Vikey Auto

Vikey Auto is a model id that holds three models: light, medium and heavy. Each time you send a new message, Vikey picks the model that fits the task. A quick question goes to the light model, a feature to the medium one, a large refactor to the heavy one.

Available model ids

Model id	Light	Medium	Heavy
vikey/agent-gpt	GPT-6 Luna	GPT-5.6 Terra	GPT-6.1 Sol
vikey/agent-coding	Claude Haiku 5.5	Claude Sonnet 5.5	Claude Opus 5.5
vikey/agent-mocin	DeepSeek V4.1 Flash	DeepSeek V4.1 Flash 1017	DeepSeek V4.1 Flash 1017

Use it like any model, for example "model": "vikey/agent-gpt" in a request, or claude --model vikey/agent-coding.

The current list and the model at each level are in the Auto Routing menu of the dashboard. In the same menu you can change the model for each level to your liking.

Which level for what

Level	Example tasks
Light	Questions, running tests or commands, commits, typos, one-line changes
Medium	One feature, page, endpoint or script. Bugs with a visible cause.
Heavy	Whole apps, changes across many files, bugs with an unclear cause, security, payments, data migrations

The full description of each level is also in the Auto Routing menu.

How the choice works

• Picked once per new message. The tool steps inside one task (the agent reads files, runs commands, then continues) keep the same model until the task is done. The model does not change mid-task.
• Helper requests do not count. Small requests from the agent, such as making a session title, go straight to the light model with no selection fee.
• If a model has problems, the request moves to the next level automatically.
• If selection is unavailable, the medium model handles the request and no selection fee is charged.

History trimming

Coding agents resend the whole history and every tool result at each step. Over time much of it is no longer used, for example files read for an earlier task.

When you send a new message, Vikey Auto also reviews large old tool results. Those that no longer matter are replaced by a one-line note before they reach the model. Your own messages are never removed. In plain chats (not agents), old history is dropped when the topic changes.

In our tests with real coding agents, the next turn’s prompt shrank by about 51% in Claude Code and opencode, and 53% in Codex. Results depend on session length: largest in long sessions and on topic changes, small in short sessions.

While the agent is still working on one task nothing is trimmed, so caching keeps working. The exception: once a prompt is very large (above about 80 thousand tokens), Vikey can trim in the middle of a task.

Prompt content is not stored

To remember what was already reviewed, Vikey stores only short fingerprints (hashes) for 6 hours. Prompts, code and answers are not stored.

Cost

Part	Cost
Tokens	The price of the model that handled the task
Model selection	Rp10 per new message
Mid-task trimming	Rp10 per trim, only when the prompt is already very large

The selection fee is charged only when the request succeeds. In Usage and the transaction history, it is a separate line with the model vikey/decision.

The Auto Routing menu shows how much you saved compared with official prices, including trimmed tokens.

When you do not need Vikey Auto

For a single small standalone request, such as classifying a short text, the Rp10 selection fee can be more than the price difference. Use a fixed model there.