Kimi K3.1 Pro: Opus-Class Agent Work at Half the Price

Moonshot AI released Kimi K3.1 Pro on Oct 23, 2026. The launch blog calls it the first model in the K3.1 series and says it "matches or approaches the strongest proprietary models." Quote that as a claim. Then look at Moonshot's own table: against Claude Opus 5.5 the record is 21 wins and 21 losses over 42 shared rows, with a mean gap of -3.4 points. The wins are small. The losses are large.

What holds: agentic and office work (mean +0.7 vs Opus 5.5 over 17 point-scale rows), Terminal-Bench 4.0 at 65.7 (0.7 behind Opus 5.5 at 66.4, 7.5 ahead of GPT-6 Astra at 58.2), and a list price of $2 / $10 per 1M tokens, half of Opus 5.5 and a fifth of Astra and Claude Fable 5.1. What does not: HLE-Diamond, Terminal-Bench Science, Program Bench, and the exploit benchmarks. All scores are self-reported by Moonshot.

For prior Kimi context see Kimi K3 vs Claude Fable 5. Peer launch posts on this site include Claude Opus 5.5, GPT-6 Astra, and Claude Fable 5.1.

At a Glance

• Maker: Moonshot AI
• Released: Oct 23, 2026
• Series: first model in the K3.1 series; a post-trained version of Kimi K3with "substantially larger post-training compute" (Moonshot)
• Highspeed variant:2x the price and "2-3x the speed" (Moonshot); no tokens-per-second figure published
• Context:1,048,576 tokens (1M). Text and image input (Moonshot's docs also describe video input). Output text.
• Thinking: always on. Effort low / high / max; default max.
• Price: $2.00 / $0.20 / $10.00 per 1M (input / cached input / output), flat; a third below K3
• Access:Moonshot's API, Kimi.ai, Kimi Work, Kimi Code, plus listings on Vercel AI Gateway and OpenRouter
• Weights: open weights promised by Nov 4, 2026, not yet released, license not stated
• Table record: 16 of 47 rows led outright; 21 wins and 21 losses vs Opus 5.5. Scores self-reported by Moonshot; none independently verified by third parties or by LLM Stats.

What K3.1 Pro Is

Moonshot describes consistent gains across hundreds of RL domains from a larger post-training run, aimed at coding, 3D creation and CAD, financial research, and long-horizon agent work. Moonshot's long-horizon example is a walking-dynamics simulator built over 167 missions and 420 reviews.

Moonshot's own internal comparisons vs K3 (Moonshot's words and numbers; there is no K3 column in the launch table): Kimi Code Bench 3.0 Long-horizon 55.95 to 61.2; finance modeling 50.4 to 57.6 (+7.2); financial research 58.8 to 70.4 (+11.6); business operation 8.9 to 37, "about 4.2x"; hard hallucination stress test 23.4% to 8.3% claim-level (Moonshot cautions this is a stress test, not a daily-use error rate).

On cyber: Moonshot says it removed cyber-related data from mid-training and RL for the open-weight release, and that K3.1 Pro's offensive cyber capability "remains behind that of frontier proprietary models." Its internal xdev-bench-hard moved from 33.3 (K3) to 81.0 (90.5 for the checkpoint without cyber risk reduction).

Benchmarks

Every score below is self-reported by Moonshotfrom the launch comparison table. None of these scores has been independently verified by third parties or by LLM Stats. Peer scores are as published in Moonshot's table and may differ from each developer's own reports. K3.1 Pro runs at max reasoning effort, temperature 1.0, top-p 0.95. Several rows run competitor tooling (Claude Code, Codex): treat that as a reason to rerun on your own stack. Moonshot says these three scores come from a third-party source: Automation Bench-AA, GDPval-AA v2.1, AA-Briefcase v1.1.

Moonshot compares with Claude Fable 5.1, GPT-6 Astra, Claude Opus 5.5 (the three closed models), DeepSeek-V4.1-Flash-1017 and GLM-5.3 (the two open models). The table has no Kimi K3 column. GPT-6.1 Sol lists at the same $2 / $10 per 1M tokens (GPT-6.1 Sol launch) but is not in Moonshot's table; no score comparison here.

