OpenAI is winding down Agent Builder on Nov 30, 2026. Replacement path: continue with the Agents SDK (Python + TypeScript) or ChatGPT Workspace Agents. If you built agents in Agent Builder's visual canvas, plan a code-based migration over Q3/Q4. The Agents SDK is the more durable foundation — same primitives, full control, and now has TypeScript parity (shipped May 6). (OpenAI deprecations)

OpenAI is sunsetting its Evals product. Two dates to track: Oct 31, 2026 — existing evals become read-only (no new runs). Nov 30, 2026 — the Evals dashboard and API are fully shut down. No first-party replacement announced. Export any eval results, datasets, and grading configs before Oct 31. Plan to move to OSS frameworks (OpenAI Evals OSS repo, LangSmith, Braintrust, Inspect AI) or build a thin internal eval harness against your model gateway. This is a meaningful gap — worth flagging to anyone running production eval pipelines on OpenAI Evals. (OpenAI deprecations)

The v1/prompts API and OpenAI-managed reusable prompt objects shut down Nov 30, 2026. Replacement: move prompt content into your application code. What it means: OpenAI is exiting the prompt-registry business — treat prompts as application configuration (Git + config files, DB rows, feature-flag system) rather than an OpenAI resource. Action: inventory any code referencing v1/prompts, move prompts to your own store, and version them in your repo for reproducibility. (OpenAI deprecations)

OpenAI announced shutdown of gpt-image-1-mini, gpt-image-1.5, and chatgpt-image-latest on Dec 1, 2026. Replacement is gpt-image-2. This is notable because gpt-image-1 was the migration target for the May 12 DALL·E 2/3 shutdown — so anyone who just migrated to gpt-image-1-mini for cost reasons gets a second migration before year-end. Plan capacity and prompt-compatibility testing for gpt-image-2 over Q3 to avoid a rushed November cutover. No price/quality details on gpt-image-2 surfaced yet. (OpenAI deprecations)

The entire Gemini 2.0 Flash family is now offline: gemini-2.0-flash, gemini-2.0-flash-001, gemini-2.0-flash-lite, and gemini-2.0-flash-lite-001. Migrate to gemini-3.5-flash (next-gen mid-tier, went GA May 19) or gemini-3.1-flash-lite (cost/latency-optimized). If anything in production, RAG pipelines, embedding workers, or agent tool-routing still references 2.0 Flash SKUs, it will now error — fix is a string change. Worth auditing today. (Gemini API changelog)

Anthropic released claude-opus-4-8. Standard pricing is unchanged from Opus 4.7: $5/1M input, $25/1M output. Fast mode is the headline upgrade — 2.5x speed at $10/1M input, $50/1M output, which Anthropic states is 3x cheaper than the previous fast mode tier. Three additional shipping features: (1) Messages API now accepts system entries inside the messages array, enabling mid-run system updates (useful for agents that need to re-scope mid-task without restarting the conversation); (2) Claude Code dynamic workflows (research preview); (3) effort control in claude.ai and Cowork. Opus 4.8 is now the top-tier choice for high-stakes reasoning and long-context coding agents — reserve for jobs where Sonnet 4.6 hits quality limits. (Anthropic news)

Google moved its native visual generation models to GA: gemini-3.1-flash-image (fast/cost-optimized) and gemini-3-pro-image (frontier quality). Notable new capability on Flash Image: video-as-context for video-to-image generation — pass a video and generate keyframes/derivatives without a separate vision pipeline. Preview SKUs deprecated effective May 28 and shut down June 25, 2026: migrate gemini-3.1-flash-image-preview → gemini-3.1-flash-image, and gemini-3-pro-image-preview → gemini-3-pro-image. Worth comparing against OpenAI gpt-image-1 and the older Imagen/Veo workflows. (Gemini API changelog)

gemini-3.1-flash-lite-preview is offline as of May 25, 2026, as scheduled. Any code still pointing at the preview SKU will now error — migrate to the GA model name gemini-3.1-flash-lite. If you missed this and are seeing outages today, the swap is a string change in your model parameter; no schema or pricing differences. (Gemini API changelog)

OpenAI released workload identity federation: trusted workloads can exchange externally issued identity tokens for short-lived OpenAI access tokens without storing long-lived API keys. This is a practical security + ops win for agent backends running on managed compute (CI/CD, Kubernetes, serverless): fewer secrets to rotate, reduced blast radius, and better alignment with enterprise IAM. Action: plan migration from static OpenAI API keys for production runtimes. (OpenAI API changelog)

