Granite 3.3 8B (Non-reasoning)
AI model by IBM. Real-time pricing and benchmark data.
Benchmarks
Editorial Analysis
Profile. Granite 3.3 8B (Non-reasoning) is a budget-tier option from IBM at $0.25/M output tokens. Despite the low price, it scores competitively on measurable benchmarks — making it a candidate for high-volume production workloads where unit cost dominates.
Latency. On latency, Granite 3.3 8B (Non-reasoning) reports 14 tok/s output speed and 28160ms time-to-first-token. Output speed is on the slower end; for real-time chat consider pairing with a streaming front-end or routing latency-sensitive requests to a faster model.
Cost model. For a workload of 1M input tokens + 100K output tokens, Granite 3.3 8B (Non-reasoning) costs $0.06 per million input-equivalent requests. Multiply by your monthly request volume to project the bill, and compare against the speed/correctness profile above to decide if the trade-off is worth it for your workload.
Editorial summary generated from public pricing and benchmark data (Artificial Analysis API). All numbers are sourced — see the Artificial Analysis methodology for benchmark definitions. Last refreshed 2026-08-24.
Compare with similar models
| Model | Input | Output | Speed |
|---|---|---|---|
Granite 3.3 8B (Non-reasoning)Current | $0.03 | $0.25 | 14 tok/s |
| Qwen3.5 9B (Non-reasoning) | $0.17 | $0.25 | 87 tok/s |
| Mistral 7B Instruct | $0.25 | $0.25 | 122 tok/s |
| Llama 2 Chat 7B | $0.05 | $0.25 | — |
| Nova Lite | $0.06 | $0.24 | 151 tok/s |
| Nemotron 3.5 Lightning | $0.07 | $0.22 | 302 tok/s |