Kimi K2.5 versus Kimi K3

Two models head to head

This comparison pits two open-weight models that you can install on your own servers against each other: Kimi K2.5 and Kimi K3, both built by Moonshot AI. The first dates from January 27, 2026, the second from July 16, 2026.

All figures below come from public sources checked on October 22, 2026: models.dev and OpenRouter for prices and capabilities, Epoch AI for scores.

Performance

On our capability index, Kimi K2.5 scores 78.1 out of 100 and Kimi K3 80.7. Kimi K3 leads by 2.6 points.

On the 2 benchmarks reported for both, Kimi K2.5 wins 0 times and Kimi K3 2:

• GPQA Diamond (PhD-level science questions): Kimi K2.5 87.6%, Kimi K3 93.1%
• OTIS Mock AIME (competition mathematics problems): Kimi K2.5 92.2%, Kimi K3 97.2%

Which one costs less?

Kimi K2.5 costs $0.58 per million input tokens and $3 per million output tokens; Kimi K3 costs $3 and $15.

For a typical agent task (read 20,000 tokens, write 2,000), expect $0.018 with Kimi K2.5 versus $0.09 with Kimi K3. Kimi K2.5 is 5.1 times cheaper, a gap that adds up quickly over 1,000 tasks: $72.4 difference.

Run the numbers with your own volumes in the agent cost calculator.

Context and capabilities

Kimi K3 reads more in one go: 1,048,576 tokens of context (roughly 1,573 pages), versus 262,144 for Kimi K2.5.

Reasoning mode: Kimi K2.5 yes, Kimi K3 yes. Tool calling: Kimi K2.5 yes, Kimi K3 yes. Image input: Kimi K2.5 yes, Kimi K3 yes.

Both publish their weights: you can host them yourself.

In short

Our reading of the data:

• Choose Kimi K2.5 for a lower cost.
• Choose Kimi K3 for a higher capability index, a larger context window and a more recent release.

This verdict relies on public figures; it does not replace a trial on your own tasks.