Blog / Model watch

How to Watch New Claude Models After Launch (Sep–Oct 2026)

2026-10-18cdnping editors5 min read

When a new Claude model ships, there is no need to switch right away. Watch it for a few days first. According to Anthropic's website, the new Opus version was released on September 22, the new Sonnet version on September 28, and the new Haiku version on October 7, about two weeks apart. This article does not benchmark quality. It covers how to tell whether a model is stable and fast after launch, and how to judge whether "the model got dumber" is real or just a feeling. Everything here is current as of October 2026.

#When were the three releases?

According to Anthropic's official pages:

All three official pages say the models are available through the Claude Platform as well as AWS, Google Cloud and Azure. The Haiku page also says it targets high-volume, cost-sensitive tasks such as summarization and classification, and that it is the first Haiku model with an adjustable effort setting. For pricing, model IDs and migration notes, rely on the official pages and docs; this article does not restate them. For the full list of official announcements, see the Anthropic news page.

Separately, OpenAI's Codex changelog shows that on October 14 an older model was retired for Codex sessions that sign in with ChatGPT. The entry says sessions that use an API key are not affected (see the Codex changelog). If the model options in Codex changed for you, check that entry first.

#Why do new models tend to have problems early on?

Official docs never promise that a new model will not be rate limited. Requests pile up right after launch, so rate limits, queuing and occasional errors are not surprising. Check the official docs for exact limits. We have compiled incident records for September and October that you can compare against: AI service incidents, Sep–Oct 2026: OpenAI and Claude outage statistics. For capacity errors on the Codex side, see Codex errors in October: "Selected model is at capacity".

So the first week is not necessarily representative of how the model behaves normally. Evaluation and migration decisions can wait a little.

#How do you observe a model after launch? Five checks

#1. Is the official endpoint reachable? Check the API probe

Open /en/status/anthropic to see the results of cdnping's probes of the official API endpoint, run from Cloudflare's edge every 5 minutes, along with 30 days of history. It shows whether the endpoint is reachable. It does not mean any particular model is available. For how to read incident levels on status pages, see How to read the OpenAI and Claude status pages.

#2. How is the model itself doing? Check /models

/en/models uses OpenRouter's public API to show the latest models from each lab, with availability, latency, throughput and quantization precision (such as fp8 and fp4) across inference providers. It also records listings, delistings, price changes and quantization changes. If a new model shows up there, compare providers: the same model can look very different on different providers. See How much can uptime differ across OpenRouter providers for the same model.

#3. What did the vendor change? Check /changes

/en/changes monitors the official changelogs, pricing pages and news pages of Claude, OpenAI and others, records additions and deletions line by line, and offers RSS. For price and model-list changes, it is more reliable than memory. For a fuller approach, see How to find out right away when LLM API prices change.

#4. How slow is your own network to the Claude API? Test on the homepage

On the homepage speed test you can measure the "first connection" (including DNS and TLS) and the "round trip" (the median of 3 requests after the connection is reused) to the Claude API. These two numbers reflect your network path and have nothing to do with how fast the model is. Time to first token also includes model inference; the three are explained in Breaking down AI API latency. If things feel slower after switching to a new model, check whether these two numbers changed before you suspect the model.

#5. Did the model really get dumber? Look at several sources

Personal impressions are easily skewed by prompts, context length and whatever task you happened to be doing that day. Some suggestions:

• Fix a few test prompts of your own and run them on both the old and new model, recording the date and parameters.
• Check whether the quantization precision on /models has changed; see Did the model get dumber, or did the provider change? Inference providers and fp8/fp4 quantization.
• Look at the third-party evaluation projects linked from the /models page (Livenerf, AI Stupid Level, Is It Nerfed?, BridgeBench Nerf Bench). cdnping only provides links and does not republish their data; the conclusions are theirs, and you should cross-check at least two of them.
• Go to /en/down/claude and see whether other users are reporting "dumber" behavior or rate limiting at the same time.

#A migration checklist you can follow

1. Read the official release page and migration notes first, and confirm whether the model ID and parameter changes affect your existing calls.
2. Switch in a test environment or on a small share of traffic. Do not replace everything at once.
3. Keep the old model's configuration so you can roll back if something breaks.
4. For the first three days, check /en/status/anthropic and /en/models once a day, and record error rates and latency.
5. If you set a custom API endpoint in your client, see Custom API endpoints for Claude Code, Codex and OpenCode.

When several models run side by side, you can route by task difficulty; see How to do multi-model routing.

#References

• Anthropic: Opus release page, Sonnet release page, Haiku release page, news page
• OpenAI: Codex changelog