Three releases at once. gemini-3.5-flash is GA — next-gen mid-tier model positioned between Flash-Lite and 3.1 Pro for general coding/RAG/agent work. Managed Agents (public preview) runs autonomous, stateful agents inside secure Google-hosted Linux sandboxes — Google operates the runtime, you orchestrate. antigravity-preview-05-2026 is a Managed Agent that can plan, code, manage files, and browse the web inside the sandbox. Together this is Google's enterprise answer to OpenAI Agents SDK and Anthropic Managed Agents — fully hosted, no sandbox infrastructure to run. Worth evaluating against the OpenAI/Anthropic stacks if you've been blocked on agent runtime ops. (Gemini API changelog)

Account-led GA of Secure MCP Tunnel — lets ChatGPT web, Codex, the Responses API, and AgentKit connect to private/on-prem MCP servers via a customer-hosted tunnel client. No public exposure of internal MCP endpoints, no inbound firewall holes, and works across the full OpenAI agent surface. Material unlock for regulated industries (finance, healthcare, gov) that couldn't expose internal tools to public AI APIs. Pair with self-hosted MCP servers for the most secure tool-use posture available from OpenAI today. (OpenAI API changelog)

Two major adds to Claude Managed Agents. Self-hosted sandboxes (public beta): tool execution runs in your infrastructure (or via Cloudflare / Daytona / Modal / Vercel) while orchestration stays on Anthropic — data and code never leave customer infra. MCP tunnels (research preview): lightweight gateway uses a single outbound connection (no inbound firewall rules, no public endpoints, encrypted end-to-end) and works in both Managed Agents and the Messages API. Same enterprise security pattern OpenAI shipped today — the three vendors are now competing on identical agent-runtime feature surfaces. (Claude Managed Agents update)

All dall-e-2 and dall-e-3 snapshots are shut down as of May 12, 2026. Migrate any remaining image-generation calls to gpt-image-1 (frontier quality) or gpt-image-1-mini (cost/latency-optimized). Anything still pointing at DALL·E snapshots will now error — audit production code, marketing pipelines, and any user-facing image features. (OpenAI deprecations)

Anthropic announced Claude Platform on AWS — native AWS billing and IAM, with Messages, Files, Batches, Managed Agents, Skills, code execution, and tool use all supported. Big unlock for AWS-native shops: consolidated invoicing, IAM-scoped Claude access, and no separate vendor onboarding. Same Claude models (Sonnet 4.6, Opus 4.6/4.7) under the hood. Worth evaluating if you've been routing around procurement to use Anthropic direct. (Claude API release notes)

New return_token_budget parameter on the Responses API web search tool allows GPT-5+ to spend more tokens reasoning over web search results before returning an answer. Directly relevant for deep-research agents and any RAG-over-web pipeline where the model was previously cutting reasoning short. Tune higher for harder multi-source questions; expect higher latency and token cost in exchange for answer quality. (OpenAI API changelog)

Three new realtime/voice models for streaming speech agents. Realtime 2 is a speech-to-speech model with configurable reasoning. Realtime Translate handles streaming speech translation. Realtime Whisper is streaming speech-to-text. Together they round out the voice-agent stack — directly relevant for call-center bots, meeting assistants, and real-time copilots where latency and translation quality matter. (OpenAI API changelog)

OpenAI is winding down self-serve fine-tuning. Effective May 7, 2026: orgs that have never run fine-tuning can no longer create training jobs. Effective July 2, 2026: orgs that haven't run inference on a fine-tuned model in the past 60 days can no longer create new jobs. Hard cutoff Jan 6, 2027: even active customers can't create new fine-tuning jobs (inference continues until the base model is deprecated). If you depend on FT for task-specific coding/RAG models, schedule periodic inference to keep eligibility, or move to alternative customization (RFT, prompting, distillation) before July 2. (OpenAI deprecations)

gemini-3.1-flash-lite is now GA — Google's speed/scale/cost workhorse for high-throughput RAG, embedding pipelines, and agent tool-routing where cost and latency dominate over peak quality. Migration timeline: gemini-3.1-flash-lite-preview deprecates May 11 and shuts down May 25. If you're still on the preview SKU, switch to the GA name within ~2 weeks to avoid outage. (Gemini API changelog)

