Today, the billing for enterprise AI has truly changed—Claude's billing split, Google's knowledge foundation, and 80% of Gemini-SQL2

When "subscription for everything" ends, when knowledge finally has a unified format, when natural language can directly query databases—these three things happening in the same week is no coincidence.

Today is June 15. If you are an Anthropic subscriber, especially on the Pro or Enterprise plan, you have most likely already received that email — starting today, automated calls such as Agent SDK, claude -p, and GitHub Actions will no longer be counted under your subscription quota, but will be separated and billed at the standard API rate.

In short, the era of "unlimited subscription" has officially come to an end today.

But I don’t want to just talk about the bill. Because two other things happened this week—Google Cloud released a knowledge standardization format called OKF, and Google Research’s Gemini-SQL2 achieved 80% on a database query benchmark—putting these three together, you can see a very clear thread: Enterprise AI is shifting from "being able to run" to "running stably and running clearly". Bill splitting is cost observability, OKF is knowledge portability, and SQL2 is data accessibility. Three dimensions, pointing in the same direction.

1. Claude Billing Breakdown: The Days of a 175:1 Subsidy Ratio Are Over

Anthropic issued an announcement on May 14, but it only officially takes effect today. The core change can be summed up in one sentence: Automated calls are separated from the subscription and billed separately at API rates.

How exactly will it be independent? "Non-interactive" usage scenarios like Agent SDK, claude -p (pipeline mode), and GitHub Actions will no longer consume your subscription quota; instead, they will be billed directly based on API usage. Your original subscription quota will only cover your interactive use within the chat interface.

The automation quota for each package is as follows:

Note this Enterprise Standard — the quota is $0. In other words, if you are on this plan and do not upgrade, Agent SDK requests will directly fail, not be throttled, but fail. You must manually enable overage billing, otherwise all automated processes will stop.

I did the math. Previously, Pro users paid $20/month and actually received API-equivalent computing power worth between $300 and $600. Think about it — $20 for $300, a subsidy ratio of 15 to 1. And for heavy users — those running Agents as their primary tool — the subsidy ratio can even reach 175 to 1. This is real, hard cash being subsidized.

This matter is actually quite reasonable. In any "unlimited" business model, cross-subsidization between light and heavy users is unsustainable. Anthropic was previously using money from light users to subsidize heavy users, and now it is merely correcting this distorted incentive.

But that's not what I want to talk about. What I find most interesting isn't the billing itself, but this signal: "The 'subscribe and use freely' model has officially ended in the AI industry." From now on, every call will be charged, and every automated task will have a cost tag. For individual users, this might just mean spending a bit more money, but for business managers, it means that AI's operational cost has shifted from "a subscription fee" to "a variable that needs to be monitored and optimized."

On the same day, OpenAI announced that new enterprise customers can use Codex for free for two months. It's clearly an attempt to snatch up corporate clients who are hesitating due to Anthropic's price hike. This is quite interesting—while one side is reeling in the net, the other is casting it out.

2. OKF: Knowledge finally has a "USB interface"

Billing breakdown addresses the question of "where the money is spent." However, enterprise AI faces a more fundamental question — "where the knowledge is."

On June 12, Google Cloud released OKF v0.1, full name Open Knowledge Format. Simply put, it is a knowledge standardization format — using Markdown files with YAML frontmatter to store and describe enterprise knowledge.

When you hear "standardized format," you might think it's just another boring specification. But I've thought about it seriously, and the pain points it addresses are very real.

Think about the knowledge distribution in a typical enterprise: metadata is in the data catalog, process documentation is in the Wiki, business rules are written in code comments, and a lot of key information exists only in the minds of veteran employees. These things are scattered across seven or eight platforms, in inconsistent formats, unsearchable, and unreadable by AI.

What did OKF do? It unified these fragments into one format: YAML headers for structured metadata, Markdown body for readable content. Vendor-neutral, human-readable, machine-readable, and version-controllable with Git.

If you're familiar with the "LLM Wiki" concept proposed by Karpathy this year (which has already garnered over 5,000 stars on GitHub), you'll find that OKF follows the same lineage. The difference is that Karpathy proposed a convention, while Google turned that convention into a standard and also provided a supporting toolchain.

