GPT-6.1 Sol in Copilot: Task Cost Depends on Agent Steps

GPT-6.1 Sol is available to all customers on Copilot Pro+, Max, Business, and Enterprise plans. GitHub offers it for agentic development and terminal work.

According to GitHub’s early tests, the model reliably completed tasks while using significantly fewer tokens and steps than models in the GPT-6 and GPT-5.6 families.

GitHub does not provide precise percentages, so companies will have to measure the effect on their own tasks themselves.

The key point in the release is the payment model: the model is billed according to the provider’s price list under usage-based billing.

Previously, expenses were capped by a fixed per-developer license fee; now there is no such ceiling.

The company pays for every token the agent consumes.

Starting July 1, managed-settings.json is available to all GitHub Enterprise Cloud customers.

The administrator centrally defines which models and plugins are allowed in VS Code and Copilot CLI.

Without such a policy, developers choose whichever model they prefer, usually the newest and most expensive one, and install any plugins they want.

A plugin with repository access can become a channel for code leakage, while the manager responsible may be someone who never saw the plugin.

CIS companies often cannot rely on Copilot, but they face the same control challenge. Their stack includes GigaChat, YandexGPT, Qwen, Claude, and proprietary models, which developers connect in whatever way they can.

At KT.Team, we address this with LLM & Security Gateway: a single point through which all model requests pass.

The gateway routes tasks by class: autocomplete goes to a low-cost model, while legacy refactoring goes to a more capable one.

It enforces team limits, masks secrets and personal data before sending them externally, and keeps an audit log.

Functionally, it is the equivalent of managed-settings.json for a heterogeneous stack, with traffic control added.

- Introduce the metric “cost per merged PR”

and calculate it by team and model. -

Create an allowlist of models by task class and enforce it through managed-settings.json or a gateway. -

Give agents an index and context: language servers, MCP, and RAG over documentation. -

Set spending limits for each team with alerts when they are exceeded.

Each item looks simple, but works only with a mature process.

You need telemetry, a single access point for models, and engineers who can connect the token bill to production outcomes.