Choose Kimi K3 on Novita AI when your workload needs a million-token context, flagship reasoning, or one model that can read text, images, and video in the same request. Choose Kimi K2.5 when your prompts fit inside 256K tokens and cost is the deciding factor: it accepts the same three input types at a fifth of K3’s per-token price. Both are serverless on Novita AI and share the same OpenAI-compatible endpoint, so the real decision comes down to context size, price, and which one is the better engineering trade-off for your traffic.
Kimi K3 vs Kimi K2.5: Quick Comparison
| Category | Kimi K3 | Kimi K2.5 | What it means |
|---|---|---|---|
| Best for | Very large working sets and flagship multimodal reasoning | Cost-sensitive multimodal and agentic workloads within 256K | Match the model to how much context you actually send |
| Availability on Novita AI | Serverless API | Serverless API | Both are ready without provisioning GPUs |
| Model ID | moonshotai/kimi-k3 | moonshotai/kimi-k2.5 | Only the model value changes between them |
| Context / max output | 1,048,576 / 1,048,576 tokens | 262,144 / 262,144 tokens | K3 holds roughly four times as much input |
| Input / output modality | Text, image, video to text | Text, image, video to text | Both read images and video; modality is not the differentiator |
| Input price (per 1M tokens) | $3.00 | $0.60 | K3 costs 5x K2.5 on input |
| Cache-read price (per 1M tokens) | $0.30 | $0.10 | Cache-heavy prompts keep the same 3x gap |
| Output price (per 1M tokens) | $15.00 | $3.00 | K3 costs 5x K2.5 on output |
| Migration risk | New model ID and a new production baseline | Established behavior and tuned prompts | Test before switching a live default |
Specs and pricing from the Novita AI model pages and pricing page, checked October 15, 2026.
Pricing Comparison on Novita AI
Novita AI currently lists seven Kimi versions, and they do not all share the same price or context window. The table below is the whole lineup, not just the two models in the title, because the cheapest version for your prompt may not be the newest one.
| Model | Context | Input (per 1M tokens) | Cache read (per 1M tokens) | Output (per 1M tokens) |
|---|---|---|---|---|
| Kimi K3 | 1M | $3.00 | $0.30 | $15.00 |
| Kimi K2.7 Code | 256K | $0.95 | $0.19 | $4.00 |
| Kimi K2.6 | 256K | $0.80 | $0.16 | $3.40 |
| Kimi K2.5 | 256K | $0.60 | $0.10 | $3.00 |
| Kimi K2 Thinking | 256K | $0.60 | $0.15 | $2.50 |
| Kimi K2 Instruct 0905 | 256K | $0.60 | Not listed | $2.50 |
| Kimi K2 Instruct | 128K | $0.57 | Not listed | $2.30 |
Pricing from the Novita AI pricing page, checked October 15, 2026.
Two things stand out. First, K3 is the only version priced in a different bracket: its output rate is five times K2.5’s, and its input rate is five times higher as well. Second, the K2 line is tightly packed. K2.5, K2 Thinking, and K2 Instruct 0905 all list $0.60 input and $2.50-$3.00 output, so a version change inside the K2 family usually costs single-digit cents per million tokens, not dollars.
The gap becomes concrete at volume. For 1M input tokens and 1M output tokens with no cache hits, K3 costs $18.00 and K2.5 costs $3.60. If most of that input is a stable system prompt, tool schema, or repository summary that hits the cache, the cache-read rate applies instead: $0.30 per million on K3 versus $0.10 on K2.5. Caching narrows the absolute difference but keeps K3 roughly three times more expensive on the cached portion.
What Changed From Kimi K2.5 to Kimi K3
K3 is not a small revision of K2.5. Three things changed, and only one of them is price.
Context. K2.5 supports 262,144 tokens of input and output. K3 lists 1,048,576 tokens for both, so it holds about four times as much text before you have to summarize or retrieve. For an agent that loses state during compaction, or a task that spans a whole repository, that is the difference between one request and several.
Scale. K2.5 is a 1-trillion-parameter mixture-of-experts model with 32B parameters active per token. K3 is a 2.8-trillion-parameter model built with Kimi Delta Attention and Attention Residuals. More parameters do not automatically mean better answers on your prompt, but they are part of why the flagship sits in a higher price tier.
Positioning. K3 is Moonshot AI’s current flagship. K2.5 was the flagship of its generation, and it stays on Novita AI as the value option in the same family, alongside the coding-focused K2.7 Code and the reasoning-focused K2 Thinking.
That combination is the reason for the 5x jump: you are paying for a much larger context window and a newer, larger model, not just for a version number. Whether that is worth it depends entirely on whether your workload can use the extra context or the newer reasoning.
Performance and Developer Takeaways
The context window is the difference you will feel first, and the one most likely to change your architecture. If your application spends real effort chunking documents, summarizing repositories, or trimming agent state to fit 256K tokens, K3 removes that work. If your prompts are short and text-only, the extra context buys you nothing and the price gap is pure cost. K3 also runs a larger model, 2.8T parameters against K2.5’s 1T, which is part of what the higher price pays for, though a bigger model does not automatically answer your specific prompt better.
| Signal | Kimi K3 | Kimi K2.5 |
|---|---|---|
| Context window | 1,048,576 tokens | 262,144 tokens |
| Total parameters | 2.8T | 1T |
| Input modalities | Text, image, video | Text, image, video |
There is no shared head-to-head benchmark for the two models, so the numbers that matter are the ones from your own prompts. A quick way to compare before you commit:
- Answer quality on the tasks you actually ship, not on generic leaderboards.
