Baidu Ernie 4.5 21B A3b Thinking
Novita AI chatBaidu Ernie 4.5 21B A3b Thinking is a Novita AI chat model.It supports a 131,072-token context windowwith up to 65,536 output tokens.Input is priced at $0.0700/M tokens and output at $0.280/M tokens. Capabilities include reasoning. Route Baidu Ernie 4.5 21B A3b Thinking via Future AGI's Agent Command Center for unified observability, caching, and 15 routing strategies including cost-optimized fallback.
Estimate Baidu Ernie 4.5 21B A3b Thinking spend
Pick a workload, fine-tune the sliders, and see the monthly bill.
Estimate uses $0.0700/M input · $0.2800/M output. Provider pricing changes. Production costs vary with retries, streaming overhead, and tool-call rounds.
Want this for free? Cache + route via Agent Command Center — first 100K requests and 100K cache hits free every month.
Pricing
Per-token rates, expressed in USD per 1M tokens. Verified May 12, 2026.
| Input | $0.0700/M | |
| Output | $0.280/M |
Limits
- Context window
- 131,072 tokens
- Max input
- 131,072 tokens
- Max output
- 65,536 tokens
- Modalities
- text
Capabilities
- Function calling — not advertised
- Parallel tool calls — not advertised
- Vision input — not advertised
- Audio input — not advertised
- Audio output — not advertised
- PDF input — not advertised
- Streaming ✓ supported
- Structured output — not advertised
- Prompt caching — not advertised
- Reasoning ✓ supported
Where it's strong
- +multi-step reasoning and analysis tasks
Watch out for
- !high cost — input + output rates are in the top 88% of priced chat peers; consider a cheaper sibling for high-volume workloads
- !agentic workflows — no advertised function-calling; use a tool-capable model and route via Agent Command Center for fallback
- !strict structured output — no JSON-schema enforcement, expect retry loops
Benchmarks pending
We haven't logged public benchmark scores for Baidu Ernie 4.5 21B A3b Thinking yet. Have one to contribute? Submit a source — citations help us prioritise.
Call Baidu Ernie 4.5 21B A3b Thinking via Agent Command Center
One OpenAI-compatible endpoint. Routing, fallback, semantic caching, guardrails, and cost tracking come along for the ride. First 100K requests + 100K cache hits free every month.
agentcc / @agentcc/client). Per-call metadata — provider, cost, latency, cache hit, request id — is returned on x-agentcc-* response headers, so any HTTP client can read it.# Baidu Ernie 4.5 21B A3b Thinking via the Agent Command Center Python SDK
# pip install agentcc
import os
from agentcc import AgentCC
client = AgentCC(
api_key=os.environ["AGENTCC_API_KEY"], # from app.futureagi.com → Settings → API Keys
base_url="https://gateway.futureagi.com/v1",
)
resp = client.chat.completions.create(
model="novita-ai/baidu-ernie-4-5-21b-a3b-thinking",
messages=[{"role": "user", "content": "Hello, Baidu Ernie 4.5 21B A3b Thinking!"}],
)
print(resp.choices[0].message.content)
print(f"Tokens: {resp.usage.total_tokens}")
# Per-call gateway metadata is returned on x-agentcc-* response headers.
# When you need it programmatically, use .with_raw_response to get them:
raw = client.chat.completions.with_raw_response.create(
model="novita-ai/baidu-ernie-4-5-21b-a3b-thinking",
messages=[{"role": "user", "content": "Same call, but I want the headers."}],
)
print("Provider:", raw.headers.get("x-agentcc-provider"))
print("Latency:", raw.headers.get("x-agentcc-latency-ms"), "ms")
print("Cost: ", raw.headers.get("x-agentcc-cost"), "USD")
print("Cache: ", raw.headers.get("x-agentcc-cache"))AGENTCC_API_KEY with a key fromapp.futureagi.com.Gateway docs ↗Compare with similar models
Baidu Ernie 4.5 21B A3b Thinking doesn't have a public Arena ELO score yet, so we group by provider only — quality-tier comparisons need a benchmark.
- Baidu Ernie 4.5 VL 28B A3b ThinkingNovita AI · $0.390/M in · $0.390/M out · 131,072 ctx
- OpenAI GPT Oss 120BNovita AI · $0.0500/M in · $0.250/M out · 131,072 ctx
- Zai Org Glm 4.6vNovita AI · $0.300/M in · $0.900/M out · 131,072 ctx
- Qwen Qwen3 Omni 30B A3b ThinkingNovita AI · $0.250/M in · $0.970/M out · 65,536 ctx
FAQ
How much does Baidu Ernie 4.5 21B A3b Thinking cost?
Input is priced at $0.0700 per 1M tokens and output at $0.280 per 1M tokens (Novita AI, last verified May 12, 2026).
What is the context window of Baidu Ernie 4.5 21B A3b Thinking?
Baidu Ernie 4.5 21B A3b Thinking supports a 131,072-token context window with up to 65,536 output tokens.
Does Baidu Ernie 4.5 21B A3b Thinking support function calling?
Baidu Ernie 4.5 21B A3b Thinking does not currently advertise function-calling support. For agentic workloads, prefer a tool-calling-capable model and route via Agent Command Center for fallback.
Is Baidu Ernie 4.5 21B A3b Thinking good for production?
Baidu Ernie 4.5 21B A3b Thinking is well-suited for multi-step reasoning and analysis tasks. Consider alternatives if you need high cost — input + output rates are in the top 88% of priced chat peers; consider a cheaper sibling for high-volume workloads.
How can I route to Baidu Ernie 4.5 21B A3b Thinking with fallback?
Use Agent Command Center: a single OpenAI-compatible endpoint that supports cost-optimized routing, latency-aware retries, model fallback, and shadow traffic. Configure once, swap models without app changes.
Useful links for Baidu Ernie 4.5 21B A3b Thinking
Official sources, independent benchmarks, and pricing aggregators — no random search-engine guesses.
Third-party evals — verify the marketing.
Cross-check our number against the rest of the ecosystem.