03-16), v0.12.4 (2026-03-18), and v0.12.5 (2026-03-19). These releases include multiple MCP/streaming fixes (normalizing canceled MCP invocations to tool errors, retry improvements for transient MCP and streamable-http failures), exposure of auth/httpx client factory hooks for SSE/StreamableHttp, and other stability/documentation fixes.
Vercel added MiniMax M2.7 (standard and high-speed variants) to AI Gateway. The high-speed variant offers ~100 tokens/sec at ~2x cost. The changelog entry notes model IDs to use in the AI SDK and reiterates AI Gateway features such as provider routing, BYO key, and observability.
Migration guide: Migrate from Amazon Nova 1 to Amazon Nova 2 on Amazon Bedrock (Mar 18, 2026)
AWS published a migration/technical blog (Mar 18, 2026) explaining how to migrate Amazon Nova 1→Nova 2 on Amazon Bedrock. It includes recommended migration paths (Nova 1 Lite/Pro/Premier → Nova 2 Lite), API/model ID changes, new capabilities (extended thinking, built-in tools, 1M token context), pricing examples for Nova 2 Lite, and breaking/integration notes (e.g., parameter restrictions when using high reasoning effort).
GPT-5.4 mini rolled out in ChatGPT; used as fallback for rate limits (Mar 18, 2026)
OpenAI announced the rollout of GPT-5.4 mini in ChatGPT. GPT-5.4 mini is available to Free and Go users via the “Thinking” feature and will be used as a rate-limit fallback for other users; GPT-5.4 Thinking mini will be retired as a selectable option in 30 days. Enterprise customers retain options for default auto-routing.
GPT‑5.4 mini rolled out into ChatGPT (Mar 18, 2026)
OpenAI rolled out GPT‑5.4 mini into ChatGPT on March 18, 2026: GPT‑5.4 mini is now available to Free and Go users via the “Thinking” feature in the + menu, and for paid users it is available as a rate-limit fallback for GPT‑5.4 Thinking. This behavior was added to the ChatGPT documentation / release notes and documents how mini will be surfaced in different tiers.
Gemini API: Built-in Tools + Function Calling combination feature released (Mar 18, 2026)
Gemini API release notes include a March 18, 2026 entry that announces a new Built-in Tools and Function Calling Combination feature, enabling use of built-in tools and function calling together in the same request/flow. The changelog entry describes the capability being available for Gemini 3 models and provides developer guidance in the docs.
Minimax M2.5 and GLM 5 models now available on Amazon Bedrock (Mar 18, 2026)
AWS announced Minimax M2.5 and GLM 5 are now available on Amazon Bedrock (posted Mar 18, 2026). The announcement describes GLM 5 as a frontier-class general-purpose LLM for long-horizon agentic and systems-engineering tasks and Minimax M2.5 as an agent-native frontier model optimized for token-efficient reasoning and workflow completion.
The langgraph repository published new releases on 2026-03-18: langgraph v1.1.3, langgraph-sdk v0.3.12, and langgraph-checkpoint-postgres v3.0.5. Release notes mention a new "execution info" addition to the runtime, several dependency bumps, and a fix to re-use connections in checkpoint-postgres. No explicit deprecations or breaking-change notices were included in the release notes.
NVIDIA Nemotron 3 Super now available on Amazon Bedrock (Mar 18, 2026)
AWS announced that NVIDIA Nemotron 3 Super is now available on Amazon Bedrock (posted Mar 18, 2026). The page states Nemotron 3 Super is available across select AWS Regions and describes it as an open hybrid MoE model for complex multi-agent/agentic workloads, with full openness (weights, datasets, recipes) and suitability for customization and secure enterprise deployment.
langchain-anthropic==1.4.0 released (2026-03-17) — explicit caching for system messages & tool definitions
LangChain published langchain-anthropic==1.4.0 on 2026-03-17. The release notes add AnthropicPromptCachingMiddleware to apply explicit caching to the system message and tool definitions, and expose/delegate a cache_control (cache%5C_control) kwarg to the top-level Anthropic parameters. No explicit breaking-deprecation marker was present in the release notes.
Increased error rates and latency for Gemini API (Mar 17, 2026)
Google AI Studio's status page reports increased error rates and latency for the Gemini API on March 17, 2026. Google identified the issue, deployed mitigations, and marked the incident as resolved in their status history.
Multiple reports and a status incident reported elevated errors and partial outages for Claude services on March 17–18, 2026. Users and media (Downdetector, news sites, and community posts) reported disruption; Anthropic posted status updates pointing to elevated error rates and mitigation work.
Mistral launches Forge — enterprise platform for building custom frontier models
Mistral launched Forge, an enterprise platform that enables organizations to build, train, and continuously improve frontier-grade models grounded in their proprietary data. Forge supports dense and Mixture-of-Experts architectures, multimodal inputs, reinforcement-learning pipelines, and is positioned for agent-first workflows and enterprise control over models and data.
ChatGPT release notes updated: model picker simplified and availability notes (Mar 17, 2026)
The ChatGPT release notes were updated (March 17, 2026) to include UI changes to the model picker — a simplification to make choosing reasoning level easier — and to document model availability/behavior in ChatGPT (including mini being used for Thinking or as a fallback).