Where it leads

• BrowseComp: 92.1 (best in the table; +1.3 vs Opus 5.5).
• HLE-Full w/ tools: 70.0 (+2.3 vs Opus 5.5).
• Agents' Last Exam (ALL): 43.2 and CLI: 44.2.
• Automation Bench-AA: 73.3 (+3.8 vs Opus 5.5).
• OfficeQA Pro v2: 64.4(Moonshot calls it "state-of-the-art").
• GDP.pdf w/ python: 85.7, HANDBOOK.md: 44.6, Harvey Legal Agent: 23.7 (public 1,251-task set, all-pass rate), MMMU-Pro w/ python: 90.0, ZeroBench w/ python: 63.0 pass@5, GDP.pdf: 34.6, SWE Marathon v1.1: 51.7.
• Internal leads (Moonshot-built): Kimi Code Bench 2.0, Business Operation, DECK-Arena.

Where it trails

• Program Bench: Opus 5.5 leads by 33.5 (65.0 vs 31.5).
• Terminal-Bench Science 0.1: Opus 5.5 leads by 25.2 (63.3 vs 38.1); Astra leads the row at 68.1.
• HLE-Diamond: Opus 5.5 leads by 21.5 (60.9 vs 39.4); Astra leads the row at 65.8.
• ExploitBench: Opus 5.5 leads by 16.3 (91.0 vs 74.7); Astra leads the row at 100.
• ExploitGym: Opus 5.5 leads by 13.8 (34.5 vs 20.7).
• HLE-Full: Opus 5.5 leads by 8.7 (59.8 vs 51.1).
• PostTrain Bench v1.2: Opus 5.5 leads by 7.3; Fable 5.1 leads the row at 44.6.
• APEX-Agents: Opus 5.5 leads by 6.3 (73.5 vs 67.2).
• OSWorld V2.1: Opus 5.5 leads by 2.2; BioMysteryBench by 1.3; GDPval-AA v2.1 1846 vs 1866; AA-Briefcase v1.1 1678 vs 1803.

Head to head on rows both report: vs Claude Opus 5.5 21 / 0 / 21 (external only 16 / 0 / 21); vs GPT-6 Astra 23 / 0 / 19; vs Claude Fable 5.1 27 / 0 / 13; vs DeepSeek-V4.1-Flash-1017 42 / 0 / 2 (DeepSWE and the internal xdev-bench-hard); vs GLM-5.3 32 / 0 / 0. K3.1 Pro leads 16 of 47rows outright (13 of 39 once Moonshot's 8 internal benchmarks are set aside). Of those 16 leads, 13 are agentic or office work.

By job: agentic and office is the win (mean +0.7 vs Opus 5.5, +4.4 vs Astra, +2.9 vs Fable on point-scale rows both report). Coding is split: it leads SWE Marathon and is within 0.7 of Opus on Terminal-Bench 4.0, but averages -6.8 vs Opus and -5.5 vs Astra over 9 and 8 rows because of Program Bench, Terminal-Bench Science and PostTrain Bench. Knowledge and vision trails the closed models (HLE-Full 51.1 vs Opus 59.8; HLE-Diamond 39.4 vs Astra 65.8; GPQA 94.2 sits above Opus 93.9 but below Astra 96.1). Cyber trails by design. Against the open peers it is not close.

Eight of the 47 rows are Moonshot's own internal benchmarks (it leads 3, trails 5). Mark those as internal wherever they appear: Kimi Code Bench 2.0, Kimi Code Bench 3.0 Long-horizon, Business Operation, DECK-Arena, Design Bench, both Kimi Hallu Bench rows, and xdev-bench-hard. On factuality (internal, lower is better): Kimi Hallu Bench Hard 8.3% vs Opus 15.0% and Astra 5.5%; without tools 5.4% vs Opus 6.8% and Astra 4.2%: second to Astra on both.

Pricing and Tiers

Moonshot's flat first-party price is $2.00 / $0.20 / $10.00 per 1M (input / cached input / output). Versus Kimi K3($3.00 / $0.30 / $15.00), K3.1 Pro is exactly two thirds of K3's price on every line: a third cheaper. See Moonshot's pricing page.

