Skip to main content
0 reactions

Fixed: “Unknown model name ‘claude-sonnet-5-5’, using ‘cl100k_base’ encoding as fallback” in Rasa

A designed card reading Rasa Fix: Unknown Model Name Warning, with a warning sign and two labels: shown with Claude models, and harmless, nothing to change.
Fixed: the unknown model name warning in Rasa

When a Rasa Pro agent uses Claude, rasa inspect keeps printing Unknown model name 'claude-sonnet-5-5', using 'cl100k_base' encoding as fallback: once when the server starts, and again with every message. Your agent still works. Here’s what the warning means, and why there’s nothing to fix.

What I Was Doing

I was building my first Mantle agent with Rasa Pro 3.21.4 and Claude Sonnet 5.5, following Rasa Tutorial — Part 2: Create Your First Mantle Agent. I started the agent with rasa inspect and chatted with it in the Rasa Inspector.

What I Expected

A server log with no warnings about my model.

What I Got

The warning appeared when the server started, then again with each message I sent:

Output (the lines that matter)Plain text
2026-11-18 23:33:44 INFO     root  - Starting Rasa server on http://0.0.0.0:5005
2026-11-18 23:33:45 WARNING  rasa.shared.utils.tiktoken_utils  - [warning  ] Unknown model name 'claude-sonnet-5-5', using 'cl100k_base' encoding as fallback.
2026-11-18 23:33:45 INFO     root  - Rasa server is up and running.
2026-11-18 23:39:17 WARNING  rasa.shared.utils.tiktoken_utils  - [warning  ] Unknown model name 'claude-sonnet-5-5', using 'cl100k_base' encoding as fallback.
2026-11-18 23:39:33 WARNING  rasa.shared.utils.tiktoken_utils  - [warning  ] Unknown model name 'claude-sonnet-5-5', using 'cl100k_base' encoding as fallback.
A Terminal window running rasa inspect: the server starts, and a warning says Unknown model name claude-sonnet-5-5, using cl100k_base encoding as fallback, once at startup and again with each message.
The unknown model name warning, at startup and with each message.

In the Inspector, the agent answered every message.

What I Checked

The warning names where it comes from: rasa.shared.utils.tiktoken_utils, a helper in Rasa Pro’s own code. Mantle uses it to count the tokens in each prompt, so the prompt stays within its budget, max_prompt_tokens in agent.yml, which is 8,000 tokens unless you change it (agent.yml reference (opens in a new tab)).

The helper counts with tiktoken (opens in a new tab), a tokenizer library that its README describes as “for use with OpenAI’s models”. It only knows OpenAI model names. In Rasa Pro 3.21.4, the helper adds a few more OpenAI names, and for any other name, such as a Claude model, it falls back to the general cl100k_base encoding and logs this warning.

Why It Happens

Rasa needs a rough count of each prompt’s size before it sends the prompt. tiktoken doesn’t know Claude’s model names, so Rasa estimates the size with cl100k_base instead, and says so every time it counts. The warning is about Rasa’s estimate on your computer, not about your connection to Claude.

The Fix

There’s nothing to change. Rasa Pro 3.21.4 has no setting that adds Claude to the names it knows, so the warning appears with every Claude model. Leave it, and keep using your agent.

After the Fix

With the warnings still in the log, the Inspector’s Inspect mode shows every LLM call going to claude-sonnet-5-5, a tool call, and the agent’s reply:

The Rasa Inspector conversation panel in Inspect mode: LLM calls to claude-sonnet-5-5 with their durations, a call to the get_recent_transactions tool, and the agent’s reply listing four transactions.
Every LLM call goes to claude-sonnet-5-5, and the agent replies normally.

Common Mistakes

MistakeDo this instead
Taking the warning for a failed Claude connectionCheck the Inspector: the replies arrive, and each LLM call names your Claude model.
Switching to an OpenAI model to silence itNot needed. The warning doesn’t change your agent’s replies.
Editing Rasa’s files inside .venv to add Claude’s nameLeave them as they are. Reinstalling Rasa Pro replaces them anyway.

FAQ

Does the warning cost money?

No. tiktoken runs on your computer, so counting a prompt doesn’t call Claude.

Does it appear with other Claude models?

Yes. It appears for any model name tiktoken doesn’t know, which includes every Claude model, such as claude-opus-5-5 and claude-haiku-5-5.

Why does it repeat with every message?

Rasa counts the prompt again for each message, and logs the warning each time it counts.

Conclusion

The “Unknown model name … using ‘cl100k_base’ encoding as fallback” warning means Rasa estimated a prompt’s size with a general tokenizer, because tiktoken doesn’t know Claude’s model names. Your agent keeps working with Claude, so you can leave it.

Additional Resources

Rasa, Claude, OpenAI and other names mentioned are trademarks of their respective owners. JD Bots isn’t affiliated with or endorsed by Rasa Technologies, Anthropic or OpenAI.

0 comments