Article

Kimi K3.1 Pro Explained: The Chinese Open-Weight Model Nipping at Claude Opus 5.5

Kimi K3.1 Pro Explained: The Chinese Open-Weight Model Nipping at Claude Opus 5.5

On October 23, 2026, Beijing's Moonshot AI released Kimi K3.1 Pro, the first model in its new K3.1 series. Moonshot calls it the company's "most capable 3T-class open-weight model to date," and its own numbers put the model within a whisker of the closed frontier: 65.7% on Terminal-Bench 4.0, versus 66.4% for Anthropic's Claude Opus 5.5 and 58.2% for OpenAI's GPT-6 Astra. The catch is that the weights aren't public yet — Moonshot says full weights land by November 4 — and the developers on Chinese forums, who are usually the first to adopt domestic models, spent most of their energy arguing that it costs too much.

This is one of those launches where the launch post and the community reaction tell two different stories. Both are worth reading, and they answer different questions: the blog tells you what the model can do, the forums tell you whether anyone will pay for it.

What Moonshot actually shipped

K3.1 Pro is not a new base model. Moonshot's tech blog is explicit: it is a post-trained version of Kimi K3, with "substantially larger post-training compute" — much more reinforcement learning, run across what the company calls "hundreds of downstream domains" — on the same underlying model. The specs, per Moonshot's launch announcement and tech blog:

  • 1-million-token context window, enough to hold an entire codebase in one pass
  • 2.8 trillion total parameters, 104 billion active per token (mixture-of-experts), with multimodal text-and-image input
  • Two flavors: Kimi K3.1 Pro, and Kimi K3.1 Pro Highspeed, which Moonshot says runs 2–3× faster with the same capabilities
  • Live now across the Kimi product family: Kimi.ai, the Kimi mobile apps, Kimi Work (desktop app, v3.1.0+), Kimi Code, and the Kimi API
  • Open weights promised by November 4, 2026 — today, the API is the only way to run it

One unusual detail stands out in the launch post. Moonshot devotes an entire section to cybersecurity: during development, the team deliberately identified and removed cyber-related data from both mid-training and reinforcement learning, because it wants the open weights to not push the frontier of offensive hacking capability. After the cleanup, the model scores 81.0% on Moonshot's internal xdev-bench-hard cyber benchmark, versus 90.5% for a checkpoint that skipped the data removal. It is rare for a lab to publish a "we made the model weaker on purpose" table in a launch blog, and the transparency is notable.

The benchmark story, with honest caveats

The headline number is Terminal-Bench 4.0, a test that puts coding agents on realistic terminal tasks: 65.7% for Kimi K3.1 Pro, against 66.4% for Claude Opus 5.5 and 58.2% for GPT-6 Astra. Moonshot also claims 61.2% on its internal Kimi Code Bench 3.0 long-horizon test (up from 55.95% for K3), 92.7% on the BenchCAD 3D-modeling benchmark (behind Opus 5.5's 96.2% and GPT-6 Astra's 95.9%), a state-of-the-art 64.4 on OfficeQA Pro v2 for financial document work, and a drop in claim-level hallucination rate from 23.4% to 8.3% on its internal stress test.

Independent measurement exists too. Artificial Analysis scores Kimi K3.1 Pro at 55 on its Intelligence Index — behind only Claude Opus 5.5 (58), Claude Sonnet 5.5 (56), and Muse Wonder Full (56), and eleven points above Kimi K3, per The Decoder's coverage of the ranking. If the weights land on November 4 as promised, it would be the highest-scoring open-weight model available, shrinking the open-to-closed gap from twelve points to three.

Read the fine print before getting excited. Every score above is Moonshot's own, and the company evaluated its model inside Anthropic's Claude Code harness while taking rivals' numbers from their published reports — which is standard practice but not independent verification. Several of the strongest claims (the finance tests, the hallucination numbers) come from internal benchmarks nobody outside Moonshot can run. As one widely-shared hands-on writeup notes, Artificial Analysis found the model unusually verbose, which means the effective cost per completed task can be higher than the per-token price suggests. Benchmark tables are marketing material until independent evaluators reproduce them. Moonshot's look strong, and the AA index score is genuinely independent.

What it costs

The pricing is the part that turned the community reaction sour. Per Moonshot's API price list, as widely reported by independent outlets: $2 per million input tokens and $10 per million output tokens, with cached input at $0.20 — about a third cheaper than Kimi K3's $3/$15. The Highspeed variant doubles those rates ($4/$20), with cached input staying at $0.20.

