Repository navigation
Conversation
… 3 rows in per-token units
|
|
Greptile SummaryCorrects inflated WandB model pricing by storing costs in dollars per token and updates two GPT-OSS models to current published rates.
Confidence Score: 5/5The PR appears safe to merge because both pricing registries consistently apply the corrected per-token values. The changed values use the registry’s dollars-per-token convention, documented active-model rates corroborate the corrections, and the runtime fallback copy remains synchronized with the primary registry.
|
| Filename | Overview |
|---|---|
| model_prices_and_context_window.json | Corrects 13 WandB pricing entries to per-token units, including current GPT-OSS rates; no actionable issues identified. |
| litellm/model_prices_and_context_window_backup.json | Mirrors the primary registry’s WandB pricing corrections exactly for bundled fallback use. |
Reviews (1): Last reviewed commit: "fix(wandb): 13 rows priced 100,000x too ..." | Re-trigger Greptile
|
Cross-check that doesn't need the vendor page:
Paths are |
Codecov Report✅ All modified and coverable lines are covered by tests. 📢 Thoughts on this report? Let us know! |
|
Rolled into the single rolling registry PR #37902 (maintainers keep one open |
What's wrong
13 of the 16
wandb/*rows store the price per 1M tokens divided by 10, in a field that is read as dollars per token.wandb/openai/gpt-oss-120bcurrently prices at0.015/token — $15,000 per 1M.The tell is inside the same provider block: three rows already use per-token units.
wandb/MiniMaxAI/MiniMax-M2.53e-07wandb/moonshotai/Kimi-K2.56e-07wandb/deepseek-ai/DeepSeek-V3.10.055Check that doesn't depend on trusting me
Divide the 13 rows by 100,000 and the ones W&B still lists land on the published rate exactly, to the cent — nothing rounded, nothing nudged (pricing, read 2026-08-23):
deepseek-ai/DeepSeek-V3.15.5e-07/1.65e-06meta-llama/Llama-3.1-8B-Instruct2.2e-07/2.2e-07meta-llama/Llama-3.3-70B-Instruct7.1e-07/7.1e-07Qwen/Qwen3-Coder-480B-A35B-Instruct1e-06/1.5e-06Four independent rows landing exactly is not what a wrong divisor looks like.
The remaining rows (GLM-4.5, both Qwen3-235B, DeepSeek R1-0528 and V3-0324, Llama-4-Scout, Phi-4-mini) are no longer on the pricing page, so they keep the price they were entered with, only in the right unit.
Two rows where the price itself also moved
gpt-oss-120bandgpt-oss-20bare the only rows where the unit fix alone would still be wrong — W&B has lowered both since they were added. They are set to today's published numbers: $0.03 / $0.17 and $0.03 / $0.13 per 1M. This is the only place a number comes from anywhere but the existing row. If you'd rather keep this PR mechanical, revert those four values and the rest still stands.Scope
Both copies of the registry, 13 rows, 26 values. No other provider, no key added or removed, no schema change.
Note on the earlier PRs
#37930 was closed as superseded by #37902. I re-read main today: the 13 rows are unchanged, and #37902's diff to the registry doesn't contain them. Re-filing against current main rather than reopening a branch that had drifted 86 commits behind.