- Output length, since K3’s output price is where the cost difference concentrates.
- Cache hit rate, because the cached portion is billed at a different rate than fresh input.
- Latency and time to first token under your own concurrency, since a larger model can change both.
- Structured-output and tool-calling reliability if your pipeline depends on either.
How to Access Kimi K3 and Kimi K2.5 APIs on Novita AI
Both models are serverless on Novita AI and use the same OpenAI-compatible chat completions endpoint, so switching between them means changing one field: the model value. The base URL and API key stay the same.
Both are live on Novita AI:
A minimal request against K3; change the model string to route to K2.5 or any other Kimi version in the table above.
from openai import OpenAI
client = OpenAI(
api_key="YOUR_NOVITA_API_KEY",
base_url="https://api.novita.ai/openai",
)
response = client.chat.completions.create(
model="moonshotai/kimi-k3",
messages=[{"role": "user", "content": "Summarize the tradeoffs between the Kimi K3 and Kimi K2.5."}],
max_tokens=1024,
temperature=0.7,
)
print(response.choices[0].message.content)If you want the first request, pricing checks, and troubleshooting in one place, start with the Kimi K3 API quick start. For the coding-focused variant, see the Kimi K2.7 Code launch guide.
Best Use Cases
| If your workload… | Choose | Why |
|---|---|---|
| Sends more than 256K tokens of context per request | Kimi K3 | Only K3 in the family reaches 1M tokens |
| Spans a whole repository or a long document set without chunking | Kimi K3 | The larger window replaces retrieval and summarization work |
| Is multimodal and price-sensitive within 256K | Kimi K2.5 | Same text, image, and video input at a fifth of the per-token cost |
| Runs coding agents and repository tasks | Kimi K2.7 Code | A coding-focused sibling with a lower output rate than K3 |
| Needs cheap, long reasoning chains | Kimi K2 Thinking | The lowest output rate in the 256K tier |
| Mixes easy and hard requests | Route by tier | Keep K2.5 as the fast path and escalate only large-context work to K3 |
Migration Notes
Switching from K2.5 to K3 is a one-line change, but a few checks keep it safe:
- Confirm the current model ID and pricing on the model page before you ship; the Kimi lineup has changed prices more than once this year.
- Update the
modelvalue and keep everything else the same, since both models share the endpoint and message format. - Re-check your max-token settings. A 1M-token window invites larger prompts, which raises input cost even when the per-token rate is unchanged.
- Measure cache behavior before assuming savings; cache-read pricing only applies when your request pattern actually reuses cached content.
- Keep the model ID, context limit, and pricing in configuration rather than hardcoded, so you can roll back to K2.5 without a redeploy.
- Run a prompt regression set and compare quality by task type, not only aggregate win rate.
Conclusion
Kimi K3 and Kimi K2.5 sit at two ends of the same family. K3 is the choice when context is the bottleneck: a 1M-token window, a 2.8T-parameter flagship, and multimodal input justify a per-token price five times higher than K2.5. Kimi K2.5 is the better default when your prompts fit inside 256K tokens, because it handles the same input types and the same API for a fifth of the cost. For most production systems the answer is not one model but a routing rule: send routine and mid-size multimodal traffic to K2.5, and reserve K3 for the requests where its larger working set removes real engineering work.
Check the current model pages for pricing and IDs before you commit a cost model to production, and test both models on your own prompts before switching a live default.
FAQ
What is the difference between Kimi K3 and Kimi K2.5?
Kimi K3 is the newer flagship with a 1,048,576-token context window, 2.8T parameters, and text, image, and video input. Kimi K2.5 is the previous-generation model with a 262,144-token window, 1T parameters, and the same three input types. K3 costs five times more per token on both input and output.
How much does Kimi K3 cost on Novita AI?
As of October 15, 2026, Kimi K3 lists $3.00 per million input tokens, $0.30 per million cache-read tokens, and $15.00 per million output tokens.
How much does Kimi K2.5 cost on Novita AI?
As of October 15, 2026, Kimi K2.5 lists $0.60 per million input tokens, $0.10 per million cache-read tokens, and $3.00 per million output tokens.
How many Kimi versions are available on Novita AI?
Seven: Kimi K2 Instruct, Kimi K2 Instruct 0905, Kimi K2 Thinking, Kimi K2.5, Kimi K2.6, Kimi K2.7 Code, and Kimi K3. Each has its own context window and pricing.
Is Kimi K3 worth the price increase over Kimi K2.5?
Only if your workload uses the larger context window or the newer flagship. If your prompts fit inside 256K tokens, K2.5 delivers the same input modalities for a fifth of the per-token cost, so K3 is hard to justify. Test both on your own prompts before deciding.
Do Kimi K3 and Kimi K2.5 use the same API?
Yes. Both are serverless on Novita AI and use the same OpenAI-compatible chat completions endpoint, so you switch between them by changing only the model value.
