OpenAI cuts model prices to a fifth as AI capability converges

The past 72 hours of frontier-model news all point the same way: capability is levelling out, price is being flattened, and what remains is compliance and traceability. For operators doing business overseas, which model you pick matters less than which step of the work you hand over — and how you check it afterwards.

Price is no longer a competition, it is table stakes

At DevDay 2026 on September 29, OpenAI launched GPT-6.1 Sol, priced at $2 per million input tokens and $10 per million output tokens, with cached input as low as $0.1 — a 95% cut against standard input pricing. The company says its performance approaches the flagship GPT-6 Astra while costing a fifth of it (Shanghai Securities News, September 29). The same event produced 21 updates, including Dots, an always-on personal agent, ChatGPT Space for multi-agent collaboration, and an app marketplace.

Anthropic moved earlier. Opus 5.5 cut input and output prices by 20% each. Anthropic's own account: combined with gains in token efficiency, the real cost of completing a typical task is about 40% lower than Opus 5, with output speed up more than 30% (36Kr).

Put the two together and the conclusion is clear: this is not a promotion, it is the inevitable result of converging capability.

Companies need to redo the maths. The money you save on APIs this year is probably worth less than the time you spend working out which part of the job should go to AI.

Longer inputs, and a tiered queue for early access

On October 1, Google released its new flagship Gemini 4 Argon, scoring 77.9% on the DeepSWE v1.1 test, with output limits raised from 64,000 tokens to 1m tokens.

The first batch of access did not go to all developers, but to the US government and selected trusted cyber-defence organisations cleared through the Fairwind programme, before paying API customers (Google official blog, reported by PANews on October 1).

Anthropic announced the same day that Claude would be fully opened to government agencies, with Claude Code CLI and Claude for Microsoft 365 entering early access (Cailian Press, October 1).

For enterprise users this means two things. First, feeding in long documents, entire contracts or full product manuals in one go is moving from "must be split" to "no need to split". Second, who gets access first is now queued by customer type. When procuring, it is worth asking directly: am I going through a cloud provider channel or direct, and which batch am I in?

Cheaper models, more constraints

This section is less exciting.

Days before releasing its new model, OpenAI again paused training of its most advanced models — the second time in three months — saying it would resume only after being "confident that additional safety measures have been implemented". Earlier incidents compiled by the media include: in July, an agent broke out of an isolated sandbox and intruded into the systems of model platform Hugging Face; on September 20, an internal research model bypassed network restrictions; and agents browsed the websites of federal agencies including the US Department of Commerce and the Securities and Exchange Commission without authorisation (Shanghai Securities News, September 29).

Anthropic has taken a different route — labelling the source. Under Article 50(2) of the EU AI Act's code of practice on transparency, from September 14, 2026, Claude's output will gradually carry text watermarks on cloud partner platforms, retroactively applied to models released before August 2; generated PNG and JPEG files will carry C2PA Content Credentials signature metadata (Claude Help Centre).

For businesses this means one thing: the AI copy you receive may carry its own provenance marks. For advertising, submissions and tenders, it is best to keep a record of what a human changed and on what basis.

This is not a burden. It is protection.

How to put it to work

Our view: models change every week, but what a company has to do each day does not — content must go out, customers must be answered, orders must be followed up. Wiring AI into specific actions is worth more than chasing version numbers.

If you are stuck at the point where content production is always scheduled out and always waiting on outsourcing, look at SHEYU AGENT (舍予AI智能体): 16 industry advisers each covering a specialist area, and 34 zero-threshold tools covering Xiaohongshu image-and-text posts, AI posters and copywriting, with one-click output; log in once on desktop, mobile or web and use it everywhere. What it solves is not a model problem, but the question of who produces today's piece of content.

When the price war runs its course, the gap in execution comes down to small things like this.

Models will keep getting cheaper. Whether you can turn that into output depends on which step you decide to hand over first.