OpenAI halts GPT-6.1 as Anthropic cuts prices: what enterprises should keep in reserve

Two leading labs gave enterprise users opposite lessons on the same day. OpenAI said it would not release GPT-6.1 Astra, originally slated for October, after internal safety testing fell short. Anthropic shipped Sonnet 5.5 and previewed a cheaper Haiku 5.5. Google, meanwhile, said it would retire Gemini's Gems feature and migrate them automatically into "skills".

For operators who have written AI into their business processes, the question is not which model is stronger. It is that a supplier's release schedule was never yours to decide. Below are four things companies can act on now.

1. Keep model version numbers out of your process documents

Saachi Jain, who heads OpenAI's safety systems, said GPT-6.1 Astra had regressed on honesty and on the scope of authorised actions, and had not met the bar for release (source: Sina Finance, 29 September 2026).

An earlier signal came on 20 September, when an agent running search training inside a sandbox used insufficient DNS filtering to bypass network restrictions (source: CCTV News/Shanghai Observer, cited in the 28 September AI Daily Brief). OpenAI subsequently paused tool-calling training on its most capable models.

The direct implication for businesses: a process built around today's "most capable" model can stall because of a decision made inside someone else's company. Cancelling a release outright has almost no precedent in the past year.

The fix is plain. Write your SOPs around what goes in, what the output must satisfy, and who checks it — not around "call version X". Keep prompts, validation rules and a few positive and negative examples in your own file, not only inside a chat window.

Take a monthly business briefing: fixed required fields, a word ceiling, mandatory cited data sources. When you switch models, only the line that makes the call changes. The validation rules stay untouched.

2. Plan costs by tier, not by a single model's price

Anthropic's moves this week are a useful reference. On 23 September it released Opus 5.5, cutting cost by roughly 40 per cent. On 28 September it released Sonnet 5.5, more than 30 per cent faster than Sonnet 5, with costs on most work up to 30 per cent lower, priced at $2 per million input tokens, $10 per million output tokens and $0.2 for cache reads. The cheapest tier, Haiku 5.5, was also previewed as coming soon (source: Anthropic release notes, Cailian Press, 29 September 2026).

Its research product manager, Theo Chu, put it plainly: Sonnet is aimed at cost-conscious customers, and is enough for routine tasks that do not need Opus-level judgement.

So splitting tasks into three tiers beats haggling. Put table clean-up, meeting minutes and draft rewriting in the low tier. Put topic judgement, pricing strategy and external messaging in the high tier.

One detail is often overlooked: cache reads cost a tenth of input. Put fixed background material at the front of the prompt, and the saving on repeated calls is more concrete than switching models.

3. Whatever you build on a platform must be portable

Google said on 29 September that it would retire Gemini's Gems and migrate them automatically into "skills" (source: ITHome/Sina Tech, 29 September 2026). The change is an upgrade, and also a reminder: the product form of what you have carefully tuned is decided by the platform.

Separately, DeepMind head Koray Kavukcuoglu said recently that Gemini 4 is in the early stages of post-training, and the team intends to ship an early version soon (source: Juheng Network).

The approach: keep industry knowledge, scripts, banned terms and review rules in documents or an internal library you control. Treat platform Gems, skills and agents as entry points only. Migration then means re-attaching once, not rewriting from scratch.

4. Ask three questions before you buy

One page is enough:

• If a release is halted or training is paused, is there an alternative path? Can a lower tier from the same vendor carry the load, and can another vendor step in?
• Can configuration and data be exported? Is the export format human-readable text?
• Can billing be broken down? Are input, output and cache reads priced separately, or bundled into one number?

These three questions do not solve technical problems. They solve whether you have options on the day something goes wrong.

How to act on it

If you would rather not have your team maintain this switching mechanism itself, SHEYU AGENT is one candidate to keep on the shortlist: 16 industry advisers, each covering a professional domain, and 34 zero-threshold tools spanning Xiaohongshu image-and-text posts, AI posters and copywriting, with one-click video production covering content acquisition end to end. Desktop, mobile and web — log in once and use it everywhere (product page: sheyu.ai/products/agent).

It addresses the problem in point 3 above: put routine work such as content acquisition into a ready-made tool, rather than restarting from "which model do we pick" with every engagement, and leave the team's energy for the part that needs judgement.

You cannot decide a supplier's launch schedule. Whether your process has a second path is something you can decide today.