Kimi K3.1 Pro on Magicdoor: Workflow, Pricing and Limits

Direct answer: Choose Kimi K3.1 Pro for a text or image-based coding, planning or knowledge-work task, then evaluate its answer against an explicit check. It is the current Kimi selection. Kimi K3 and Kimi K3 (Renewable) are retired: do not select them for new work or assume Pro uses their old routes or prices.

Updated October 28, 2026. See the Kimi model page for current specifications and rates.

Start with a task Pro can actually accept

1. Open chat, sign in if required, and open the model selector.
2. Choose Kimi K3.1 Pro. Opening its details is informational; use the selection action to change the answer model. The chat link does not preselect Pro.
3. Send text or an image, for example a screenshot and “identify the visible error, propose a patch, and list tests I should run.” Paste a small code excerpt as text instead of assuming a source-file upload or code interpreter is supported.
4. Ask for an acceptance check: tests for a patch, cited sources for current facts, or an explicit list of uncertainties. Search must be available and enabled for current web research; generated code is not proof that it was executed.
5. Review the recorded response model and cost when available, and the running chat cost. Use Usage to inspect balance and metered work. Retry or change models deliberately if the route is unavailable.

Pro is not one of Experimental Auto's four answer models. Select it manually when its identity matters; the Auto guide explains automatic model selection.

What changed from K3

A saved retired-K3 conversation can be read, but you must select a current model before continuing. Legacy K2.x names still normalize to the retired K3 identity; they do not silently upgrade to Pro. Retired K3 model-detail URLs are not current destinations. Use the model directory or the current Pro link above.

For renewable-powered inference, choose a currently supported GLM or DeepSeek route in the Renewable guide. Pro has no Renewable option.

Estimate cost using tokens, not prompts

An uncached 10,000-input plus 2,000-output-token example costs $0.04: $0.02 input + $0.02 output. This assumes no cache reads or writes, and is not an average cost per answer. A cache-read example with 8,000 of those input tokens reported as cached costs $0.0256. Token categories must come from actual usage; do not add the full new-input and cache-write price twice for the same tokens.

Basic is $6/month, with $2 included usage balance in the first month and $1 on renewal. Plus is $12/month with $7 monthly usage balance. Included monthly balance resets; purchased top-ups do not expire. Tools can add separate charges. A failed save, cancellation or credit stop does not promise that already completed model or search work is free. See the model-cost guide for units, tiers and other models.

Inputs, reasoning and practical limits

• Text and image input, text output. No native image, audio or video output. Image creation uses a separate Magicdoor tool/model.
• No PDF input or hosted code interpreter. Choose a compatible model for PDFs or executable data work; changing models does not make every old attachment compatible.
• Always-on reasoning. Reasoning is always enabled. The model picker has no adjustable reasoning-effort control or thinking-off switch for Pro.
• Shared 1,048,576-token context. Input, reasoning and output share the window. Magicdoor allows an input budget of 1,000,000 tokens and caps output at 32,768 tokens; neither means unlimited history or an additional million output tokens. Compaction and provider-specific history filtering affect what is actually sent.
• Search, image creation and Canvas are tools, not native output modalities. Their availability depends on the active model and current controls. This is not a promise to reproduce every Moonshot demo, third-party agent environment or file workflow.

Privacy and continuation boundaries

Pro requests use Moonshot AI with no-training enforcement. If the required privacy or continuation protections are unavailable, the service stops rather than silently switching to another inference provider.

Vercel's no-training documentation distinguishes system credentials from BYOK and no training from retention. Magicdoor does not expose a bring-your-own-Moonshot-key workflow. The no-training filter is not a promise that Magicdoor stores no chat history.

Saved Pro conversations retain private reasoning and completed tool turns for continuation by their owner. Shared responses and conversation forks do not include that private continuation data, so they may not continue in the same way as the original conversation. Ordinary visible messages and account storage have separate handling. This does not make all chat storage end-to-end encrypted, and it does not guarantee lossless context when switching to a different model.

If Pro cannot continue, inspect the error and select a compatible model explicitly. Choose a model compatible with the conversation and its attachments. Do not expect retired K3 or Renewable K3 to be restored, or assume that Pro is part of Auto's fallback pool.

Sources and scope

Source-checked October 28, 2026:

• Gateway endpoint metadata: token rates, text/image input, text output and shared context metadata match the registered contract.
• Moonshot quickstart: reasoning, vision and tool-use context. Upstream capabilities do not automatically become Magicdoor controls.
• Vercel no-training policy: enforcement, system credentials and the retention distinction.

Compare answers against your own acceptance checks rather than assuming that a newer version is better for every task.

FAQ

Is Kimi K3.1 Pro available on Magicdoor?

Yes. Select Kimi K3.1 Pro in the chat model picker. Route availability still depends on service capacity and your account access. Kimi K3 and Kimi K3 (Renewable) are retired, not fallback choices.

What does Kimi K3.1 Pro cost?

The registered rates are USD $2 input, $0.20 eligible cached input, $2 cache writes and $10 output per million tokens. An uncached 10,000-input plus 2,000-output-token request costs $0.04 before subscription or tools. Cache hits and response length are not guaranteed.

Can I use PDFs, code execution or Renewable with Pro?

Pro accepts text and image input and produces text. It does not expose PDF input, a hosted code interpreter or a GreenPT Renewable route. Search, image creation and Canvas are separate Magicdoor tools when available, with their own capability and billing boundaries.

Does no training also guarantee no retention?

No. The route pins Moonshot AI through Vercel AI Gateway system credentials and requires no-training enforcement. Provider retention and Magicdoor chat storage are separate. Private replay encryption is not end-to-end encryption or a platform-wide zero-retention promise.