Kimi K3 API Guide: Pricing, 1M Context, Model IDs & Production Fit

Kimi K3 at a glance

The table below contains only information documented by Moonshot on July 16, 2026.

Kimi K3 model IDs depend on the access channel

The string used to call K3 is not universal. Moonshot's direct developer platform and the Kimi Code subscription product use different identifiers, entitlements, and billing systems.

Kimi Code's plan limits also explain why two developers can select K3 but see different context capacity. The official Kimi Code documentation lists no K3 access for Andante, a 256K K3 limit for Moderato, and up to 1M for Allegretto and higher plans.

What changed from the K2 generation

K3 is not simply K2.6 with a larger context setting. Moonshot describes it as its most capable model and highlights a new architectural combination: Kimi Delta Attention, a hybrid linear-attention approach, and Attention Residuals.

For product teams, the practical changes are easier to understand in workload terms:

Who should evaluate Kimi K3 now?

K3 has a credible early fit when the value of completing a difficult task is much larger than the token bill.

This is why a benchmark headline alone will not decide K3's production value. Teams need to measure whether the model completes hard tasks with fewer retries, less human correction, or fewer route escalations.

Direct API pricing and practical cost

Moonshot lists one price across the full K3 context range rather than separate short- and long-context tiers:

These are Moonshot direct API prices, not EvoLink prices. EvoLink pricing cannot be published until the route, provider channel, billing wrapper, and production behavior are verified.

Do not mix three different cost channels:

• Moonshot Open Platform uses the token prices above.
• Kimi Code uses membership quotas and plan entitlements rather than this direct API price table.
• Third-party gateways and aggregators set their own availability, provider routing, and prices. Community posts quoting dollar prices are useful demand signals, but they are not evidence of Moonshot or future EvoLink pricing.

Example cost scenarios

The examples below use Moonshot's direct list price and exclude retries, tools, storage, taxes, gateway pricing, or failed requests.

Caching changes the economics materially. A stable 500K-token repository or knowledge prefix costs about ¥10 as uncached K3 input but about ¥1 when the input hits cache. The output remains the expensive part, so applications should still control tool loops, verbosity, retry policy, and completion budgets.

K3 versus K2.6 direct list price

K3's uncached input is about 3.1 times the K2.6 price, while output is about 3.7 times the K2.6 price. That gap supports a routing policy: use K3 when task difficulty, context size, or tool complexity warrants escalation, and keep less expensive models for routine work.

The production metric that matters: cost per successful task

List price is only the beginning. A premium model can still be economical when it prevents retries or completes work that a cheaper route cannot finish. Conversely, a strong model can be a poor default if it overthinks simple tasks or produces long expensive answers.

Measure K3 with:

The useful comparison is not “which model has the cheapest token?” It is “which route completes this workload at the lowest accepted-result cost within the required latency and reliability limits?”

Developer controls: reasoning, tools, and multimodal input

Reasoning behavior and request controls

K3 always reasons. Moonshot's documentation says:

• reasoning_effort is the K3 reasoning parameter.
• Only the max level is currently supported.
• The K2.x thinking parameter should not be used for K3.
• Streaming responses separate reasoning_content from final content.
• Multi-turn and tool workflows must return the complete assistant message to the next request.

That last rule is a common agent integration failure. If an application stores only the final text and drops reasoning-related response fields or tool calls, the next turn may not preserve the state expected by the model.

Tool use: K3's most important developer feature

K3 adds two controls that matter for production agents:

1. tool_choice can require a tool call when the workflow must retrieve data or perform an action before answering.
2. Dynamic tool loading lets an application inject a tool definition into the conversation only when it becomes relevant.

The recommended pattern for a large tool catalog is:

This is more than a token optimization. Large overlapping tool catalogs often reduce selection accuracy. Dynamic loading turns tool discovery into part of the agent architecture.

Multimodal input and current constraints

K3 accepts text, images, and video, but developers should note the input rules:

• Visual message content must use the documented object-array format.
• Public image URLs are not supported in the K3 quickstart guidance.
• Images can be supplied as base64 data.
• Uploaded files can be referenced through Moonshot's ms:// file mechanism.
• Video inputs can be uploaded and referenced as files.

These are direct Moonshot interface details. A future EvoLink route may expose different file handling or compatibility behavior, so applications should not assume the vendor upload flow will be identical on the gateway.

Long context, caching, and common access issues

1M context does not remove the need for context engineering

A million-token window makes larger tasks possible, but it does not make every token useful. Sending an entire repository, tool catalog, log archive, and conversation history can still increase cost, latency, and distraction.

Use the larger window deliberately:

The best K3 workflow may use fewer carefully selected tokens than the maximum window allows.

Why Kimi K3 may show 256K, return 401, or use more quota

Several launch-week problems come from mixing API behavior with Kimi Code subscription behavior.

