Head-to-head showdown: Cerebras — Qwen 3 32B ($0.40 in / $0.80 out per 1M) vs GPT-5.6 Luna ($0.20 in / $1.20 out per 1M). Cerebras — Qwen 3 32B is 1.2× cheaper across standard token mixes, with 128K vs 1.1M context windows.
qwen-3-32b-cerebras · cerebras
gpt-5.6-luna · openai
Benchmark
| Workload Scenario | Cerebras — Qwen 3 32B | GPT-5.6 Luna | Price Delta |
|---|---|---|---|
| 1M input tokens (raw text) | $0.40 | $0.20 | +$0.20 |
| 1M output tokens (generation) | $0.80 | $1.20 | −$0.40 |
| 1M tokens · 70% input / 30% output mix | $0.52 | $0.50 | +$0.02 |
| Standard chat turn (4K in / 800 out, 50% cached) | $0.00224 | $0.0014 | +$0.00084 |
| Monthly scale (10K requests / day) | $672.00 | $420.00 | +$252.00 |
Negative difference = Cerebras — Qwen 3 32B is cheaper. Positive = GPT-5.6 Luna is cheaper.
Scaling Curve
Total cost of a token volume at 70/30 input/output split (uncached). Log-log scale.
Verdict
GPT-5.6 Luna has the cheaper input rate, while Cerebras — Qwen 3 32B has the cheaper output rate. The crossover happens when output makes up about 33% of your total tokens.
Below that share (retrieval, summarization, extraction — lots of context in, little text out) GPT-5.6 Luna is cheaper. Above it (generation, translation, coding — long completions) Cerebras — Qwen 3 32B wins.
FAQ
On input, GPT-5.6 Luna is cheaper ($0.20/M vs $0.40/M). On output, Cerebras — Qwen 3 32B is cheaper ($0.80/M vs $1.20/M). For workloads where more than 33% of tokens are output, the output-cheaper model wins overall.
A chat-style request (4,000 input + 800 output tokens, 50% cached) costs $0.00224 on Cerebras — Qwen 3 32B and $0.0014 on GPT-5.6 Luna — GPT-5.6 Luna is 1.6× more expensive for that workload.
Cerebras — Qwen 3 32B supports 128,000 tokens (8.2K max output); GPT-5.6 Luna supports 1,050,000 (128K max output). GPT-5.6 Luna fits 8.2× more context, which matters for long documents and agents.