Moonshot's Kimi K3.1 Pro Cuts Pricing 33% and Tops Terminal-Bench Among Open Models

Moonshot AI released Kimi K3.1 Pro today, October 23, posting a 4th-place finish on the Artificial Analysis Intelligence Index and reducing API pricing 33% below its predecessor — while scheduling open-weights release by November 4, twelve days from now. The model scores 55 on the AA Intelligence Index against Claude Opus 5.5's leading 58, a gap of three points that represents the narrowest distance between an open-weights system and the proprietary frontier that independent evaluation has recorded to date.

The release comes 99 days after Kimi K3 launched in July, and the architectural approach is deliberately conservative: the base mixture-of-experts structure is unchanged. What changed is the post-training pipeline — specifically, a substantially larger reinforcement learning compute budget applied across hundreds of downstream domains. The result is a model that independently benchmarks as the strongest agentic coding system currently available in any weight class.

A 2.8-Trillion-Parameter Base, Retrained From the Outside In

Kimi K3.1 Pro inherits the K3 architecture intact: 2.8 trillion total parameters organized as 896 experts, with 16 experts active per token during inference, yielding 104 billion active parameters per forward pass. The sparse expert routing is the core economic mechanism — because only 16 of 896 expert sublayers fire per token, inference cost scales with active parameters (104B) rather than total model size (2.8T). A dense model of equivalent total size would be prohibitively expensive to run.

What distinguishes K3.1 Pro from K3 is not architecture but post-training scale. Moonshot's technical blog describes "substantially larger post-training compute compared to K3" delivered through large-scale RL across hundreds of downstream task domains. The engineering improvements span the full RL stack: training framework, long-context inference efficiency, sandbox infrastructure for agent evaluation, rollout scheduling, and environment quality checks.

The model introduces a new "medium" reasoning effort level, joining the existing low, high, and max tiers. Developers can now choose four points on the cost-quality tradeoff curve rather than three. Maximum effort remains the default. One notable constraint: K3.1 Pro operates in thinking mode only — the API does not support disabling reasoning for lower-latency responses.

At 1,048,576 tokens of context, K3.1 Pro matches K3's 1M-token window, which remains one of the largest context offerings at this capability tier. The 1M context is reserved for Plus-tier subscribers on the Kimi consumer product, with the standard 128K window available for lower subscription tiers.

Terminal-Bench Leadership and the Independent Evidence That Matters

The benchmark result that carries the most weight is not Moonshot's own. Artificial Analysis, the independent AI evaluation firm, placed K3.1 Pro first among all measured models on Terminal-Bench 4.0 with a score of 67% — ahead of Claude Sonnet 5.5 (64%) and above Moonshot's own reported figure of 65.7% (vendor-run tests used the Claude-Code harness).

Terminal-Bench 4.0 evaluates agentic coding tasks executed in a real terminal environment, making it one of the more practically grounded software development benchmarks currently in use. The gap between Moonshot's self-reported 65.7% and Artificial Analysis's 67% is within normal harness variation; the directional finding is consistent — K3.1 Pro leads this benchmark. For context, vendor-reported figures from Moonshot place Claude Opus 5.5 at 66.4% on the same benchmark, meaning the two systems are essentially tied on agentic coding when measured independently.

On Artificial Analysis's broader Intelligence Index, K3.1 Pro scores 55, placing 4th globally behind Claude Opus 5.5 (58), Claude Sonnet 5.5 (56), and Muse Wonder Full (56). K3, the prior version, scored 44 on the same index — an 11-point gain across 99 days of post-training iteration.

On the GDPval-AA agentic Elo (Artificial Analysis), K3.1 Pro scores 1,846 — up 312 points from K3's 1,534, and 20 Elo points behind the leader Claude Opus 5.5 (1,866). AutomationBench-AA, which evaluates real SaaS platform automation, rose from 58% to 73%.

Moonshot's own benchmarks cover several vertical domains: BenchCAD 92.7% with tool use (approaching Claude Opus 5.5's 96.2%); OfficeQA Pro v2 financial document analysis at 64.4% (company-claimed state of the art); internal financial modeling improved from 50.4 to 57.6. The Kimi Hallu Bench Hard stress test shows the claim-level hallucination rate dropped from 23.4% for K3 to 8.3% for K3.1 Pro — a 64% reduction. That figure is company-reported on an internal benchmark, but the direction is consistent with RL post-training research on factual accuracy.

