Best LLMs for Structured Data Extraction
Turning emails, invoices and forms into JSON at volume.
Typical request this page prices
- Quality floor: Top 60% of priced models by Intelligence Index (≥ 11.9).
- 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 | $0.101 |
| Ling 3.0 Flash InclusionAI | 20.1 | - | OpenRouter 3 providers | $0.021 / $0.063 cached $0.0042 | $0.102 |
| Qwen3.5 4B Alibaba | 13.1 | 20tok/s | Empirio Labs 2 providers | $0.040 / $0.070 cached $0.020 | $0.178 |
| Ling-3.0-flash-Fin InclusionAI | 22.6 | 317tok/s | OpenRouter 3 providers | $0.042 / $0.123 cached $0.0084 | $0.202 |
| Mercury 2.5 Inception | 12.3 | 833tok/s | OpenRouter 1 provider | $0.040 / $0.150 cached $0.0040 | $0.207 |
| Gemma 4 26B Google · 262K context | 16.7 | 101tok/s | OpenRouter 7 providers | $0.042 / $0.220 cached $0.021 | $0.275 |
| DeepSeek V4 Flash DeepSeek · 1.0M context | 34.3 | - | Deep Infra 19 providers | $0.060 / $0.180 cached $0.015 | $0.294 |
| Nemotron 3.5 Lightning NVIDIA | 12.9 | 290tok/s | Crusoe 3 providers | $0.050 / $0.200 cached $0.030 | $0.296 |
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 | $23.44 |
| Claude Sonnet 5.5 Anthropic | 56 | 135tok/s | Anthropic 6 providers | $2.00 / $10.00 cached $0.100 | $11.72 |
| Claude Fable 5.1 Anthropic | 53.4 | 72tok/s | DigitalOcean 4 providers | $10.00 / $50.00 cached $0.250 | $58.30 |
| GPT-6 Astra OpenAI | 52.7 | 63tok/s | Velokey 7 providers | $3.00 / $15.00 cached $0.300 | $17.76 |
| GPT-6.1 Sol OpenAI | 51.8 | 65tok/s | Microsoft Azure 6 providers | $2.00 / $10.00 cached $0.100 | $11.72 |
| Claude Opus 5 Anthropic | 50.8 | - | Velokey 7 providers | $2.00 / $10.00 cached $0.200 | $11.84 |
| Claude Fable 5 Anthropic · 1.0M context | 49.6 | - | Velokey 6 providers | $4.00 / $20.00 cached $0.400 | $23.68 |
| Muse Spark 1.3 Meta | 48.1 | 115tok/s | OpenRouter 2 providers | $1.25 / $4.25 cached $0.150 | $6.23 |
Compare every model and provider on the LLM API pricing table.
How to choose an LLM for structured data extraction
Extraction pipelines send a schema, a few examples and one document, and expect a JSON object back. The instructions and examples repeat on every call, so part of the input is cacheable; the document itself is not. The estimate assumes about a third of input is a cached prefix.
Accuracy is measured against your own schema, not a general benchmark. A model with a modest intelligence score but reliable structured-output support often beats a higher-scoring model that occasionally emits malformed JSON. The quality floor here is deliberately below the median so smaller, faster models stay in the comparison.
At high volume the per-request cost is what matters. Small differences in output rate add up when every document returns several hundred tokens of JSON.
Frequently Asked Questions
How do I pick a model for extraction without a benchmark for my schema?
Label a few dozen representative documents, run the cheapest few models from this list against them, and compare field-level accuracy. Move up the quality list only if the cheaper models miss fields you need.
Does structured output mode cost more?
Providers generally bill structured or JSON-mode output at the normal token rates. The cost difference comes from output length, so keep schemas compact and avoid asking for explanations alongside the data.
Can I use batch APIs for extraction?
Yes, if documents can be processed asynchronously. Batch rates are lower than standard rates at most providers that offer them; check each model’s page for current batch pricing.