A small but useful plugin for Codex that surfaces Platform access and API setup guidance directly inside the coding workflow. No API breaking changes — just smoother onboarding for devs building against OpenAI inside Codex. (OpenAI API changelog)

Anthropic announced significant API rate limit increases for Claude Opus models, effective immediately. Not a model spec or pricing change, but it directly reduces throttling pain for high-volume Opus workloads (long-context coding agents, document understanding, agent orchestration). Doesn't change Sonnet 4.6 limits. (Anthropic news)

Gemini Interactions API is renaming outputs to steps and changing response_format configuration. The new schema becomes default on May 20, 2026, and the legacy schema is removed entirely on June 6, 2026. This is a breaking change for any agent framework built on the Interactions API — plan migration tests before May 20 and confirm production is updated before June 6 to avoid outages. (Gemini API changelog)

The OpenAI Agents SDK has shipped a TypeScript implementation alongside the existing Python SDK, including sandbox agents and the open-source harness. Significant for Node/TS-native agent stacks — reduces friction for web apps integrating tool-using agents and standardizes harnessing/testing patterns. If you maintain agent infrastructure in TypeScript, evaluate parity with your Python setup before migrating production workloads. (OpenAI API changelog)

New moving-target alias that tracks the current Instant model used in ChatGPT. Useful for evaluating ChatGPT-style behavior and quick regression checks, but explicitly NOT for production — the underlying snapshot rotates without notice. For production, keep pinning specific snapshots (or use GPT-5.5 for the recommended frontier line). (OpenAI API changelog)

Gemini API File Search now supports image embedding and retrieval via gemini-embedding-2 — first-party image RAG without rolling your own pipeline. Grounding metadata also improved with new media_id and page_numbers fields, enabling proper image- and page-level citations in attributed-answer UIs. If you're using Gemini File Search for RAG, consider indexing images and surfacing the new metadata as inline citations. (Gemini API changelog)

Event-driven completion notifications for the Batch API and other long-running operations — no more polling. Cuts idle compute and simplifies orchestration for async workloads (batch evals, offline RAG indexing, large multimodal processing). Worth migrating any polling-based batch workflows to webhooks. (Gemini API changelog)

GPT-5.5 (snapshot gpt-5.5-2026-04-23) is now available in the Responses and Chat Completions APIs. 1,050,000 token context window, 128K max output, Dec 1 2025 knowledge cutoff. Pricing: $5/1M input, $0.50 cached, $30/1M output (standard, <272K context). Prompts >272K input tokens are billed at 2x input / 1.5x output for the full session. Reasoning effort supports none/low/medium/high/xhigh. Marketed as OpenAI's strongest model for spreadsheet creation, polished frontend code, hard math, document understanding, tool use, and multi-source research. Roughly 2x the price of GPT-5.4 — community reports suggest output efficiency gains may partially offset the increase. Plus tier users hit token limits in minutes on heavy workloads, so budget carefully. (Introducing GPT-5.5, API docs)

GPT-5.5 Pro is available in the Responses API (not Chat Completions). Pricing: $30/1M input, $180/1M output (standard, <272K context); $60/$270 per 1M for long context. Same 1M context window as GPT-5.5. Use case: high-stakes reasoning, agent orchestration, and tasks where accuracy beats latency or cost. At $180/1M output, reserve for jobs where a wrong answer is more expensive than the model. Same price as GPT-5.4 Pro — the upgrade is free at the Pro tier. (OpenAI pricing)

gemini-embedding-2 is now generally available via the Gemini API. Production-ready multimodal embeddings for RAG, semantic search, and classification workloads — no longer preview-only.

OpenAI added new deprecation entries for legacy model snapshots. Check the deprecations tab and update any pinned snapshot references in production code.

deep-research-preview-04-2026 and deep-research-max-preview-04-2026 released. New capabilities: collaborative planning, visualization support, MCP server integration, and File Search. Designed for multi-step research agents that need tool access and structured outputs.

OpenAI launched gpt-image-2 for /v1/images/generations and /v1/images/edits. Features token-based image pricing and Batch API support (50% discount). Replaces prior image generation endpoints for production use.

Anthropic released Claude Opus 4.7 via the Claude Platform. Successor to Opus 4.6 with improved reasoning and coding performance.

Google released Gemini 3.1 Flash TTS Preview, adding native text-to-speech capabilities to the Gemini API. Relevant for voice-first agents and multimodal applications.

