Bug: Billing uses GPT-5.5 rates for budget models (Minimax M2.7, DeepSeek 4 Pro)

[aleksandrpalem-prog]

Summary

When using budget models (Minimax M2.7, DeepSeek 4 Pro), billing is calculated at GPT-5.5 rates instead of the selected model's rates.

Evidence

Chat	Model	Billed	Expected	Overcharge
pip dependency fix	Minimax M2.7	$9.68	$0.53	$9.15 (18x)
Open Interpreter	DeepSeek 4 Pro	$4.47	$0.49	$3.98 (9x)
Khoj setup	Mimo 2.5 Pro	$10.85	~$10.50	✅ Correct

Total overcharge: ~$13.13

Example calculation (Step 119, Minimax M2.7)

• Input: 144,602 tokens, Output: 168 tokens
• Billed: $0.96
• Expected (Minimax M2.7): $0.052
• GPT 5.5 rate: $0.96 ✅ exact match

Models API pricing (from /api/models)

Model	Input ($/M)	Output ($/M)	Tier
Minimax M2.7	$0.36	$1.44	$ (budget)
DeepSeek 4 Pro	$2.52	$5.28	$$
GPT 5.5	$6.00	$36.00	$$$

Impact

Budget models are being billed at flagship prices. Users selecting cheap models get unexpectedly high charges.

Environment

• OS: Windows 11
• Date: 2026-05-18/19
• Chat API response IDs: chatcmpl-* format (OpenAI-compatible)

Suggested fix

1. Use selected model's pricing for cost calculation
2. Add transparency: show which model's rates are being used
3. Warn when cost deviates significantly from expected

[debanjum]

Is this for Pipali or the Khoj project? Did the tasks you mention require the agent to read images? We auto fallback to use vision capable models like GPT or Kimi in such cases, this maybe why you're seeing the higher costs.

[aleksandrpalem-prog]

This is Pipali feedback.
My questions were not related to images and screenshots of the screen. My tasks are solved in the terminal or by text files.
For me, when the Minimax M2.7 model is enabled, GPT 5 is included in the costs.5. I do not know if I am the only one? Test it yourself

[aleksandrpalem-prog]

And even if Pipali needs to connect flagship AI models to solve the problem, is it possible to warn the user and request permissions? how is it implemented in other cases?

[debanjum]

Hey @aleksandrpalem-prog, thanks for the clarifications. I tested this on my end and noticed that the Minimax M2.7 model had become flaky. This makes Pipali auto fallback to use GPT, the default model. I've made changes to improve the reliability and speed of the Minimax model. This increases the cost of Minimax a bit but it works more consistent.

Try it out and let me know if it's working for your now?

[debanjum]

And even if Pipali needs to connect flagship AI models to solve the problem, is it possible to warn the user and request permissions? how is it implemented in other cases?

Yes, we should find some way to either warn the user if we switch to a more expensive model or fallback to a similar tier model instead. Let me think about the best way to deal with this. We'd want something that'll generalize well