Third-party coding clients may expose their own context controls. A “1M-capable” account does not guarantee that every client automatically sends the maximum window.

Production evaluation and remaining unknowns

What remains unverified

Moonshot's launch documentation is unusually concrete about the API, but several important questions remain open:

Early community demos describe strong coding and web-building results, but they also mention slow responses, overthinking, and launch-night errors. These are useful test ideas, not reliable model facts or production guarantees.

A production evaluation plan

Do not evaluate K3 with one impressive prompt. Build a small workload that represents real customer value.

Phase 1: establish baselines

Choose 20–50 tasks across:

• repository understanding;
• multi-file code changes;
• bug localization;
• long-document synthesis;
• structured extraction;
• multimodal review;
• multi-step tool use.

Record the current route's success rate, latency, input/output tokens, retry count, and reviewer effort.

Phase 2: test K3 directly

Use the official Moonshot route only if direct vendor testing fits your procurement and data policy. Keep prompts, tools, and acceptance criteria consistent. Record cache behavior separately from first-turn uncached cost.

Phase 3: decide the future routing role

K3 does not need to replace the baseline model to be valuable. It may be best as:

• an escalation route after a cheaper model fails;
• a repository-scale coding route;
• a long-document synthesis route;
• a specialist for tool-heavy workflows;
• a premium user-selectable option.

Phase 4: verify the EvoLink route when available

When EvoLink publishes a verified K3 route, repeat a smaller compatibility suite:

• model ID and request shape;
• reasoning field preservation;
• streaming behavior;
• tool calls and structured output;
• image and video input;
• 1M context handling;
• billing and cache accounting;
• timeout, retry, and fallback behavior.

How EvoLink users should prepare

EvoLink's value is not to turn every newly released model into an automatic default. A unified API gateway should make model choice, cost control, provider fallback, and migration easier.

While K3 is not yet integrated, teams can:

1. Review the current EvoLink model catalog instead of assuming a K3 route exists.
2. Build workload-specific acceptance tests that can be replayed when K3 becomes available.
3. Keep model selection behind configuration rather than hard-coding one provider.
4. Define cost and latency thresholds for escalation to a premium model.
5. Read the developer guide to AI model routing for a broader routing framework.
6. Use the Kimi K3 availability tracker for EvoLink-specific status.

The production opportunity is model optionality: route ordinary work economically, reserve premium reasoning for tasks that benefit from it, and keep a fallback when one provider is slow or unavailable.

FAQ

Is Kimi K3 officially released?

Yes. Moonshot published Kimi K3 in its official model list, quickstart, pricing documentation, and direct API examples on July 16, 2026.

What is the official Kimi K3 model ID?

Is the Kimi Code k3 model ID the same as kimi-k3?

Is kimi-for-coding Kimi K3?

Why does Kimi Code show only 256K context for K3?

Why does Kimi Code return 401 for K3?

A 401 can indicate that the current membership tier does not include the requested model or context entitlement. Check the plan, refresh authentication, and retry in a new session.

Why can quota usage rise after switching to K3?

Kimi Code says context cache is not reused across model switches. Starting a new session avoids carrying a long conversation into a route whose previous cache is no longer valid.

How large is the Kimi K3 context window?

Moonshot documents a 1,048,576-token context window.

Does Kimi K3 support images and video?

Yes. Moonshot documents native visual understanding with text, image, and video input. Direct API visual inputs must follow Moonshot's documented file or base64 formats.

Can Kimi K3 disable reasoning?

How much does Kimi K3 cost?

Moonshot's direct list price is ¥2 per million cached input tokens, ¥20 per million uncached input tokens, and ¥100 per million output tokens. These are not EvoLink prices.

Is Kimi K3 open source?

An official K3 weight release and license were not found in the reviewed launch sources. Do not describe K3 as open source until Moonshot publishes the weights and license.

Is Kimi K3 available on EvoLink?

Should teams replace K2.6 or another current model with K3?

Not automatically. K3 is substantially more expensive at Moonshot's direct list price. Test it on difficult long-context, coding, research, multimodal, and tool-heavy tasks, then choose a specialist, escalation, or premium routing role.

What should developers test first?

Start with tasks that currently fail because of repository size, document volume, long tool histories, or complex reasoning. Measure accepted-result cost and reviewer effort, not only output quality.

Sources

• Moonshot: Kimi K3 quickstart and model overview
• Moonshot: Kimi K3 direct API pricing
• Moonshot: current model list
• Moonshot: model parameter reference
• Moonshot: Kimi K3 tool-calling best practices
• Moonshot: direct API rate-limit tiers
• Kimi Code: supported model IDs, plans, context limits, and switching behavior

Community discussions were used only to identify evaluation questions and early adoption concerns. Model facts, prices, IDs, limits, and API behavior in this review are based on Moonshot's official documentation.