Peer prices are from each vendor's own pricing page on Oct 23, 2026. GPT-6 Astra applies a long-prompt surcharge over 272K input tokens: 2x input, 1.5x output. For 1M input plus 1M output tokens: K3.1 Pro $12, Highspeed $24, Opus 5.5 $24, Astra and Fable $60. Highspeed's price equals Opus 5.5's standard price, while Opus 5.5 fast mode is 2x Highspeed and GPT-6 Astra Fast is 5x Highspeed.

Terminal-Bench 4.0 against price: 65.7 at $2 / $10 vs Opus 5.5 66.4 at $4 / $20 vs GPT-6 Astra 58.2 and Claude Fable 5.1 57.9 at $10 / $50. Per-token price is not cost per task, and K3.1 Pro at max effort may use more tokens. Moonshot published no token-usage figures.

Weights and Access

Moonshot calls it "our most capable 3T-class open-weight model to date" and says "The full model weights will be released by Nov 4." As of Oct 23, 2026 there are no public weights, no Hugging Face repo, no GitHub repo, and Moonshot has not stated a license. Frame it honestly: open weights promised by Nov 4, 2026, not yet released, license not stated. Today this is an API model. As precedent, not as a promise: Kimi K3's weights are on Hugging Face under a custom Kimi K3 license that adds a separate-agreement requirement for large model-as-a-service businesses, and Moonshot has not said K3.1 Pro will use the same terms.

Where it runs today: Moonshot's API, Kimi.ai, the Kimi apps (iOS, Android, HarmonyOS), Kimi Work desktop 3.1.0 or later (Windows and Apple silicon Macs), Kimi Code (the model picker), Kimi Business, Vercel AI Gateway and OpenRouter (both list Moonshot AI as the provider at the same prices). Zero Data Retention is available for API customers (Moonshot). Specs and effort controls are in Moonshot's K3.1 Pro quickstart.

Should You Switch?

From K3:K3.1 Pro is a third cheaper and Moonshot's own internal results improve on finance, long-horizon coding and factuality. Moonshot published no external-benchmark deltas against K3, so test on your own tasks. Thinking is always on and the default effort is max (the slowest and most token-hungry setting), so for latency lower the effort level or use Highspeed.

Standard vs Highspeed: pay 2x for 2 to 3x the speed on interactive or human-in-the-loop agents; stay on standard for batch and background work. Cached input costs the same on both.

Pick K3.1 Pro for:long-horizon agent, research and document or finance workflows (BrowseComp, Agents' Last Exam, OfficeQA Pro v2, Automation Bench-AA, HLE w/ tools), where it matches or beats Opus 5.5 at half the price and GPT-6 Astra at a fifth.

Prefer Claude Opus 5.5 or GPT-6 Astra for: hardest-reasoning science questions (HLE-Diamond, GPQA-style work), program synthesis (Program Bench), terminal science tasks, vulnerability research, and perception-heavy vision (PerceptionBench, ZeroBench); also Opus 5.5 where APEX-Agents or OSWorld-style computer use matters.

Test GLM-5.3 or DeepSeek-V4.1-Flash-1017 if price per token dominates: both list far cheaper, and DeepSeek wins DeepSWE, but K3.1 Pro wins every other shared row against GLM-5.3 and all but one external row against DeepSeek. If you need weights today, K3.1 Pro is not that model yet. Evaluate on your own workload and harness.

FAQ

Outlook

On Moonshot's own table, Kimi K3.1 Pro is an Opus 5.5-class agent and document model at half Opus's list price. It is not a frontier-knowledge or exploit model. The 21-to-21 split with Opus 5.5 and the -3.4 mean gap make that plain: the agentic and office block holds, the hard science and exploit rows do not. The price is the point. The weights are only promised by Nov 4 with no license stated, so today this is a $2 / $10 API model. All scores self-reported by Moonshot.

Open questions: whether the weights ship by Nov 4 and under which license; independent reruns on neutral harnesses (especially Terminal-Bench 4.0, where a 0.7-point gap on one benchmark is thin evidence either way, and the internal benchmarks Moonshot designed); whether Moonshot publishes token-usage and Highspeed throughput numbers; how far the "K3.1 series" goes beyond Pro; how Opus 5.5 and GPT-6 Astra respond on price.