At 01:15 Beijing time on October 24, 2026, Kimi announced on its official X account that it has released Kimi K3.1 Pro, along with K3.1 Pro Highspeed. The short version:
- K3.1 Pro is K3 with more post-training compute behind it: 2.8T total parameters, 104B active (MoE), a 1M-token context window, and multimodal support.
- Kimi says the Highspeed version is 2–3x faster than K3.1 Pro.
- The official API model IDs are
kimi-k3.1-proandkimi-k3.1-pro-highspeed. - Our platform (aicodewith) currently supports
kimi-k3, and K3.1 Pro is not live yet. The Claude Code guide and hands-on test below both usekimi-k3, which works end to end. Once K3.1 Pro is live, you only need to swap the model ID.
Key Points from the Launch
Everything below comes from the Kimi tech blog and the official post:
| Item | Official statement |
|---|---|
| Release time | 01:15 Beijing time on October 24 (time of the main post) |
| Positioning | Original: "our most capable 3T-class open-weight model to date" |
| Parameters | Original: "2.8T total / 104B active params (MoE), Multimodal" |
| Context | Original: "a 1-million-token context window" |
| Highspeed | Original: "delivers 2-3× the speed of K3.1 Pro" |
| Availability | Original: "available today on Kimi.ai, Kimi Work, Kimi Code, and the Kimi API" |
| Open weights | Original: "Full model weights will be released by Nov 4, 2026." |
| API model ID | Original: "select kimi-k3.1-pro / kimi-k3.1-pro-highspeed" |
Enterprise note: the announcement states that Kimi Enterprise admins must manually enable K3.1 Pro for their organization. The original reads "access is off by default at launch".
Official Self-Reported Benchmarks
All scores below are self-reported in the official Kimi blog. K3.1 Pro's coding results were measured under the Claude Code harness with reasoning effort set to max:
| Benchmark | K3.1 Pro | Comparison |
|---|---|---|
| Terminal-Bench 4.0 | 65.7% | GPT-6 Astra 58.2%, Claude Opus 5.5 66.4% |
| Kimi Code Bench 3.0 Long-horizon (Kimi internal benchmark) | 61.2% | K3 55.95% |
| BenchCAD (with tools) | 92.7% | Claude Opus 5.5 96.2%, GPT-6 Astra 95.9% |
How K3.1 Pro Differs from K3
| Item | Kimi K3 | Kimi K3.1 Pro |
|---|---|---|
| Relationship | Base model | Official text: "a post-trained version of Kimi K3, with substantially larger post-training compute" |
| API model ID | kimi-k3 | kimi-k3.1-pro / kimi-k3.1-pro-highspeed |
| Kimi Code Bench 3.0 Long-horizon (self-reported) | 55.95% | 61.2% |
| Hallucination rate (Kimi Hallu Bench Hard, self-reported, claim-level) | 23.4% | 8.3% |
| Available on our platform (aicodewith)? | Yes, verified with a key from our platform | Not live yet |
Kimi's own caveat on the hallucination numbers is that this is a stress test built around error-prone scenarios: "The reported rates do not represent overall error rates in daily use."
For how Codex users can connect K3, see Connecting Kimi K3 to Codex. This post only covers Claude Code.
Connecting Claude Code
1. Create an API Key on our platform
After signing in, create an API Key in the console. Refer to the model list for the models currently available.
2. Write to the global settings.json
Open ~/.claude/settings.json(on Windows, %USERPROFILE%\.claude\settings.json), and in env add the base URL, key, and model name. Note that ANTHROPIC_BASE_URL mustnot have /v1at the end.
