Kimi K3.1 Pro: $2/$10, 1M context, open weights by Nov 4. Should you switch from Opus 5.5?

Moonshot dropped Kimi K3.1 Pro on Oct 23 and it's sitting at ~900 points on Hacker News. The headline: an open-weight model that lands within a point of Claude Opus 5.5 on a coding agent benchmark, at a fraction of the price.

Let's separate the signal from the launch-day gloss.

The numbers (vendor-reported)

From Moonshot's own post:

Also claimed: a 1M-token context window, and a claim-level hallucination rate on their internal "Hallu Bench Hard" dropping from 23.4% (K3) to 8.3%.

Two caveats before you tweet "Opus killer":

1. Terminal-Bench was run with the Claude Code harness for Kimi. Fair, but it's their run vs. leaderboard maxima for everyone else.
2. Hallu Bench is internal, and they say so themselves: the rates "do not represent overall error rates in daily use."

Price

Per the Kimi platform pricing doc (quoted in the HN thread) and third-party listings:

• kimi-k3.1-pro: $2 / MTok in, $10 / MTok out, cached input $0.20
• kimi-k3.1-pro-highspeed: 2-3x faster, $4 / $20

One HN user ran a Pac-Man bake-off: K3.1 Pro tied Sonnet 5.5 at a score of 93, for $0.68 and 11 minutes vs. $1.61 and 14 minutes. One anecdote, one task. But it points the same direction as the price sheet.

Independent check: another commenter pointed to Artificial Analysis putting it at an index of 55, 4th overall, ahead of GPT-6 Astra and Fable 5.1. Same commenter notes it burned 150M tokens to run the index vs. a median of 87M. It's verbose. Per-token price is not per-task price, so measure on your own workload.

Switching is nearly free

The API is OpenAI-compatible, so trying it is a config change:

from openai import OpenAI

client = OpenAI(
    base_url="https://api.moonshot.ai/v1",  # check platform.kimi.ai for your region's endpoint
    api_key=os.environ["KIMI_API_KEY"],
)

resp = client.chat.completions.create(
    model="kimi-k3.1-pro",
    messages=[{"role": "user", "content": "Refactor this module to remove the global state."}],
)
print(resp.choices[0].message.content)

In Kimi Code you switch with /model. It's also on OpenRouter per search listings.

The interesting part: they scrubbed cyber data

This is the most unusual thing in the post. Moonshot says vulnerability-research ability rose steadily as post-training scaled, and they traced it to cyber content sitting in the general mid-training corpus and cyber tasks in the RL environment mix. For the open-weight release they removed both.

Their evidence, on an in-house exploit benchmark (xdev-bench-hard, pass@3):

• Same pipeline, no data removal: 90.5%
• Final K3.1 Pro: 81.0%
• K3 (previous): 33.3%

So the removal cost ~9 points, yet the model is still far above its predecessor, because better general coding generalizes into exploit-writing anyway. Moonshot says it still trails frontier proprietary models that have been public for 3+ months.

The HN skeptics have a point: how much "cyber data" is just ordinary systems knowledge? You can't cleanly delete the ability to reason about memory layouts. Treat this as risk reduction, not a guarantee. And if you do defensive work, expect to hit the same ceiling you hit with other hosted models.

The catch: quotas and "open" timing

• Weights aren't out yet. Promised "by Nov 4." Until then it's an API model. Don't plan self-hosting around a date.
• Subscription limits are getting roasted. One commenter hit the 5-hour quota in 47 minutes on the $99 plan with a single session. Another said it lasts noticeably longer than before but still caps if you blast it. If you're building on it, use the pay-per-token API, not the plan.
• Identity weirdness. One tester saw the reasoning trace referring to itself as ChatGPT at times and mentioning "the evaluator." Probably distillation residue; irrelevant for most tasks, funny for the rest.
• Enterprise admins have it off by default on Kimi Business.

Should you switch?

My take:

• Agentic coding on a budget: yes, run an A/B this week. If it holds anywhere near the Terminal-Bench number, a 2x-5x cost cut matters at agent scale. Log tokens per solved task, not per request.
• Planner/worker split: one commenter uses Kimi to plan and DeepSeek to implement. That's a sensible pattern to steal.
• Regulated or high-stakes code: keep your frontier default until you've replicated results yourself. Vendor benchmarks are marketing until someone else reruns them.
• Self-hosting: wait for the weights. A ~3T-class MoE is not something you run on your workstation; per third-party listings it's about 2.8T total / 104B active, which still means a serious GPU cluster.

The bigger story isn't one leaderboard. Open-weight models at 3T scale are now inside a point or two of the closed frontier on agentic coding, at a fifth of the price. Whatever your moat is, "we have access to the best model" is not it anymore.

Run your own evals. Then decide.

Sources: Moonshot K3.1 Pro announcement, HN discussion, OpenRouter listing.