ERNIE-4.5-21B-A3B-Thinking
baidu/ernie-4-5-21b-a3b-thinkingPar baidu · famille: ernie · sorti 2025-09-19 · fin de connaissance: 2025-03
Prices in USD per 1M tokens. Unknown means the provider does not publish per-token pricing.
Capacités
Model fit scores
0–100 · higher is betterThese scores reward declared capabilities, context size, price and provider availability — they are not benchmark results. Use them as a directional signal alongside your own evaluation.
Coding14
- Tool calling0/40
- Structured output0/20
- Reasoning10/10
- Context window (100K → 1M)2/20
- Provider availability2/10
Agents32
- Tool calling0/35
- Structured output0/25
- Reasoning15/15
- Output token limit15/15
- Provider availability2/10
JSON / structured output29
- Structured output / JSON mode0/50
- Tool calling0/20
- Temperature control10/10
- Price-friendly for high-volume19/20
Cost efficiency74
- Headline price (log-scaled)74/95
- Has prompt-cache pricing0/5
Long context46
- Context window (100K → 2M)36/90
- Has published price for full window10/10
Production-readiness70
- Number of independent providers10/40
- Has published per-token price20/20
- Context window ≥ 8K15/15
- No data inconsistencies across providers10/10
- Official model (not derivative)15/15
Cost Efficiency Index
Open full calculator →Estimated cost using the recommended provider's headline rate. Each scenario fixes average input/output tokens — the assumptions are shown in the third column.
| Scenario | Cost | Assumption |
|---|---|---|
RAG answer per 1,000 RAG answers | $0.49 < $0.01 per request | 5K input tokens (query + 4 retrieved chunks of ~1K each) and a 500-token answer. Typical SaaS knowledge-base bot. |
Support ticket triage per 10,000 tickets | $0.98 < $0.01 per request | 1K input tokens (ticket body + system prompt) and a 100-token JSON classification reply. High-volume customer support. |
Data extraction per 1,000 documents | $0.28 < $0.01 per request | 2K input tokens (a single document page) and a 500-token JSON extraction. ETL / invoice / form pipelines. |
Code review per 1,000 PRs | $0.84 < $0.01 per request | 8K input tokens (diff + surrounding files) and a 1K-token review comment. PR-bot workloads. |
Agent step per 1,000 steps | $1.01 < $0.01 per request | 12K input tokens (long-running tool history) and a 600-token tool-call decision. Cost per agent step. |
Détail des tarifs
Tarif recommandé de novita-ai · baidu/ernie-4.5-21B-a3b-thinking
Disponible chez 2 fournisseurs
| Fournisseur | ID modèle fournisseur | Entrée / 1M | Sortie / 1M | Contexte | Publié le |
|---|---|---|---|---|---|
| NovitaAI novita-ai | baidu/ernie-4.5-21B-a3b-thinking | $0.070 | $0.280 | 131K | 2025-09-19 |
| Kilo Gateway kilo | baidu/ernie-4.5-21b-a3b-thinking | $0.070 | $0.280 | 131K | 2025-09-19 |
Frequently asked questions
How much does ERNIE-4.5-21B-A3B-Thinking cost?
ERNIE-4.5-21B-A3B-Thinking costs $0.070 per 1M input tokens and $0.280 per 1M output tokens, sourced from novita-ai. Cache reads, audio tokens and >200K-context tiers (where applicable) are listed in the Pricing detail block above.
What is the context window of ERNIE-4.5-21B-A3B-Thinking?
ERNIE-4.5-21B-A3B-Thinking has a context window of 131K tokens, with a max output of 66K tokens per reply. This is the total combined size of prompt + completion.
Does ERNIE-4.5-21B-A3B-Thinking support tool calling?
No. ERNIE-4.5-21B-A3B-Thinking does not support tool calling (function calling). If your workflow requires it, look at the /capabilities/tool-calling list for alternatives.
Does ERNIE-4.5-21B-A3B-Thinking support structured output / JSON mode?
Support for structured output / JSON-schema-constrained decoding is not reported for ERNIE-4.5-21B-A3B-Thinking in our data source. Verify with baidu's official documentation before relying on it in production.
Can ERNIE-4.5-21B-A3B-Thinking accept image input?
No. ERNIE-4.5-21B-A3B-Thinking only accepts text as input. If you need image input, see our /capabilities/vision list for current vision-capable models.
Is ERNIE-4.5-21B-A3B-Thinking open-weight?
Yes. ERNIE-4.5-21B-A3B-Thinking's weights are publicly available, so you can self-host or fine-tune. Note that open weights ≠ open source — the training data and code are typically not released.
What are the best alternatives to ERNIE-4.5-21B-A3B-Thinking?
If ERNIE-4.5-21B-A3B-Thinking doesn't fit, consider ERNIE 4.5 300B, ERNIE 4.5 VL 28B, ERNIE 4.5 VL 424B A47B. Each one targets the same use case — see the Related links below for direct head-to-head pages.
Where does this data come from?
All numbers come from the public models.dev API and are normalised into a single canonical model record. We re-pull daily and write any changes (price, context, capability) to the /changelog page.
Explore more
More baidu models
- ERNIE 4.5 300B$0.28 in / $1.10 out
- ERNIE 4.5 VL 28B$0.14 in / $0.56 out
- ERNIE 4.5 VL 424B A47B$0.42 in / $1.25 out
- ERNIE 4.5 300B A47B$0.28 in / $1.10 out
- ERNIE 4.5 21B A3B$0.07 in / $0.28 out
Capability lists this model is in
Dernière mise à jour :
Data is sourced from models.dev and normalized for comparison. Prices and capabilities may change. Always verify critical production decisions with the provider's official documentation.