The setup that works on our platform today (the model is kimi-k3):
{
"env": {
"ANTHROPIC_BASE_URL": "https://api.aicodewith.ai",
"ANTHROPIC_AUTH_TOKEN": "<你的 API Key>",
"ANTHROPIC_MODEL": "kimi-k3",
"ANTHROPIC_DEFAULT_SONNET_MODEL": "kimi-k3",
"ANTHROPIC_DEFAULT_OPUS_MODEL": "kimi-k3",
"ANTHROPIC_DEFAULT_HAIKU_MODEL": "kimi-k3",
"CLAUDE_CODE_SUBAGENT_MODEL": "kimi-k3"
}
}
For K3.1 Pro, just swap in the official model ID kimi-k3.1-pro(for Highspeed, use kimi-k3.1-pro-highspeed). The model IDs come from the Kimi tech blog , which says "select kimi-k3.1-pro / kimi-k3.1-pro-highspeed":
"ANTHROPIC_MODEL": "kimi-k3.1-pro"
K3.1 Pro update: Our platform doesn't have K3.1 Pro yet, so using kimi-k3.1-pro will return the error described in the "model not found" section below. We'll update the test results here once it's live.3. Switch with /model
After launching Claude Code, type /model to switch models. The config above points the default models for Sonnet, Opus, Haiku, and subagents all to kimi-k3, so after switching, or when Claude Code launches a subagent, it never lands on a model our platform doesn't have.
4. Verify the connection
Claude Code shows this prompt on startup:
Claude Code doesn't recognize the model "kimi-k3", so it assumes a context window of 200k tokens.
This is Claude Code's notice for non-Claude model IDs and doesn't affect calls. To verify, give it a small task and check the response.
Hands-on test: writing a function and running tests with kimi-k3 in Claude Code
This test used kimi-k3, not K3.1 Pro. It ran in the early morning of October 24 Beijing time, on Claude Code version 2.1.309, with the settings.json above.
Task: in an empty directory, write a Python function slugify(s)(lowercase, convert spaces and underscores to hyphens, strip other characters, collapse repeated hyphens, trim leading and trailing hyphens), then write unittest tests and run them.
Result:
- It took about 72 seconds over 5 turns, and the model reported in the response was
kimi-k3. - It generated
slugify.pyandtest_slugify.py, with 8 test cases in total. We re-ran them locally and gotRan 8 tests ... OK.
The core code it generated:
def slugify(s):
s = s.lower()
s = re.sub(r"[\s_]", "-", s)
s = re.sub(r"[^a-z0-9-]", "", s)
s = re.sub(r"-{2,}", "-", s)
return s.strip("-")
How to Choose Between K3.1 Pro and Highspeed
The only difference the official docs call out is speed: Highspeed "delivers 2-3× the speed of K3.1 Pro for more responsive interactions." They don't say whether the two differ in capability, so here's one way to choose:
- Interactive code edits and frequent back-and-forth: Try Highspeed first for shorter wait times.
- Long tasks, cross-file refactors, and long-running unattended jobs: Use K3.1 Pro, and compare the results of both on your own tasks before you decide.
For pricing, go by the official Kimi pricing page or this site's model list as displayed on the day you check.
Common Errors
401: Invalid Key
First check that ANTHROPIC_AUTH_TOKEN was pasted in full, has no extra spaces, and isn't overridden by an environment variable of the same name that takes precedence over settings.json. Restart Claude Code after making changes.
429: Too Many Requests
This usually means too many requests in a short time. Wait a moment and retry, or reduce the number of parallel tasks and subagents.
model not found: Model Does Not Exist
If you use a key from this site to request a model that isn't live yet, Claude Code actually outputs:
API Error: 400 {"error":"模型 kimi-k3.1-pro 不存在或未上线，请前往 https://aicodewith.ai/zh/dashboard/pricing 查看可用模型列表 (request_id: ...)"}
When that happens, change the model name back to kimi-k3, and check whether ANTHROPIC_DEFAULT_*_MODEL and CLAUDE_CODE_SUBAGENT_MODEL are also set to models that are available.
FAQ
What is the API model ID for Kimi K3.1 Pro?
The official IDs are kimi-k3.1-pro and kimi-k3.1-pro-highspeed.
Can I use K3.1 Pro on this site right now?
Not yet. This site currently supports kimi-k3, and this article will be updated once K3.1 Pro goes live.
Should the base URL include /v1?
In Claude Code, set ANTHROPIC_BASE_URL to https://api.aicodewith.ai, without /v1.
Is the "doesn't recognize the model" message at startup a problem?
No, it doesn't affect calls. It's just Claude Code's notice for non-Claude model IDs.
When will the open weights for K3.1 Pro be released?
From the official statement: "Full model weights will be released by Nov 4, 2026."
How do I connect Kimi to Codex?
See Connecting Kimi K3 to Codex.
