What happened
Moonshot released Kimi K3.1 Pro on 24 October 2026, making the model available through the OrcaRouter catalogue. The vendor publishes a rate card of $2.00 per million input tokens and $10.00 per million output tokens, with cache‑read pricing identical to Anthropic’s ($0.20 per million). By contrast, Claude Opus 5.5 is priced at $4.00 per million input and $20.00 per million output. Both models accept a 1 million‑token and expose a “reasoning effort” setting (max for Kimi, medium for Opus). In the only shared the vendors publish—Terminal‑Bench 4.0—Kimi scores 65.7 % versus Opus’s 66.4 %, a 0.7‑point gap measured under different effort settings. Moonshot also reports improvements over its predecessor: coding suite scores rose from 55.95 % to 61.2 %, CAD scores are 72.8 % versus Opus’s 73.0 %, and rates dropped from 23.4 % to 8.3 % with tools. The model’s weights are promised as open‑weight by 4 November 2026, though they are not yet publicly released.
Moonshot announced that Kimi K3.1 Pro is now live and can be accessed via OrcaRouter. The model’s rate card lists $2.00 per million input tokens and $10.00 per million output tokens, with cache‑read pricing of $0.20 per million tokens—identical to Anthropic’s cache‑read price. Cache‑write costs are also lower for Kimi ($2.00 for a five‑minute entry, $4.00 for an hour) compared with Anthropic’s $5.00 and $8.00 respectively.
Both Kimi K3.1 Pro and Claude Opus 5.5 accept a 1 million‑token and always perform a reasoning step before answering. Kimi defaults to its highest reasoning setting (max), while Opus defaults to medium. The models expose a reasoning‑effort field that users can adjust, but the default settings affect token usage and therefore cost.
In the only shared published by the vendors—Terminal‑Bench 4.0—Kimi scores 65.7 % and Opus scores 66.4 %. The scores were generated under different effort settings (Kimi at max, Opus at xhigh) and on vendor‑controlled harnesses, so the gap is not a pure measure of model quality. Independent testing by Artificial Analysis reported Opus at 59.6 % under a different configuration, highlighting the variability of benchmark results.
Moonshot also shared internal metrics: its long‑horizon coding suite improved from 55.95 % to 61.2 %, its CAD score is 72.8 % (versus Opus’s 73.0 %), and its rate on a hard set fell from 23.4 % to 8.3 % when tools are enabled. These figures suggest notable gains over the earlier Kimi K3 checkpoint.
Source details: orcarouter.ai ↗
Why it matters
The launch creates a direct, lower‑cost alternative to Anthropic’s flagship Opus tier for developers who need high‑throughput reasoning or agentic workloads. Because pricing is per token, the half‑price advantage translates into roughly 50 % lower compute spend for comparable token usage, assuming the models generate similar token volumes. The near‑par scores suggest that cost savings may be achieved without a large sacrifice in performance for many tasks, especially those measured by Terminal‑Bench. Open‑weight availability (expected 4 Nov) could further lower barriers for on‑premise deployments, giving enterprises a path to avoid vendor lock‑in. However, the comparison is complicated by differing default reasoning settings and vendor‑run benchmark methodologies, meaning real‑world performance will vary by workload. The shared routing key on OrcaRouter also means users can dynamically split traffic between the two models, leveraging price differentials while maintaining fallback capability.
Cost: At half the per‑token price, Kimi K3.1 Pro can reduce spend on high‑output workloads such as code generation, long‑form reasoning, or autonomous agents, provided token usage patterns are similar.
Performance parity: The sub‑one‑point difference on Terminal‑Bench indicates that developers may achieve comparable results without paying double, though workload‑specific testing is required because effort settings differ.
Open‑weight promise: Moonshot’s commitment to release the model weights by 4 Nov opens the possibility for on‑premise deployment, which is rare for a model positioned against a hosted‑only service like Opus. This could be a strategic advantage for regulated industries.
Operational flexibility: Both models are now reachable through the same OrcaRouter key, enabling dynamic traffic routing. Users can send cost‑sensitive requests to Kimi and reserve Opus for tasks where its longer evaluation history or fallback mechanisms are preferred.
Interactive Mechanism: How It Actually Works
Explore the underlying technology behind this development interactively.
Which component of an AI application is the machine-learning model itself?
What to watch next
Watch for the November 4 open‑weight release, which will let organizations host Kimi K3.1 Pro locally and potentially drive broader adoption. Independent results (e.g., from AI‑Analysis or other third parties) will be needed to confirm the claimed performance gap under comparable effort settings. Pricing adjustments from either vendor could shift the cost advantage, so monitoring rate‑card updates is essential. Finally, observe how developers configure the reasoning‑effort dial, as higher effort can erode the price benefit by increasing token consumption.
Open‑weight release on 4 Nov: Confirmation of weight availability and any licensing terms will affect adoption curves.
Third‑party benchmarks: Independent evaluations (e.g., from AI‑Analysis, MMLU, or other standard suites) will clarify whether the vendor‑reported scores hold under uniform settings.
Rate‑card changes: Both Moonshot and Anthropic may adjust token pricing; monitoring OrcaRouter listings will reveal real‑time cost dynamics.
Reasoning‑effort usage: How developers set the effort level in production will influence actual token consumption and thus the realized cost advantage.