Google has released a complete set of things: a specification repository, a BigQuery enhancement Agent (a "producer" used to generate OKF files), an HTML visualization tool (a "consumer" used to read OKF files), and three example packages for GA4, Stack Overflow, and Bitcoin.

For CIOs, the significance is straightforward — knowledge is no longer locked within a single platform. Your Wiki can be migrated, your documents can be version-controlled alongside code, and your AI Agent has a standard knowledge input format. I’d call it the "USB interface of the enterprise knowledge bus" — previously, each device required a different driver to connect, but now it’s plug-and-play.

3. Gemini-SQL2: What Does 80% Mean

OKF solves the problem of "how to store knowledge," so what about "how to query data"? Coincidentally, also on June 12, Google Research released Gemini-SQL2.

This model has been specifically optimized based on Gemini 3.1 Pro to do one thing: convert natural language to SQL. On the BIRD benchmark, it achieved an execution accuracy of 80.04%, ranking first among single models and significantly outperforming the solutions from OpenAI and Anthropic.

What does the number 80% mean? BIRD is not a simple "SELECT * FROM table" test. It includes real business databases with complex table structures, field ambiguities, and nested business rules. To understand a question like "the three SKUs with the highest return rate in the East China region last quarter," you first need to know how the return table relates to the SKU table, what the definition of the East China region is, and whether the return rate metric is based on amount or quantity.

In plain terms, this is no longer "writing SQL"; it's "understanding the business."

There are three direct application scenarios I see:

• Self-service Reports—Business personnel can directly ask "How do sales figures compare across regions this month?" without needing to submit tickets and wait for the data team's schedule.
• Enterprise Search — Turning databases into conversational knowledge sources, complementing the document knowledge stored in OKF
• SaaS Data Q&A — Customers don’t need to learn your report interface; they can just ask directly

It becomes even more interesting when viewed alongside OKF. OKF addresses the storage and retrieval of unstructured knowledge, while SQL2 handles the querying and answering of structured data. One manages "documents," the other manages "data." Together, they truly complete the knowledge coverage for enterprise AI.

Three Clues, One Direction

Back to the beginning. Why did I say it’s not a coincidence that these three things happened in the same week?

Claude billing breakdown allows enterprises to see exactly where every penny of AI spending goes for the first time—this is cost observability. OKF provides a unified storage and transfer format for enterprise knowledge—this is knowledge portability. Gemini-SQL2 enables non-technical personnel to directly "query the database"—this is data accessibility.

These three dimensions together point to the same thing: Enterprise AI is moving from "demos that work" to "production-ready". And the prerequisite for production readiness is precisely that costs are clear, knowledge is findable, and data is queryable.

This is not a technological upgrade; it is the formation of infrastructure.

Advice for Business Managers

Finally, let me say something practical. If you are a CIO or a technical decision-maker in a company, considering these three things together, I suggest you do three things right now:

First, immediately audit your AI automation call costs. Claude's pricing took effect today, and it's only a matter of time before other providers follow suit. Pull out your Agent call frequency, token consumption, and peak usage periods, and calculate how much it would actually cost at API rates. If this number makes your heart race, it means you've been "using for free" under subsidies, and now it's time to do budget planning.

Second, start evaluating OKF or similar knowledge standardization solutions. There's no need to wait for OKF to mature to version 1.0; you can pilot it with a business unit right now. Organize your core business knowledge in the OKF format, and you'll find that the mere act of "organizing" can expose countless blind spots in your enterprise knowledge management. Once you have a knowledge base in a standard format, it will be much easier to switch AI vendors, change platforms, or implement RAG in the future.

Third, focus on the maturity of natural language data query capabilities. 80% of Gemini-SQL2's performance is based on lab data, which may drop by about 30% in production environments. But even with only 56% accuracy, it represents a qualitative leap for business teams that submit data tickets every day. I suggest you find a department with the most frequent reporting needs and conduct a small-scale proof of concept—let business personnel directly query data using natural language and see how many manual tickets can be replaced.

Three steps: first, calculate costs clearly; second, organize knowledge systematically; third, open up data smoothly. What follows will come naturally.