Beijing, October 24, 2026 – Moonshot AI has unveiled its latest large language model, Kimi K3.1 Pro, with a significant reduction in standard token pricing. The company slashed input token costs by one-third and output token costs by one-third as well, a move that could make the model more attractive for high-volume applications. However, independent testing reveals that the overall cost of running benchmark tasks remains roughly unchanged from its predecessor, Kimi K3.
According to a report from Artificial Analysis, the Kimi K3.1 Pro consumes more output tokens per task than its predecessor. The Intelligence Index task—a standardized benchmark—now generates approximately 55,000 output tokens, compared to 48,000 for Kimi K3. This increase in token consumption largely offsets the price reduction, keeping the average task cost at $1.98, essentially flat year-over-year.
The distinction between unit pricing and total cost is crucial for developers and enterprises that deploy AI agents for complex coding, data analysis, or autonomous workflows. While the per-token price drop is welcome, the overall expenditure depends on how many tokens the model generates and processes for each job. For coding agents that iterate through lengthy tasks and generate reasoning traces, the higher token output can diminish the financial benefit of the lower rates.
Pricing Breakdown
The new pricing structure for Kimi K3.1 Pro is straightforward: $2 per million input tokens and $10 per million output tokens, as listed in Vercel's AI Gateway catalog and independently confirmed on OpenRouter. These rates apply to standard first-party access, excluding cache discounts. For uncached input and output, the Highspeed version costs twice as much: $4 per million input and $20 per million output.
Compared to the previous Kimi K3 model, which was priced at $3 per million input tokens and $15 per million output tokens, the new rates represent a 33% reduction across the board. Yet, the average cost per task remains stable because the model now generates more tokens to achieve its results.
Performance Gains Concentrated in Autonomous Computer Use
Artificial Analysis's testing highlights that the most dramatic improvements are in tasks involving autonomous computer use. The Terminal-Bench 4.0 score jumped from 13% for Kimi K3 to 67% for Kimi K3.1 Pro, a substantial leap. The broader Intelligence Index, which measures overall model capability, rose 11 points to 55.
These gains suggest that Moonshot AI focused its post-training efforts on enhancing the model's ability to interact with computer environments, execute commands, and complete multi-step tasks without human intervention. This is a key area for AI agents that need to navigate software, manipulate files, or manage workflows.
Cost Analysis: Output Component
Using rounded output token counts, the generated-text component of a typical Intelligence Index task falls from $0.72 to $0.55—a reduction of about 24%. This calculation excludes input tokens, cache charges, and other components that contribute to the total task cost. The overall task cost remains flat because the increased token output offsets the lower per-token price.
For enterprises planning to deploy Kimi K3.1 Pro at scale, the flat total cost per task may be a more important metric than the headline token price reduction. The model's improved performance on complex tasks could justify the cost, but the financial impact will depend on the specific use case and the volume of tokens generated.
Technical Details and Availability
Moonshot AI describes Kimi K3.1 Pro as a post-trained version of Kimi K3, utilizing additional compute after the initial pretraining phase. The model retains a one-million-token context window, which is critical for processing long documents or codebases. The launch report does not specify the exact amount of additional training compute used.
The company's reinforcement learning efforts span hundreds of application domains, with announced improvements in rollout scheduling, inference, and training infrastructure. Moonshot AI is targeting programming, 3D creation, and financial research with the updated model, positioning it as a versatile tool for professional use.
According to Kimi Code documentation, both new models require thinking mode; requests with thinking disabled are rejected. Users can select low, high, or max reasoning effort for their workloads. The documentation also warns that switching models within an existing session reduces cache hits, which could affect performance and cost.
Distribution of Kimi K3.1 Pro extends beyond Moonshot's own products. Vercel added both models to its AI Gateway on October 23, supporting existing API formats such as Chat Completions, Responses, and Anthropic Messages. OpenRouter also offers hosted access, making the model widely available to developers.
Risks and Considerations
Benchmark results depend on task design, reasoning settings, and cache behavior. Production retries and real-world usage patterns may alter costs further. Moonshot has promised to release the full model weights by November 4, but this remains a schedule target. Independent operators will need to evaluate what they can deliver once the weights are available.
For now, the Kimi K3.1 Pro represents a strategic move by Moonshot AI to compete on price while pushing the boundaries of autonomous AI capabilities. The flat task costs suggest that the company is balancing affordability with performance, a delicate equation in the rapidly evolving AI landscape.