OpenAI releases GPT-5.4 mini and GPT-5.4 nano (Mar 17, 2026)
OpenAI announced GPT-5.4 mini and GPT-5.4 nano on March 17, 2026. GPT-5.4 mini is available in the API, Codex, and ChatGPT (as a Thinking model and as a rate-limit fallback) with a 400k context window; pricing in the announcement is $0.75 per 1M input tokens and $4.50 per 1M output tokens. GPT-5.4 nano is API-only and priced at $0.20 per 1M input tokens and $1.25 per 1M output tokens. The post describes performance benchmarks, recommended use cases (coding, subagents, low-latency workloads), and guidance about using mini/nano as subagents or fallbacks.
Amazon Bedrock now available in Asia Pacific (New Zealand) (Mar 17, 2026)
AWS announced Amazon Bedrock is available in the Asia Pacific (New Zealand) Region starting Mar 17, 2026. The announcement lists model availability in the region—Anthropic (Sonnet 4.5, Sonnet 4.6, Opus 4.5, Opus 4.6, Haiku 4.5) and Amazon (Nova 2 Lite)—and points customers to the Bedrock product page and region/model compatibility docs for details.
Open SWE released — open-source framework for internal coding agents (built on Deep Agents & LangGraph)
LangChain published a blog post announcing Open SWE, an open-source reference framework for internal coding agents built on top of Deep Agents and LangGraph. Open SWE packages patterns such as isolated sandboxes, curated toolsets, subagent orchestration, and middleware hooks as a reusable framework and provides installation and customization guides on GitHub.
Multiple March 2026 Claude product updates (Cowork, visuals, Office integrations, memory for free users)
Anthropic support release notes for March 2026 list several product updates: (Mar 17) Cowork persistent thread control on mobile (Pro/Max preview); (Mar 12) interactive charts/visualizations in responses; (Mar 11) improved Excel and PowerPoint integrations; (Mar 02) memory available for free users. These are documented on Anthropic's release notes page.
Gemini API changelog: revamped Usage Tiers and Billing Account spend caps announced (Mar 16, 2026)
The official Gemini API release notes were updated on March 16, 2026 to announce revamped Usage Tiers and Billing Account spend caps. The billing docs were also updated with details: tier spend caps are enforced at the billing-account level, preset values are shown (Tier 1: $250; Tier 2: $2,000; Tier 3: $20,000–$100,000+), and the docs state enforcement of tier spend caps will begin April 1, 2026.
AWS Partner Central agents powered by Amazon Bedrock AgentCore (Mar 16, 2026)
AWS announced AWS Partner Central agents powered by Amazon Bedrock AgentCore, available to Partners who have migrated to the new Partner Central experience. The post describes agentic capabilities embedded into Partner Central, MCP integration options for connecting tools, and migration/get-started guidance for partners.
New CloudWatch metrics for Amazon Bedrock: TimeToFirstToken and EstimatedTPMQuotaUsage (Mar 16, 2026)
AWS announced two new Amazon CloudWatch metrics for Amazon Bedrock — TimeToFirstToken and EstimatedTPMQuotaUsage — to help customers observe model latency and estimated token‑per‑minute quota usage. The announcement explains the metrics and how they can be used to monitor Bedrock performance and quota consumption.
Mistral joins NVIDIA Nemotron Coalition and highlights release of Mistral Small 4
Mistral announced participation as a founding member of the NVIDIA Nemotron Coalition and simultaneously highlighted the release of Mistral Small 4. The announcement frames a co-development effort with NVIDIA to train and open-source frontier-level base models using NVIDIA compute and tooling.
Mistral releases Leanstral: an open-source code agent (Apache 2.0) and free labs endpoint
Mistral published Leanstral, an open-source code agent optimized for Lean 4. They released model weights under an Apache 2.0 license, made it available via a free/near-free labs API endpoint (labs-leanstral-2603), and provided documentation and a tech report. The release includes evaluation results and a download for self-hosting.
LangChain and NVIDIA announce enterprise agentic AI platform (LangGraph + Deep Agents + NVIDIA)
LangChain announced a strategic enterprise partnership with NVIDIA to deliver an enterprise agentic AI platform that integrates LangGraph, Deep Agents, and NVIDIA infrastructure. The announcement highlights a commercial platform for building, deploying, and monitoring production-grade AI agents at scale.
Anthropic removes long-context surcharge — 1M-token contexts billed at standard per-token rates
Anthropic removed its previous long-context surcharge for Claude Opus 4.6 and Claude Sonnet 4.6, making 1,000,000-token context windows available at standard per-token pricing. The change also removes the previous requirement/beta header for very large requests and expands multimodal request capacity (higher images/PDF pages per request) as part of the rollout.
GPT-5.3 & GPT-5.4 availability and per-tier usage limits added to ChatGPT docs
The Help Center article 'GPT-5.3 and GPT-5.4 in ChatGPT' was updated to document rollout and per-tier usage limits, context window sizes, and availability. It states GPT-5.3 Instant is rolling out to all ChatGPT users and lists explicit message limits (e.g., Free: 10 messages/5 hours; Plus/Go: 160 messages/3 hours) and Thinking-model weekly limits and context window capacities by tier.
GPT-5.3 Instant update improves follow-up tone and reduces teaser-style phrasing
OpenAI updated ChatGPT release notes on March 16, 2026 to roll out a targeted update to GPT-5.3 Instant that improves follow-up tone and reduces teaser-style phrasing. The change is described as a behavior/tuning update to make follow-ups clearer and reduce abrupt or overly declarative phrasing.
Azure's status reporting recorded an availability degradation for the Azure OpenAI Service on March 16, 2026 which Microsoft marked as mitigated/resolved in their status history.