Summary
CodeBurn shows $0.00 for 27 models that are not in LiteLLM's pricing database, totaling 599.6M input tokens and 1.9M output tokens of unpriced usage. These appear as "$0" rows in the dashboard and distort the total cost (which is understated).
The affected models fall into three categories:
- Thinking/reasoning variants (
:thinking suffix) — e.g. glm-4.7:thinking, glm-5:thinking, kimi-k2.6:thinking
- Cloud variants (
:cloud suffix) — e.g. deepseek-v4-pro:cloud, minimax-m3:cloud
- TEE variants and other model IDs not in LiteLLM — e.g.
GLM-4.7-TEE, Kimi-K2.5-TEE, MiMo-V2-Flash, KAT-Coder-Pro-V1
Evidence
Top 10 unpriced models (all show $0.00)
| Model |
Input tokens |
Output tokens |
CodeBurn cost |
GLM-4.7-TEE |
389.4M |
541K |
$0.00 |
glm-4.7:thinking |
126.4M |
678K |
$0.00 |
MiMo-V2-Flash |
32.0M |
115K |
$0.00 |
Kimi-K2.5-TEE |
8.4M |
30K |
$0.00 |
qwen3-max-2026-01-23 |
7.6M |
80K |
$0.00 |
deepseek-v4-pro:cloud |
5.9M |
55K |
$0.00 |
glm-5:thinking |
5.4M |
31K |
$0.00 |
KAT-Coder-Pro-V1 |
4.3M |
21K |
$0.00 |
kimi-k2.6:thinking |
3.9M |
125K |
$0.00 |
mimo-v2-flash:free |
2.8M |
10K |
$0.00 |
Total: 27 unpriced models, 599.6M input + 1.9M output tokens
Suggested improvements
1. Suffix stripping (like the cx/ prefix fix)
Many of these models are variants of known models with a suffix that LiteLLM doesn't index:
:thinking → strip to get base model (e.g. glm-5:thinking → glm-5 → $1.40/$4.40 per 1M):cloud → strip to get base model (e.g. deepseek-v4-pro:cloud → deepseek-v4-pro → $0.43/$0.87 per 1M)-TEE → strip to get base model (e.g. GLM-4.7-TEE → glm-4.7 → known price)
2. Fuzzy matching fallback
For models like MiMo-V2-Flash, KAT-Coder-Pro-V1, qwen3-max-2026-01-23 that don't have a simple suffix relationship, a fuzzy match against LiteLLM keys (e.g. mimo-v2-flash → minimax/mimo-v2-flash) would catch most cases.
3. User-configurable aliases
CodeBurn already has codeburn model-alias for manual mapping, but it requires the user to know every unpriced model. A built-in fallback that auto-strips common suffixes (:thinking, :cloud, -TEE, -free) before giving up would reduce the manual burden significantly.
Environment
Related
Summary
CodeBurn shows $0.00 for 27 models that are not in LiteLLM's pricing database, totaling 599.6M input tokens and 1.9M output tokens of unpriced usage. These appear as "$0" rows in the dashboard and distort the total cost (which is understated).
The affected models fall into three categories:
:thinkingsuffix) — e.g.glm-4.7:thinking,glm-5:thinking,kimi-k2.6:thinking:cloudsuffix) — e.g.deepseek-v4-pro:cloud,minimax-m3:cloudGLM-4.7-TEE,Kimi-K2.5-TEE,MiMo-V2-Flash,KAT-Coder-Pro-V1Evidence
Top 10 unpriced models (all show $0.00)
GLM-4.7-TEEglm-4.7:thinkingMiMo-V2-FlashKimi-K2.5-TEEqwen3-max-2026-01-23deepseek-v4-pro:cloudglm-5:thinkingKAT-Coder-Pro-V1kimi-k2.6:thinkingmimo-v2-flash:freeTotal: 27 unpriced models, 599.6M input + 1.9M output tokens
Suggested improvements
1. Suffix stripping (like the
cx/prefix fix)Many of these models are variants of known models with a suffix that LiteLLM doesn't index:
:thinking→ strip to get base model (e.g.glm-5:thinking→glm-5→ $1.40/$4.40 per 1M):cloud→ strip to get base model (e.g.deepseek-v4-pro:cloud→deepseek-v4-pro→ $0.43/$0.87 per 1M)-TEE→ strip to get base model (e.g.GLM-4.7-TEE→glm-4.7→ known price)2. Fuzzy matching fallback
For models like
MiMo-V2-Flash,KAT-Coder-Pro-V1,qwen3-max-2026-01-23that don't have a simple suffix relationship, a fuzzy match against LiteLLM keys (e.g.mimo-v2-flash→minimax/mimo-v2-flash) would catch most cases.3. User-configurable aliases
CodeBurn already has
codeburn model-aliasfor manual mapping, but it requires the user to know every unpriced model. A built-in fallback that auto-strips common suffixes (:thinking,:cloud,-TEE,-free) before giving up would reduce the manual burden significantly.Environment
~/.local/share/opencode/opencode.db(SQLite, 1.5 GB)Related
cx/prefix stripping was already fixed forcx/gpt-5.5but not for combinedcx/gpt-5.5-xhigh(tier suffix)