Kimi K3.1 Pro Costs $2 for 1 Million Tokens: Is It the New Open Agent King?

Key takeaways

• Kimi K3.1 Pro was released on 23 October 2026 as the first model in Moonshot AI's K3.1 series, with the official blog listing dated 24 October 2026.
• It is open-weight, with a 1,048,576 token context window, and it is designed for agent work, coding, and long document tasks through an API.
• Standard Kimi API pricing is $2.00 per 1 million input tokens, $0.20 per 1 million cached input tokens, $2.00 per 1 million tokens for a 5 minute cache write, and $10.00 per 1 million output tokens.
• Reported benchmark highlights include 67.2 percent on APEX-Agents, 65.7 percent on Terminal-Bench 4.0, 75.5 percent on DeepSWE 1.1, and 94.2 percent on GPQA Diamond.
• K2 series models stop opening to new registered users from 31 October 2026, existing registered users are unaffected, and Kimi plans to take all K2 series models offline at the end of December.
• The practical trap is output cost. A long agent run can cost far more in output tokens than the $2 input headline suggests, so set a budget and test with small jobs first.

What is Kimi K3.1 Pro?

Kimi K3.1 Pro is Moonshot AI's new open-weight model for agentic work. It is the first release in the K3.1 series, and it follows Kimi K3, which readers may know from earlier open model roundups. The key change is not just a new name. K3.1 Pro is positioned as a model you call through an API for tasks that need tools, code, browsing, and long context, rather than a small model you run on a laptop.

That distinction matters. The reported parameter count is 2.8 trillion, which puts it firmly in the datacenter class. Do not plan to download it and run it on consumer hardware. The normal way to use it is through Kimi's API, or through a provider that hosts it, where you pay per token. If your main goal is private local use, a small model such as the 4GB NVIDIA model covered in our other new post is a better starting point. If your goal is a capable agent that can work through a long task, K3.1 Pro is the kind of model to test.

The 1 million token context is also central to the pitch. Kimi lists 1,048,576 tokens for K3 series models. In plain terms, that is room for very long documents, large codebases, or long agent histories in one run. Long context is useful, but it is not free. Every token you send, and every token the model returns, counts toward cost, and long histories can make agent loops expensive if you do not control them.

What do the benchmarks actually say?

Benchmarks do not prove that a model will work well on your job, but they do show where it is strong. K3.1 Pro's reported results are strongest on agent and reasoning tasks, which matches its positioning. The headline agent result is 67.2 percent on APEX-Agents, ranked first in the tracker summary. It also scored 65.7 percent on Terminal-Bench 4.0, ranked third, just behind Claude Opus 5.5 at 66.4 percent. That terminal result matters because it tests real computer work, not just answering questions.

For coding, the reported DeepSWE 1.1 score is 75.5 percent, which is a strong software engineering signal. For knowledge and reasoning, GPQA Diamond is reported at 94.2 percent, and MMMU-Pro at 84.1 percent, ranked first. For research style work, BrowseComp is reported at 92.1 percent, ranked second, and CharXiv Reasoning at 89 percent, rising to 93.4 percent with tools. Humanity's Last Exam is reported at 51.1 percent without tools and 70 percent with tools, ranked first with tools.

Read those numbers with care. A benchmark is a controlled test, not a guarantee. Tools change the result, as the Humanity's Last Exam gap shows. The right test is a small, realistic job from your own work, with a clear pass or fail check. Use benchmarks to shortlist K3.1 Pro, then test it on one task you already do, and compare the result with the model you use now.

How much does Kimi K3.1 Pro really cost?

The standard Kimi API price is the number to start with. Input costs $2.00 per 1 million tokens. Cached input costs $0.20 per 1 million tokens. A 5 minute cache write costs $2.00 per 1 million tokens. Output costs $10.00 per 1 million tokens. Kimi API is pay as you go, not a subscription, and it should not be confused with Kimi membership or Kimi Code, which are different products.

The Highspeed variant, reported through OpenRouter data on the tracker, costs $4.00 per 1 million input tokens and $20.00 per 1 million output tokens. That is double the standard input price and double the output price. Use Highspeed only if speed is worth that premium for a specific job. For most testing, standard pricing is the sensible default.