OpenAI Agents SDK now supports running agents in controlled sandboxes for safer code execution, plus explicit controls over when memories are created and where they’re stored. Key for compliance, per-tenant isolation, and secure coding agents.

gemini-robotics-er-1.6-preview released for embodied reasoning in robotics. gemini-robotics-er-1.5-preview shut down April 30, 2026 at 9AM PST — production workloads must use 1.6.

Anthropic launched a gated research preview of Claude Mythos Preview for defensive cybersecurity via Project Glasswing. Pricing for participants: $25/MTok input, $125/MTok output. Available via Claude API, AWS Bedrock, Vertex AI, and Microsoft Foundry.

Anthropic expanded its partnership with Google and Broadcom for multiple gigawatts of next-generation TPU compute starting in 2027. Signals massive infrastructure buildout for future Claude models.

Gemma 4 open-weight models are now accessible through the Gemini API, expanding options for self-hosted and fine-tuned deployments.

Google introduced Flex and Priority inference tiers for Gemini API, giving developers control over latency vs. cost tradeoffs in production workloads.

Veo 3.1 Lite Preview (veo-3.1-lite-generate-preview) launched for cost-efficient video generation. gemini-2.5-flash-lite-preview-09-2025 has been shut down — migrate to gemini-3.1-flash-lite-preview.

gemini-3.1-flash-live-preview enables audio-to-audio real-time conversations via the Live API. Key for agent-to-agent (A2A) and voice-first applications.

Lyria 3 models generate music from text and image prompts. Expands Gemini's multimodal capabilities into audio generation.

OpenAI announced deprecation of Sora 2 model aliases/snapshots and the entire Videos API. Shutdown date: September 24, 2026. Plan migrations now if using Sora in production.

Google introduced Prepay and Postpay billing plans for AI Studio, giving developers more flexible payment options for Gemini API usage.

Gemini API now supports using Gemini’s built-in tools alongside custom function calling tools in a single request; Google Maps grounding is now supported for Gemini 3 models going forward.

GPT-5.4 mini brings GPT-5.4-class capability to high-volume workloads with tool search, computer use, and compaction; GPT-5.4 nano is optimized for simple, high-volume tasks and supports compaction only.

Google introduced new usage tier structure with billing account spend caps for Gemini API. Better control for production workloads.

Added project-level spend caps for Gemini API billing in AI Studio, enabling tighter cost control for production experimentation.

First multimodal embedding model supporting text, image, video, audio, and PDF inputs in a unified embedding space. Major for multimodal RAG and unified retrieval.

Deprecated. Will be shut down March 31, 2026. Migrate to gemini-3.1-flash-lite-preview.

Gemini 3 Pro Preview was shut down; the gemini-3-pro-preview alias now points to gemini-3.1-pro-preview.

Sora API expanded with character references, longer generations up to 20s, 1080p for sora-2-pro (billed at $0.70/sec), Batch API support, and a new /v1/videos/edits endpoint. /v1/videos/{video_id}/remix will be deprecated in 6 months.

1.05M context, configurable 5-level reasoning, native computer-use API. First general-purpose model with built-in computer use. Priced at $2.50/$15 per 1M tokens. GPT-5.4 Pro at $30/$180.

gpt-4-0314, gpt-4-1106-preview, gpt-4-0125-preview, gpt-4-turbo-preview all sunset March 26, 2026. Requests auto-route to gpt-4.1.

First Flash-Lite model in the Gemini 3 series. 1M context, $0.25/$1.50 pricing. Ultra-efficient for high-volume, cost-sensitive workloads.

Nano Banana 2 (gemini-3.1-flash-image-preview) optimized for speed. Gemini 3 Pro Preview shut down March 9.

Frontier reasoning model. 1M context, $2/$12 per 1M tokens. Medium thinking level for cost/speed/performance balance. Enhanced agentic coding and multimodal analysis.

Most capable Sonnet yet. 1M context (beta), 128K max output, improved coding + computer use + long-context reasoning. $3/$15 per 1M tokens. Approaches Opus intelligence at Sonnet pricing.

GPT-4o, GPT-4.1, GPT-4.1 mini, o4-mini, and GPT-5 (Instant/Thinking) removed from ChatGPT. Still available in API. GPT-4o in Custom GPTs until March 31.

Specialized reasoning mode upgraded to solve modern science, research, and engineering challenges. Available to Google AI Ultra subscribers.