Best OpenAI migration feed: OpenAI API changelog. Official OpenAI release log for launches, pricing, and renamed features. Use it to catch live changes like pricing moves, feature renames, and model guidance shifts before rewriting a workflow.
Best Anthropic retirement source: Claude model deprecations. Official Anthropic retirement schedule and replacements. Use it before pinning Claude model IDs or teaching teams to rely on an older snapshot.
Best Gemini shutdown source: Gemini deprecations. Official Gemini API deprecation schedule. Use it when Google preview and Flash variants are moving quickly and you need the replacement path from the primary source.
Best Anthropic upgrade checklist: Anthropic migration guide. Official Claude migration guide covering current model deltas and breaking changes. Use it when you are changing Claude model IDs and need to re-check effort, tool behavior, prompt caching, or system-message differences.
Best Perplexity platform migration source: Perplexity migrate from Sonar to Agent API. Official Perplexity migration guide from Sonar to the Agent API. Use it when old Perplexity integrations still assume chat completions instead of the newer agent-first API.
Best DeepSeek migration source: DeepSeek V4 Preview release. Official DeepSeek V4 release note and alias migration notice. Use it when a provider is moving old aliases to new V4 families and you want the exact date and replacement models.
Release notes are part of the learning stack now
If you build with frontier models, release notes and deprecation pages are no longer optional reading. Model aliases change, pricing changes, feature flags move, APIs get renamed, and a previously good course can become subtly wrong if it assumes an older default.
As of Saturday, September 5, 2026, three model changes deserve fresh migration tests. GPT-6 Astra introduces async tool calls, mid-turn steering, mutable reasoning effort, new long-context behavior, and premium pricing, so compare it with GPT-5.6 on completed-task cost rather than tokens alone. Gemini 3.8 Flash is now GA for long-horizon coding and autonomous-agent workloads, while Lyria 3.5 adds new preview model IDs for clip and full-song generation. Claude Fable 5.1 remains Anthropic's newest demanding long-horizon option with model-specific prompting and migration guidance. The workflow layer is moving too: test steering behavior, hook side effects, reusable recorded skills, WebMCP permissions, and cross-session memory with the same rigor as model swaps. The rest of the current stack still matters: DeepSeek V4, Grok 4.6, Qwen3.8 and Qwen Code, Mistral Agentic Search and OCR, Cohere Parse and its ATE agent-tool dataset, Perplexity's Search and Agent APIs, Hugging Face's retrieval and agent guides, and OpenRouter's Agent SDK, Ori Eval, Ori Harness, routing, and live catalogs. For every change, pin a candidate model, run a representative eval set, compare tool behavior and total task cost, and only then update production defaults.
What to bookmark by default
Keep one model catalog page and one release or deprecation page open for each provider you actually use. For OpenAI that usually means the model guide plus the changelog or deprecations page. For Anthropic it means the Claude models overview plus release notes or model deprecations. For Gemini it means the models page plus the changelog and deprecations page. For DeepSeek it means pricing or updates plus the live model list.
When you use a routing layer such as OpenRouter, keep both the provider docs and the routed catalog open. The routed view helps you see availability, aliases, and market movement. The provider view tells you what the model is supposed to do and what changed.
How to keep apps and coding agents current
Do not rely on `latest` or family aliases blindly unless you also have evals and regression checks. A safer workflow is to watch release notes, shortlist the affected models, run a small eval set, compare latency and cost, and then update the default for your app, coding agent, or internal tool.
This matters even more for coding agents and deep-research tools because they depend on tool use, output structure, and long-running behavior. A migration that looks small on a model card can still change how an agent plans, reasons, or uses budget.
Recommended courses and resources
-
The AI Engineer Roadmap
Free tutorial · Matt Pocock · Beginner to intermediate
You want a guided path through core AI concepts, model selection, the AI engineering mindset, evals, and techniques for improving LLM-powered apps.
-
Claude Sonnet 5.5 explained
Model migration explainer · Anthropic · Intermediate to advanced
Anthropic released Claude Sonnet 5.5 on September 28, 2026. Use this guide to understand the efficiency claims, breaking API changes, and a safe migration test from Sonnet 5.
-
Ember-1 explained
Reasoning-efficient coding model explainer · Fireworks AI · Intermediate to advanced
Fireworks released Ember-1 on September 23, 2026 as a Kimi K3-based model trained to use fewer reasoning tokens. Use this explainer to evaluate the quality, cost, latency, and context-growth claims on your own agent workload.
-
Grok 4.7 explained
Frontier coding model launch explainer · SpaceXAI · Intermediate to advanced
SpaceXAI released Grok 4.7 on September 21, 2026 for coding, agentic tasks, and knowledge work. Use this guide to compare its task reliability and total cost with your current model.
-
Union Alpha revealed as Pareto 26.9
Multi-model routing launch explainer · Unbiased · Intermediate to advanced
Union Alpha was revealed as Unbiased's Pareto 26.9 on September 17, 2026: a hosted system that routes work across several models and returns one checked answer.
Common questions
Which AI release notes should I actually follow?
Follow the official changelog or release-note page for every provider you actively use, plus its deprecations or shutdown page if available. OpenAI, Anthropic, Gemini, DeepSeek, Mistral, xAI, and OpenRouter all publish enough primary-source change data now that third-party summaries should be secondary.
Where can I track model deprecations and migrations across providers?
Use OpenAI deprecations, Claude model deprecations, Gemini deprecations, DeepSeek updates and pricing, Mistral changelog, and xAI release notes. Pair those with the provider's model catalog so you can see both replacements and current IDs.
How do I keep coding agents and AI apps current when model aliases change?
Watch the release notes, keep evals for your core workflows, compare the replacement model on your own prompts, and only then update the default model or alias. Treat migrations as engineering changes, not just documentation housekeeping.