Put next to the closed frontier, that is aggressive: roughly half of Claude Opus 5.5's $4/$20 per million tokens, and a fraction of GPT-6 Astra's reported $10/$50. Moonshot's pricing strategy is unchanged from the Kimi K3 playbook — undercut the Western labs on price while chasing them on quality.

But the token math and the subscription math are different things, and Chinese developers noticed. Kimi's app subscriptions are sold in yuan-denominated tiers (the discussion centers on the ¥699 tier), and forum members complain that the tier's usage quota feels smaller than a $20 Western plan — one popular line of argument holds that the same money buys roughly five Claude Pro subscriptions' worth of capacity with more quota each. Since I could not verify Moonshot's current plan quotas against a primary source, treat that as community sentiment rather than fact — but the sentiment is loud and consistent across threads.

How Chinese developers actually reacted

A launch announcement posted to LINUX DO's frontline-news section on October 23 drew quick engagement (thread), and the first thing the commenters did was not celebrate — it was audit. Members asked which coding benchmarks the model was actually tested on, and the thread's most upvoted exchange was a small seminar on benchmark culture: with classic coding benchmarks like SWE-bench saturated and gamed, several members argued that Terminal-Bench is now the only coding-agent test that still carries weight. That is a striking verdict on the state of model evaluation: a community of working developers has independently arrived at "ignore the marketing tables, look at one number."

The second reaction was price. In the announcement thread and a larger follow-up discussion (thread, 44 posts and over 2,000 views), members called the model genuinely impressive and the pricing genuinely too high. One thread's framing question, translated loosely, was: "a day after launch, why are there so few hands-on test posts — is the price the reason?" Developers compare the ¥699 tier unfavorably with Western $20 plans, grumble that plan quotas are too small for serious agent work, and joke that they'll wait for a hypothetical Kimi K4 instead.

The cultural color is worth noting. One member crowned it "国模一姐" — roughly "the top Chinese model" — while immediately asking how it actually feels in day-to-day use, a nickname that captures the whole mood: national pride in the benchmark chart, and a wait-and-see attitude toward the product. Nobody is disputing that Moonshot built something impressive. The dispute is whether the subscription value proposition matches the benchmark story.

How to try it

The fastest free route is the Kimi app or kimi.ai, where the model is selectable in the chat interface now. For developers:

  • API: OpenAI-compatible endpoint with model IDs kimi-k3.1-pro and kimi-k3.1-pro-highspeed. Hands-on reports indicate access unlocks after a minimum $1 top-up, thinking mode is always on, and reasoning effort defaults to max. Moonshot's docs note temperature and top-p are fixed, so leave them out of your requests.
  • Kimi Code: select the model with the /model command. Regular K3.1 Pro reportedly needs a Plus plan or higher; Highspeed needs Pro or higher.
  • Enterprise: admins must enable the model — it is off by default at launch.
  • Weights: wait until November 4. Self-hosting a 2.8-trillion-parameter model needs data-center-class hardware, so for most people "open weights" means third-party hosted inference getting cheaper, not running it on a home server.

A minimal API call follows the standard OpenAI shape:

curl https://api.moonshot.ai/v1/chat/completions \
  -H "Authorization: Bearer $MOONSHOT_API_KEY" \
  -H "Content-Type: application/json" \
  -d '{
    "model": "kimi-k3.1-pro",
    "messages": [{"role": "user", "content": "Explain this codebase in one page."}]
  }'

(Replace the endpoint and key with your own; the base URL above is from an independent hands-on report, and Moonshot documents the model IDs on its official tech blog.)

The real test is November 4

Strip away the marketing tables and the forum grumbling, and Kimi K3.1 Pro is a genuinely important release for two reasons. First, it is another data point that post-training — not just bigger base models — is where the frontier is moving: Moonshot closed most of the gap to Opus 5.5 with RL compute alone. Second, the November 4 weight release would put near-frontier coding and agent capability into anyone's hands, the same kind of shock the DeepSeek release delivered in early 2025 (as we covered for DeepSeek's flash model).

The community's price complaints are the market speaking, and Moonshot would be wise to listen — benchmarks win headlines, but developers who switched to Claude Pro and back have shown they will pay for quota they can actually feel. If the weights land on schedule and independent tests confirm the scores, the price argument gets weaker: open weights make the API price irrelevant for anyone with the hardware or a cheap hosted provider. Until then, K3.1 Pro is an impressive API product with a subscription tier Chinese developers are not yet convinced about — and a November date that is now the one to watch.

Keep reading
1024 Programmers' Day: Why China's Developers Take October 24 Off
Read next →