Claude Fable 5 and Kimi K3 compared. What still matters?

As of 4 November 2026. If a production integration calls claude-fable-5 or kimi-k3, read the identifier carefully. Both models remain documented. Claude Fable 5.1 and Kimi K3.1 Pro are now available too. Reading the original summer comparison as a current leaderboard would put the wrong versions head to head.

What has changed

Anthropic’s API catalogue marks Fable 5 active, legacy. Existing applications can continue to call it. Anthropic points developers to Fable 5.1 for migration. Moonshot has delivered the previously promised K3 weights, with its own model card and licence. Kimi Code announced K3.1 Pro on 23 October. A download of K3 therefore says nothing about whether you can download K3.1 Pro. Fable 5 catalogue entry · Moonshot’s model card · Kimi Code release history

Long context is useful when a task needs many sources. It does not guarantee the model will notice the decisive footnote. In a contract review, score the model on the relevant clause, the exception that limits it, and whether it correctly says that a term is absent. Anthropic’s specifications · Kimi API documentation

An API cost calculation

Anthropic lists Fable 5 at USD 10 per million input tokens and USD 50 per million output tokens. Moonshot quotes USD 3 per million uncached input tokens and USD 15 per million output tokens for K3. Assume 50,000 input tokens and 5,000 billed output tokens. The token component comes to USD 0.75 and USD 0.225 respectively. This illustration excludes taxes, separate cache writes, tools, batch discounts and retries. A completed task may take multiple calls. Fable pricing · K3 rates · Kimi cache and billing rules

Caching also matters to existing customers. Anthropic lists the same standard input and output rates for Fable 5.1 but a lower cache-read price. An agent that carries the same source documents through many steps may therefore produce a different bill from a one-off question. Anthropic’s Fable 5.1 guidance

What the weights let you do

Moonshot reports 2.8 trillion parameters for K3. A rough lower bound puts that number in perspective. If each parameter occupied exactly four bits, the weights alone would take about 1.4 terabytes in decimal units. Runtime state and memory for long inputs come on top. This is a scale illustration rather than a hardware specification. A team may have the right to run the model itself and still find a dedicated cluster uneconomic. K3 technical description · Deployment and licence

Decide on the workload you actually have

Choose three real tasks with known answers. Give the earlier model and the candidate the same inputs and tool permissions. Have someone assess the outputs without seeing which model produced them. Log retries, elapsed time and corrections. The figure that matters to a team is cost per accepted result. Repeated attempts can erase a token saving. Higher spend can be justified when it demonstrably saves review time. How AI costs add up