Pricing That Aligns With the Self-Hosted Economics Argument

Kimi API pricing for K3.1 Pro is $2.00 per million input tokens, $10.00 per million output tokens, and $0.20 per million cache-hit tokens — down from K3's $3.00/$15.00/$0.30. The 33% reduction applies consistently across all three pricing components.

A Highspeed variant (API ID: kimi-k3.1-pro-highspeed) runs at 2–3× the throughput of the standard model, priced at $4.00/$20.00/$0.20 per million tokens. The Highspeed tier is primarily suited for latency-sensitive production pipelines rather than cost-optimized batch workloads.

Artificial Analysis estimates the cost per intelligence task at approximately $2.11 for K3.1 Pro. The apparent paradox — 33% cheaper tokens with similar task cost to K3 — resolves at the token level: K3.1 Pro generates approximately 55,000 output tokens per complex agentic task versus K3's 48,000, a 14% increase consistent with more extended reasoning chains. Lower per-token price and higher token consumption offset each other.

Zero Data Retention is available for API customers who opt in, which addresses one dimension of the data-handling concern inherent to any Chinese-headquartered provider. Enterprise deployments on Kimi Work must have K3.1 Pro explicitly enabled by administrators — the model is off by default in enterprise configurations.

Where the Open-Weights Gap Has Arrived

When K3.1 Pro's weights release by November 4, it will become the highest-scoring openly available model on the AA Intelligence Index — by a margin that matters. K3, the current open-weights leader, scores 44 on that index. K3.1 Pro at 55 will sit 3 points below the best proprietary model (Claude Opus 5.5 at 58).

For high-volume deployments, the economics of that 3-point gap are meaningful. Self-hosting K3.1 Pro eliminates per-token API costs entirely. The 104B active parameter profile is manageable on enterprise multi-GPU clusters — MoE architectures have different memory footprint characteristics than dense models of the same active parameter count, but the practical hardware requirements are comparable. Enterprises choosing between frontier API access and self-hosting will face a narrower capability justification for paying the premium than at any prior point in the frontier era.

Deliberate Limits and Known Weaknesses

Moonshot AI's technical blog details the cybersecurity capability disclosure at length. The final K3.1 Pro weights score 81.0% on xdev-bench-hard pass@3, which evaluates real-world cybersecurity exploitation tasks. An intermediate version without the deliberate data-removal intervention scored 90.5% on the same benchmark — a capability level Moonshot describes as extending the "absolute frontier" of AI-assisted cyberoffense beyond publicly available models.

Moonshot's stated policy is not to release weights that extend the absolute frontier of cybersecurity capability. The company applied data removal at both mid-training and RL stages to reduce the final model's exploit performance. The gap from 90.5% to 81.0% is intentional and explicitly acknowledged. K3, for comparison, scored 33.3% on the same benchmark — so K3.1 Pro at 81.0% still represents a substantial improvement, just not the unconstrained maximum.

The business operations benchmark improvement deserves the same scrutiny applied to very large percentage gains. K3 scored 8.9 on Moonshot's internal Business Operation evaluation; K3.1 Pro scores 37 — approximately 4.2× higher. Dramatic relative gains from low baselines are common in RL post-training and indicate strong improvement, but the absolute K3 score suggests the prior model's business workflow capability was limited enough that proportional gains may overstate practical readiness.

All vendor-reported coding benchmarks, including the Terminal-Bench 65.7% figure, were run using the Claude-Code harness. This is a methodological limitation Moonshot acknowledges: the evaluation infrastructure itself is built around Claude's tool-calling patterns, which may not be fully neutral across models. Artificial Analysis's independent Terminal-Bench figure (67%) is the cleanest available data point for cross-model comparison.

Read more: Anthropic names Moonshot in AI reasoning-theft defense

Open Weights by November 4 and What Follows

The K3.1 Pro open weights are scheduled to arrive by November 4, 2026 — 12 days from today. K3's weights followed its API release by 11 days; K3.1 Pro matches that cadence, suggesting Moonshot is institutionalizing a cycle: launch the post-trained model at API, then release weights within two weeks.

The 11-point Intelligence Index gain across 99 days of post-training work represents a meaningful improvement rate without changing the underlying architecture. If that iteration rate holds, the practical question for the field is not whether K3.1 Pro is good today, but what K3.2 or K4 looks like when a similar RL compute budget is applied again.

The 3-point proprietary-to-open gap on the AA Intelligence Index reflects a real convergence between what commercial APIs can offer and what the open ecosystem can deploy. Within 12 days, developers will be able to download and run a system Artificial Analysis ranks within 5% of the global frontier — and whether that gap closes further depends on whether Moonshot's RL infrastructure can keep outpacing proprietary post-training programs.