SHEYU.AIAs GPT-4o Retires and GPT-5.5 Winds Down, Enterprises Need Three Tables Before Picking a Model
Old models are retiring, introductory pricing is expiring and usage policies are changing. Three tables — retirement, cost window and compliance — should guide enterprise model choices.
Recent public announcements point to three things at once: old models leaving service, pricing windows closing and usage policies being rewritten. None of them changes whether a company should use AI, but all of them will shape January's bill and the year-end compliance review.
We think any enterprise choosing a large model should first prepare three tables: a retirement table, a cost-window table and a compliance-boundary table.
1. Retirement table: chat and API are two separate ledgers
Start with the conclusion: the old models inside ChatGPT have already been cleared twice, but change on the API side is far smaller than in the chat product.
OpenAI's Help Center version notes are explicit. On 13 February 2026, GPT-4o, GPT-4.1, GPT-4.1 mini and o4-mini, together with the previously announced retirement of GPT-5 (Instant and Thinking), were all discontinued from ChatGPT. The same notes contain a line that is often overlooked: the API currently has no changes.
The chat-side turnover continues. A model comparison summary shows that on 9 July 2026 GPT-5.6 opened fully, splitting a generation into three tiers for the first time: flagship Sol, balanced Terra and lightweight Luna. The main chat model on paid plans became 5.6 Sol, while the default model for free-tier chat switched to 5.6 Luna on 6 August 2026. GPT-5.5 has also been scheduled to retire on 14 October — that is, this month.
Earlier still, OpenAI released GPT-6 Astra on 3 September 2026, opened in batches and called "GPT-6 Pro" inside ChatGPT, aimed at Pro plans above $100 and at Business/Enterprise. A 23 September update on OpenAI's news hub noted that Airbnb is expanding its use of GPT-6 Astra.
How to use this table: create one row for every model you run, with only three columns — where it runs now, its official retirement date and the replacement.
An example. Suppose an overseas-facing team's customer service relies on the chat version for human-assisted replies during the day. After the February clear-out, the old model options simply disappeared and the workflow had to be rebuilt. But if its back-end automatic replies run through the API, it was barely affected. Separating these two paths removes half the panic.
2. Cost-window table: introductory pricing ends this year
Google states in its Gemini API documentation that Gemini 3.8 Flash, 3.7 Flash and 3.6 Flash enjoy introductory pricing on Google AI Studio and Gemini Enterprise Agent Platform until 31 December 2026, with standard pricing applying from January 2027.
For operators the meaning is direct: this is a window for testing with real business volume, not demo data.
What to do: pick a task with stable historical data, such as product descriptions or email replies, run a full month of real data through it, and record only two numbers — volume processed and the human rework rate. Finish before year-end and you will have something to compare.
On the image side, one item is worth noting. Google AI for Developers' Gemini API documentation lists Gemini Nano Banana 2.1: input accepts text, images, video and PDF, output is images and text, with an input token limit of 131,072 and an output limit of 32,768. It supports image generation and search grounding, but not function calling. Teams producing e-commerce hero images and posters should not watch text models alone.
3. Compliance-boundary table: usage policy takes effect on 12 November
Anthropic published its 2026 version of the Usage Policy on 8 October 2026, effective 12 November 2026.
The official line is that most changes clarify existing rules. The stated reason: over the past year Claude has taken on longer, more independent work, and the policy needs to set out the boundaries that match these new capabilities. At the same time, Anthropic announced a three-year, $150m commitment to support the Genesis Mission and launched Anthropic Cyber Mission.
For enterprises: treat the policy as a checklist and verify one thing only — whether your automated tasks are "long-running with little human intervention".
Suppose you let a model autonomously gather material, rewrite and publish for several hours with no human review in between. Such workflows should have their boundaries reconfirmed against the wording of the new policy. Do this check now, rather than reworking processes after 12 November.
4. Capability news is not a reason to migrate
On 7 October, a report by New Scientist, republished by Sina Finance, said OpenAI released 722 mathematics papers at once, including proofs and disproofs of various mathematical propositions. The company did not name the model used, saying only that it was an "internal frontier model".
The academic reaction is worth noting for enterprises. Francis Johnson of University College London spent 25 years researching the Wolf D(2) problem and wrote two books on it, and in the end it was he who gave up. He said AI cracking it so quickly was surprising.
Kevin Buzzard of Imperial College London cautioned that 30 of the results fall in number theory, his field, and only seven appear to be of outstanding quality, with only one passing Lean formal verification — Lean being a computer tool that can verify mathematical derivations without error.
Put the two remarks together: capability is genuinely advancing, but "released" and "verified" are two different things. The reason to switch models should be better results on your own evaluation set, not a loud headline.
How to put this into practice
Back to the most practical question: most operating teams do not have the bandwidth to track every monthly retirement schedule and pricing adjustment. This kind of work is better handled at the tool layer.
SHEYU AGENT (舍予AI智能体) brings together 16 industry advisers and 34 zero-threshold tools. Xiaohongshu image-and-text posts, AI posters and copywriting can all be produced in one click, and a single login works across desktop, mobile and web. Content teams no longer need to hard-wire their workflow to one chat model, which removes one more reason to chase retirement tables and rewrite configurations.
One judgement to take away: models will keep changing, but the three tables — who is using them, how much they cost and where the boundaries lie — are assets you should maintain yourself for the long term.