Kimi K3 scored 57. How close is it to Claude and GPT? Moonshot AI has introduced Kimi K3, a model...Kimi K3 scored 57. How close is it to Claude and GPT? Moonshot AI has introduced Kimi K3, a model...
The network for creativity
Join 1.25M professional creatives like you
Connect with clients, get discovered, and run your business 100% commission-free
Creatives on Contra have earned over $150M and we are just getting started
Back to feedPost
Steve's avatar
Qorpia logo
Steve Castillo
• Jul 20
Kimi K3 scored 57. How close is it to Claude and GPT?
Moonshot AI has introduced Kimi K3, a model focused on coding, agents, and tasks that require maintaining context over long periods.
Its numbers are impressive:
- 2.8T parameters using a Mixture-of-Experts architecture.
- A 1-million-token context window.
- Text and image input.
- Support for large codebases and long-running workflows.
But the most interesting results do not come from Moonshot alone.
Artificial Analysis independently evaluated Kimi K3 and gave it a score of 57 on its Intelligence Index. In the original comparison, it finished close to some of today’s most advanced models:
- Claude Fable 5: 60
- GPT-5.6 Sol: 59
- Kimi K3: 57
- Claude Opus 4.8: 56
The index combines nine evaluations covering agents, coding, scientific reasoning, and general knowledge. It cannot represent every possible use case, but it provides more context than relying on a single benchmark.
Is it affordable too?
Not exactly.
Kimi K3 costs $3 per million input tokens and $15 per million output tokens. That makes it considerably more expensive than Kimi K2.6 and several alternatives whose complete model files can be downloaded.
However, the comparison changes when we look at Anthropic’s most advanced models:
- Kimi K3: $3 input and $15 output.
- Claude Opus 4.8: $5 input and $25 output.
- Claude Fable 5: $10 input and $50 output.
Per token, Kimi K3 costs approximately 40% less than Opus 4.8 and 70% less than Fable 5. Artificial Analysis also calculated an average cost of $0.94 per evaluated task, compared with $1.80 for Opus 4.8.
But price is not the whole story.
Artificial Analysis measured a 51% hallucination rate, up from 39% for Kimi K2.6. Moonshot also acknowledges that K3 may act too proactively when instructions are ambiguous.
Moonshot plans to make the complete trained model available for independent download and deployment, but those files are not available yet. The company says they will be released on July 27, 2026.
Kimi K3 does not outperform every leader, and it is not the cheapest option available. What makes it interesting is that it delivers performance close to Claude and GPT while remaining less expensive than Anthropic’s most advanced models.
Would you use it for coding today, or wait until the complete model becomes downloadable?
Sources: Moonshot AI · Artificial Analysis · Kimi pricing · Anthropic pricing
And remember: good software starts with good decisions. See you in the next one.
artificialintelligenceSoftware Developmenttechnews
Post image
Back to feed
The network for creativity
Join 1.25M professional creatives like you
Connect with clients, get discovered, and run your business 100% commission-free
Creatives on Contra have earned over $150M and we are just getting started