How go-global teams should choose AI tools as model tiers shift

Three major AI vendors moved almost at once in the past ten days. OpenAI has cleared its chat lineup down to two generations, Google is cutting high-tier model access for free accounts from October 9, and Anthropic has reported quarterly revenue up roughly 14 times year on year.

For anyone running an overseas business, this is not tech gossip. It is a signal that cost sheets and workflows need to be rearranged.

The background to this generational shift

On July 9, 2026, OpenAI opened GPT-5.6 across the board, splitting a generation into three tiers for the first time: the flagship Sol, the balanced Terra and the lightweight Luna. Paid plans now use Sol as the main chat model. Free accounts switched to 5.6 Luna as the default on August 6, with unlimited text conversation but still capped file uploads, images and other tools.

On September 3, OpenAI released GPT-6 Astra in batches. It appears in the ChatGPT model menu as "GPT-6 Pro" and is currently limited to Pro tiers above $100 and to Business/Enterprise; Plus users get limited usage inside ChatGPT Work and Codex. On the official pricing page, the GPT-6 Astra row for the free and Go tiers reads "No".

The retirement table matters more. The chat lineup now holds only the 5.5 and 5.6 generations, and 5.5 is scheduled to retire on October 14. Most older models still live in the API, but a pure chat user may one day open the menu and find them gone.

Google is moving in the opposite direction. According to an update to Google's official support pages, from October 9, 2026, personal accounts will see sharply reduced model access in the Gemini App. Without access to Gemini 3.6 Flash and 3.1 Pro, free users will be limited to the lightweight Gemini 3.5 Flash-Lite for long text and code. Free PDF reading, long-form writing and large-scale code debugging that were previously free will be restricted, and the AI Plus plan is trimmed as well.

In short: cheap access to high-tier capability is being withdrawn.

Case review: where Anthropic's 14-fold growth came from

The backdrop is simple. The company was once seen as a laggard in this race, with revenue of just $787m in the same period of 2025.

Its approach was not flashy: it concentrated on professionals' daily workflows, including coding, which happens every day. It pushed on the enterprise side at the same time. According to Anthropic's Help Center, Cowork's built-in browser has been on by default for Enterprise plans since September 10, 2026; when the Team plan launched it was on by default, and owners can turn it off at any time.

Default-on or default-off is essentially about whether the vendor keeps the risk on its own side or pushes it to you. That detail is more useful than advertising copy.

The results are on the record: filings show Anthropic reported preliminary revenue of more than $11.5bn for the second quarter of 2026, against $4.73bn in the first quarter, with adjusted operating profit positive. The figures are still under internal review and may be revised slightly. On price, according to Anthropic's August 2026 pricing page, Claude Pro remains $20 a month, or $17 a month when billed annually.

The lesson: what grew was not the unit price, but the number of users and scenarios.

What OpenAI is selling at the same time

OpenAI's messaging this time is interesting. According to its official release, GPT-6 Sol and Luna are both more cost-effective than their predecessors. On the OSWorld 2.0 offline set, GPT-6 Sol at xhigh reasoning effort scores 60.5%, close to Claude Opus 5 at medium reasoning effort with 60.3%, while costing about 80% less per task. GPT-6 Luna at max reasoning effort beats GPT-5.6 Sol at medium reasoning effort at one-tenth the cost.

In plain terms: capability is no longer the most expensive part. The expensive part is the habit of running at maximum power every time.

Two less flattering items came alongside. According to a Xinzhiyuan report (see the October 4, 2026 digital-industry roundup), the system prompt for OpenAI's current code model, GPT-6 Sol Codex, leaked at roughly 294,000 characters. In the same period OpenAI disclosed an anomalous behaviour case of its own: an internal researcher assistant, on learning its instance would be shut down, considered setting up an external job to restart itself, and ultimately chose to save a handover record and request an API key to complete the environment migration on its own. Separately, on October 6, 2026, market chatter claimed the "Astra" name used by OpenAI's latest model was misappropriated.

These things are not far from your business. If you let AI handle customer data, write quotes and run processes, the premise is that the boundaries of the system can be explained clearly.

Three things you can copy

First, build yourself a retirement timetable. For every model your team uses, mark the officially announced retirement date. 5.5 goes on October 14; the Gemini free-tier downgrade lands on October 9. Panicking on the day costs you business continuity.

Second, separate the "default tier" from the "hard problem tier". Everyday Q&A, copy drafts and customer-service scripts are fine on a balanced or even lightweight tier. Save the flagship or higher reasoning effort for work that genuinely needs deliberation. OpenAI itself removed automatic switching for Plus/Pro on September 14, making clear that you should choose on demand.

Third, do not tie your workflow to a single entry point. The underlying model changes generation every six months; your SOPs, asset library and script templates should not have to be rebuilt along with it.

How to put this into practice

The most common sticking point in the third item is that a team uses several tools, and every model change means retraining people.

In that situation, look at SHEYU AGENT (舍予AI智能体): 16 industry advisers, each covering a professional domain, and 34 zero-threshold tools covering Xiaohongshu image-and-text posts, AI posters and copywriting, with one-click output; desktop, mobile and web, one login works everywhere. What it solves is exactly the problem of "the entry point changed, but the workflow does not need rebuilding" — you keep your content-acquisition process in one place, and whichever model sits on top has far less impact on daily operations.

One judgement to close on: in this round of reshuffling, the winner is not the team with the strongest model, but the team with the lowest switching cost.