Best LLMs for Classification and Content Moderation
Tiny outputs at very high volume — the cheapest capable model wins.
Typical request this page prices
- Quality floor: none — every priced chat model is compared.
- Costs use each provider’s current standard rates, with cached tokens at the cached-input rate where one is published. Speed is from Artificial Analysis where measured.
Lowest cost that meets the bar
Models that clear the quality floor, ordered by estimated cost for the request above.
| Model | Intelligence | Speed | Cheapest host | Input / output per 1M | Per 1,000 requests |
|---|---|---|---|---|---|
| Ling-3.0-flash-VL InclusionAI | 24.6 | 147tok/s | OpenRouter 3 providers | $0.021 / $0.062 cached $0.0042 | $0.0085 |
| Ling 3.0 Flash InclusionAI | 20.1 | - | OpenRouter 3 providers | $0.021 / $0.063 cached $0.0042 | $0.0085 |
| GPT-OSS-20B OpenAI · 128K context | 10 | 285tok/s | OpenRouter 12 providers | $0.018 / $0.090 cached $0.0090 | $0.0095 |
| Granite 4.0 H Micro IBM · 131K context | - | - | OpenRouter 1 provider | $0.017 / $0.112 | $0.012 |
| Llama 3.2 1B Meta · 128K context | 4.8 | - | Novita AI 5 providers | $0.020 / $0.020 | $0.012 |
| Llama 3.1 8B Meta · 128K context | 6.9 | - | Novita AI 5 providers | $0.020 / $0.050 | $0.013 |
| Qwen3.7 Flash Alibaba | - | - | OpenRouter 2 providers | $0.030 / $0.130 cached $0.0060 | $0.013 |
| Granite 4.2 3B IBM | 9.1 | 227tok/s | Deep Infra 1 provider | $0.030 / $0.120 cached $0.0075 | $0.013 |
Highest-scoring options
Ordered by Artificial Analysis Intelligence Index, with the cost of the same request.
| Model | Intelligence | Speed | Cheapest host | Input / output per 1M | Per 1,000 requests |
|---|---|---|---|---|---|
| Claude Opus 5.5 Anthropic | 57.6 | 97tok/s | Deep Infra 6 providers | $4.00 / $20.00 cached $0.200 | $1.56 |
| Claude Sonnet 5.5 Anthropic | 56 | 135tok/s | Anthropic 6 providers | $2.00 / $10.00 cached $0.100 | $0.780 |
| Claude Fable 5.1 Anthropic | 53.4 | 72tok/s | DigitalOcean 4 providers | $10.00 / $50.00 cached $0.250 | $3.82 |
| GPT-6 Astra OpenAI | 52.7 | 63tok/s | Velokey 7 providers | $3.00 / $15.00 cached $0.300 | $1.21 |
| GPT-6.1 Sol OpenAI | 51.8 | 65tok/s | Microsoft Azure 6 providers | $2.00 / $10.00 cached $0.100 | $0.780 |
| Claude Opus 5 Anthropic | 50.8 | - | Velokey 7 providers | $2.00 / $10.00 cached $0.200 | $0.810 |
| Claude Fable 5 Anthropic · 1.0M context | 49.6 | - | Velokey 6 providers | $4.00 / $20.00 cached $0.400 | $1.62 |
| Muse Spark 1.3 Meta | 48.1 | 115tok/s | OpenRouter 2 providers | $1.25 / $4.25 cached $0.150 | $0.484 |
Compare every model and provider on the LLM API pricing table.
How to choose an LLM for classification & moderation
Classification, routing, sentiment and moderation calls send a short instruction plus one item and get back a label. Output is a handful of tokens, so the input rate is effectively the price. These workloads run at millions of calls a day, which makes small per-token differences material.
There is no quality floor on this list: the right model is the cheapest one that reaches acceptable accuracy on your labelled samples, and small models often do. The instruction and label set repeat on every call, so the estimate assumes half the input is cached.
Latency also matters when classification sits in a request path, for example routing a user message before answering it. Compare time to first token where it is shown.
Related Reading
GPU Marketplaces Are Cheapest, Until They Aren't
We analyzed the marketplace-style providers tracked by ComputePrices to find where marketplaces actually beat traditional GPU clouds, where they do not, and what the cheapest listing really buys.
Top 8 Cheapest Cloud GPU Providers for AI Training
Explore affordable cloud GPU providers that offer competitive pricing and performance for AI training tasks, ideal for developers and businesses.
Frequently Asked Questions
Is an LLM the right tool for classification at all?
For a fixed label set with plenty of training data, a fine-tuned small classifier is usually cheaper and faster. LLMs are useful when labels change often, training data is scarce, or the categories need judgement that is hard to encode.
Why does this list include models with no intelligence score?
Because no quality floor applies here. Small models that benchmark providers have not scored can still classify well, and excluding them would hide some of the cheapest options.
How should I evaluate models for moderation?
Use a labelled sample that includes the borderline cases you care about, and measure false negatives and false positives separately. A model that is cheaper per call but needs a second pass on borderline items may cost more overall.