Here is a simple worked example. Suppose an agent job uses 100,000 input tokens and returns 20,000 output tokens. On standard pricing, input costs 0.20 dollars and output costs 0.20 dollars, for a total of 0.40 dollars. On Highspeed pricing, the same job costs 0.40 dollars for input and 0.40 dollars for output, for a total of 0.80 dollars. If the same input is cached, the cached portion at standard pricing costs only 0.02 dollars per 100,000 tokens, before any cache write cost. That is why repeated work on the same documents can be much cheaper than fresh work every time.

| Job example | Standard input | Standard output | Standard total | Highspeed total | | 100,000 input tokens, 20,000 output tokens | $0.20 | $0.20 | $0.40 | $0.80 | | 500,000 input tokens, 100,000 output tokens | $1.00 | $1.00 | $2.00 | $4.00 | | 100,000 cached input tokens, 20,000 output tokens | $0.02 | $0.20 | $0.22 plus any cache write | $0.80, before cache rules |

The table shows the main lesson. Output is the expensive direction. An agent that loops, retries, or writes long reports can quickly cost more than a single chat reply. Set a maximum token budget per job, limit retries, and log the first ten jobs before you scale anything.

What is the K2 deadline trap?

The deadline is the most urgent part of this story for existing Kimi users. Kimi's pricing docs state that K2 series models stop opening to new registered users from 31 October 2026. Existing registered users are unaffected by that part of the change. Kimi also plans to take all K2 series models offline at the end of December. That second date is the one that affects everyone still using K2.

The practical action is simple. If you have a project, script, or product that calls a K2 model, check the model name in your code now. Do not wait until December. Test K3.1 Pro on the same job, compare the output, and update the model name and any reasoning settings before the old model goes offline. Also check any saved prompts that assume K2 behaviour, because a new model can format answers differently even when it is stronger overall.

If you are new to Kimi, the deadline is less of a trap and more of a reason to start with K3.1 Pro. Do not build a new workflow on a model that is scheduled to go offline. Start with the current series, and keep your model name in one place in your code so the next change is a one line edit, not a search through many files.

How should you test Kimi K3.1 Pro in one afternoon?

Start with a small, measurable test. Pick one task you already do, such as triaging a long document, fixing a small coding bug, or drafting a report from notes. Write down what a good result looks like before you run it. Then run the same task on K3.1 Pro and on your current model, and compare accuracy, completeness, and cost. One afternoon of controlled testing beats a week of casual chatting.

Next, test long context carefully. Give it a long document and ask for a specific fact, a summary with page references, and a list of items that meet a clear rule. Check each answer against the source. Long context models can still miss details, and the only way to know is to verify. Also test tool use in stages. Start with one tool, then add a second tool, and watch for loops, repeated calls, and wasted tokens. If the agent repeats itself, tighten the instructions and add a clear stop rule.

Finally, test reasoning effort. Kimi allows reasoning effort to be set with reasoning_effort, using low, high, or max. Low is cheaper and faster for simple jobs. High or max may help for hard reasoning, but will cost more in output tokens. Do not leave every job on max. Match the setting to the job, and record which setting you used so your cost comparison is fair.

Who should use it, and who should skip it?

Use K3.1 Pro first if you build or test agents, if you work with long documents or large codebases through an API, or if you need strong tool use and are willing to pay per token. It is also a good candidate if you are already on Kimi and need a path off K2 before December. Developers who want a strong open-weight option without running datacenter hardware themselves will find the API route the most practical.

Skip it for now if you only need everyday chat, if you want a free local model, or if you do not want to manage API costs. A subscription chatbot is simpler for casual use, and a small local model is better for private offline work. Also skip a full migration until you have tested K3.1 Pro on your own tasks. Benchmark leadership is a reason to test, not a reason to move a production system in one day.

The bottom line

Kimi K3.1 Pro is a credible new open agent model, with strong reported agent and coding scores, a 1 million token context, and clear API pricing at $2.00 input and $10.00 output per 1 million tokens on the standard tier. The headline input price is attractive, but output cost and agent loops decide the real bill. Test it on one real job, watch the token usage, and if you use K2, plan the move before the end of December deadline.

Explore the AI Basics and Tutorials hub, read Best Open Source LLMs in 2026: Kimi K3, GLM-5.3 and Qwen, and compare prices in Open Source LLM API Pricing: Kimi, GLM, Qwen and DeepSeek (2026).

Sources: Kimi API pricing docs at platform.kimi.com, Kimi research blog listing for Kimi K3.1 Pro dated 24 October 2026, and AI Release Tracker Kimi K3.1 Pro release page for 23 October 2026.

Disclosure: Links in this article are informational. No affiliate commission is claimed for Kimi API pricing links.