Introducing Kojo, our latest model
Kojo scores 92.2% on SWE-bench Verified and 86% on Terminal-Bench 2.1. You can chat with it for free at kojo.sh, and its API at api.kojo.sh starts at $0.175 per million input tokens.

Kojo is our latest model. Ahead of the launch we opened it up for free at kojo.empero.org, which is now kojo.sh. Since then it has handled billions of tokens a day at a very low serving cost. Within a few hours, about 500 accounts signed up just to farm free tokens. They cost us about $3.
Today Kojo gets its own home at kojo.sh, its API opens to everyone at api.kojo.sh, and there are plans for people who want more.
Two benchmarks
Full evaluation suites were prohibitively expensive for us, so we picked the two benchmarks with the best signal per dollar.
| Benchmark | Kojo |
|---|---|
| SWE-bench Verified | 92.2% (461 of 500) |
| Terminal-Bench 2.1 | 86% |
How we use it
Internally we have been running Kojo alongside an Opus 5.5 advisor, and two things stood out. First, its performance relative to its price was better than we expected. Second, Opus 5.5 often did not appear in the logs at all: Kojo was handling the task from start to finish without consulting the advisor, and doing it well.
That is why we now use Kojo to build the infrastructure that serves our models, and why it has been involved in training others: mainly a text-to-speech model we hope to release soon, and an architecture for local use that should train on minimal compute.
What Kojo does in the chat
- Searches the web and reads the pages it finds, so its answers use current information.
- Reads the images you attach.
- Builds things: web pages, small games, drawings and art, with a live preview in the chat.
- Works the way you do: sign in with an Empero account to give it your own instructions and skills (SKILL.md files), and to have your chats on every device.
- Shares chats: send a chat as a link, and anyone can continue it as their own.
Kojo is free at kojo.sh, without an account. Signing in keeps your chats on every device and gives you more room each day.
The API
Kojo speaks the OpenAI chat completions API, so existing SDKs and tools work once you change two settings, the base URL and the key:
from openai import OpenAI
client = OpenAI(base_url="https://api.kojo.sh/v1", api_key="sk-kojo-...")
reply = client.chat.completions.create(
model="kojo",
messages=[{"role": "user", "content": "Hello, Kojo."}],
)
print(reply.choices[0].message.content)
Streaming, tool calls and images work as they do with other OpenAI-compatible services. You create keys, follow your usage and read the documentation at platform.kojo.sh.
Pay as you go costs $0.175 per million input tokens, $1.00 per million output tokens and $0.02 per million cached input tokens. That is a launch price for two weeks from 26 October. After that it is $0.20, $1.25 and $0.02.
Plans
Choose a plan in Kojo under Settings, then Plan, or at platform.kojo.sh. You can pay by card or from your Empero balance.
| Plan | Price | Usage | API keys |
|---|---|---|---|
| Free | $0 | Chat with a daily allowance | None |
| Go | $1 a month | Starter usage | 1 |
| Basic | $5 a month | 5× Go | 2 |
| Plus | $15 a month | About 20× Go | 5 |
| Pro | $35 a month | About 35× Go | 5 |
| Max | $50 a month | 50× Go | 10 |
| Flex | Pay as you go | At the prices above | No limit |
Moving from kojo.empero.org
kojo.empero.org now forwards to kojo.sh, and your chats and sign-in come with you. If you use the API, change the base URL to https://api.kojo.sh/v1. Your keys stay the same.
What comes next
Longer term, we plan subscriptions for the API models that follow Kojo. More on that when the time comes.
Made by Kojo
Kojo drew every image below in code, as SVG and JavaScript, with no image model involved. Select one to see it larger.
Thanks for reading. Try Kojo or get an API key.











