What Baidu released
ERNIE-Light-DT-1.0 is a Baidu model, and two sellers carry it so far: Baidu itself and NanoGPT. Its release date comes from a third-party model page, not from Baidu, so treat it with some caution.
**ERNIE Light DT 1.0 is a model from Baidu**, released on Nov 10, 2026.Most of what is known about the model's purpose comes from the lab. In the description quoted on OrcaRouter's blog, Baidu stresses how the model keeps track of what came earlier in a session. No independent test in our sources has checked that claim.
Baidu says the model "improves the storage and indexing of historical information" through feedback-driven history compression, "significantly reduces...Those claims seem aimed at agents that carry a long history from one step to the next. For that workload, the price list is worth reading line by line.
Where is it cheapest?
Two sellers, one model, and a modest spread. Baidu is the designated cheapest seller, meaning the one our site names as the lowest-priced offer for the model, at $1.8325 on the mixed price.
| Seller | Input, $/M | Output, $/M | Mixed 3:1, $/M |
|---|---|---|---|
| Baidu | 1.13 | 3.94 | 1.8325 |
| NanoGPT | 1.30 | 4.50 | 2.10 |
Baidu lists $1.13 for Input and $3.94 for Output, while NanoGPT asks $1.30 for Input and $4.50 for Output. On the mixed price, NanoGPT's offer works out at 1.1 times Baidu's, and the sources give no reason for the gap.
For an agent, the line used most is often the cache read: Input the seller has already seen, billed at a lower price. Baidu charges $0.113 for it, and that rate shapes the cost of any workload that re-sends a long context.
What does an agent task cost?
A rate per million tokens says nothing about how much the model writes, and an agent's bill depends on exactly that. Take one agent step: 120,000 tokens of context re-read from the cache, 30,000 tokens of new request, and a written answer of 15,000 tokens.
At Baidu, that step costs $0.107, and a run of 1,000 of them comes to $106.56. At NanoGPT, the same thousand steps cost $122.10.
| Seller | One task, $ | 1,000 tasks, $ | Output share, % |
|---|---|---|---|
| Baidu | 0.107 | 106.56 | 55 |
| NanoGPT | 0.122 | 122.10 | 55 |
The surprise is where the money goes. The answer is the smallest of the three token counts, yet Output makes up 55% of the cost at both sellers. The cache does its job on the long context: at Baidu, those 120,000 tokens are billed at $0.113 per million rather than $1.13.
Two levers therefore act on the bill. One is the length of the answers, which carries 55% of it, so a cap on Output tokens in the prompt or the request matters. The other is the seller, which moves a thousand tasks from NanoGPT at $122.10 to Baidu at $106.56.
How long a context does ERNIE-Light-DT-1.0 accept?
Up to 1,024,000 tokens, at Baidu as at NanoGPT. Both sellers open the full window, so picking one does not mean giving up context length; an agent that needs the whole window gets it at either seller's rates.
What does a cache read cost on ERNIE-Light-DT-1.0?
Baidu bills a cache read at $0.113, and NanoGPT at $0.13. Both rates sit under each seller's Input price, but only Input the seller has already seen qualifies, so keep the long part of the context identical from step to step and put new material after it.
On InferIndex: ERNIE-Light-DT-1.0 model page