A Fivefold Token Price Drop Doesn't Pay for AI Projects

In its AI Twitter review, Latent Space describes the launch of GPT-6.1 Sol and Sonnet 5.5 as shifting the cost-quality frontier. The figures confirmed by the review are: Sol costs $2/$10 per million input/output tokens, while Astra costs $10/$50.

The same source reports that Sol outperforms GPT-6 Sol by 6.4 points.

It is unclear from the fragment we have which benchmark was used.

That is why we compare quality only after running our own tests on our own data, and do not trust figures from a tweet.

In our estimate, tokens account for a small share of the total cost of ownership in a typical project.

Most of the money goes toward data preparation, integration with accounting systems, response quality control, and security.

A cheaper model removes the most visible line item from the budget; the others remain unchanged.

Cheap output tokens create a separate effect.

When generation became five times cheaper, teams started asking the model to write more: detailed reports, longer descriptions, and multi-step reasoning.

The volume grows, while the user’s time to result may remain unchanged.

Measure TTU (time to use), meaning the time from the request to a usable result.

The number of pages generated says nothing about the result.

Prices are falling, new models are released every few weeks, and companies need their own layer between the model and the business process.

It solves four tasks. Routing.

The request goes to a model suited to the task in price and quality: classification to a cheaper model, legal analysis to a stronger one. In KT.Team’s LLM & Security Gateway, this is one policy-driven entry point.

Switching providers—OpenAI, Anthropic Claude, Google Gemini, or Sber GigaChat—comes down to editing one configuration line; the application code remains unchanged. Context.

Only part of the 20,000 input tokens in the example above is useful. RAG selects the necessary fragments, reducing the input severalfold and improving accuracy.

Such savings exceed any provider discount. Access to systems.

A model without access to 1C, Bitrix, Pimcore, or Odoo can only reason.

Through MCP, it reads stock levels, creates an order, and updates the product card.

This is where a result visible in the report appears. Security.

A cheap token encourages sending more data to the model.

The gateway masks personal data and trade secrets before data leaves the internal environment and maintains an audit trail.

First, recalculate the register of deferred AI ideas.

Some of them dropped out because of the price, so it is worth checking which ones have returned to the payback range. Second, do not tie yourself to a single model through either a contract or code.

According to a Latent Space review, two releases shifted the price–quality frontier in a single cycle, and the next shift is coming soon. Third, require one metric visible in the P&L from every pilot.

Without it, the pilot remains a show, and the person who commissioned it pays dearly.