A New AI Model Every Month: What Companies Should Build

According to Latent Space, Anthropic's annual recurring revenue from subscriptions (ARR) was $47 billion at the time of its largest funding round in May. It exceeded $65 billion in July, and the publication expects up to $100 billion by the end of 2026.

The same source mentions a target IPO valuation of $2 trillion.

These figures come from an industry roundup; auditors have not verified them.

They show the scale: companies pay substantial sums for models and rapidly increase their spending. September 24–25, AINews recorded a wave of releases: Claude Opus 5.5, the GPT-6 family, Gemini 3.8 Flash, and Xiaomi MiMo-V2.6-Pro.

According to AINews, Opus 5.5 topped SimpleBench with a score of 88.4%. SimpleBench consists of common-sense and spatial-reasoning questions: people solve them easily, while models struggled with them for a long time.

Our hypothesis is that improvement on this test means the agent can be assigned more tasks with implicit conditions that previously required a person.

This should be tested on your own tasks.

Powerful models are released roughly once a month.

For an internal AI project, this means the model chosen at the start will become outdated before the project reaches production.

A company that invests in the four layers below changes the model through configuration and a test run, without rewriting applications. ### Context

The model receives company data through RAG: searches across documents, regulations, 1C, and the PIM catalog.

These data, their completeness, and their relevance determine the quality of the response. ###

MCP (Model Context Protocol) tools are an open standard that lets an agent work with systems: create an order in 1C, update an item card in Pimcore, and find a document in EDI.

The team writes the connector once, and it works with any model that supports the protocol. ###

The LLM & Security Gateway sits between employees and models.

It masks personal data, logs requests, and routes them: sensitive tasks go to GigaChat, YandexGPT, or Qwen within the company's environment, while complex reasoning goes to Claude.

An engineer connects the new model in the gateway, and the applications work as before. ### Evaluation

A set of 50–100 real company tasks with benchmark answers.

The new model is run through it in one evening, and the decision to switch is based on the numbers, without a committee.

This material was prepared by our news-machine pipeline.

Every morning, it collects industry sources, discards outdated and inaccessible material, identifies strong signals, and writes a draft using Claude.

Checks come next: text-quality gates, source-code review before publication, dry-run mode by default, and a kill switch that removes content with one command.

We change the model inside the pipeline through configuration, without touching the checks or the interfaces between components.

We apply the same principle in PIM projects built on Pimcore and Akeneo.

The agent writes product card copy, organizes the assortment by taxonomy, and fills in filter facets. A content manager reviews the result before publication.

When a better model is released, we run it on the client's benchmark cards and switch if the metrics improve.

The value of AI for business is measured by time to usefulness (TTU): how many days pass from an idea to a moving metric.

The model race shortens this timeline for companies that have their context, tools, gateway, and own test ready.

Companies without these layers return to the comparison table with every release.

I would put model selection last in the plan: it can be reviewed monthly, while the infrastructure around the model is built once and used for years.