Best LLMs for Structured Output Extraction
Extracts structured data from text into a specified JSON schema. Pure shape-conformance — no enrichment, no rephrasing, no summarisation. Used when a downstream consumer needs schema-clean data from prose.
Models
Frontier on this task: Gemini 3.1 Pro Preview at 9.84 / 10. Quality bar at 90%: 8.86.
point-estimate floor (CI low) · upper CI (less certain) · Bars sorted by blended cost; best-value model first. Greyed rows are MEDIUM+ models whose point estimate clears the bar but whose CI low does not.
| Model | Quality score | CI low | Cost / 1k runs | vs best value |
|---|---|---|---|---|
| DeepSeek V4 Flash | 9.54 / 10 | 9.40 | $0.40 | best value |
| MiniMax M3 | 9.46 / 10 | 9.30 | $5.29 | 13x more expensive |
| GPT-5.4 Nano | 9.46 / 10 | 9.22 | $7.61 | 19x more expensive |
| DeepSeek V4 Pro | 9.54 / 10 | 9.40 | $11.52 | 29x more expensive |
| GPT-5.6 Luna | 9.51 / 10 | 9.27 | $12.15 | 30x more expensive |
| Qwen 3.7 Plus | 9.31 / 10 | 9.17 | $22.47 | 56x more expensive |
| Qwen 3.6 Flash | 8.99 / 10 | 8.49 | $23.13 | 58x more expensive |
| Grok 4.5 | 9.23 / 10 | 9.04 | $29.30 | 73x more expensive |
| Qwen 3.6 Plus | 9.65 / 10 | 9.47 | $37.25 | 93x more expensive |
| Gemini 3.5 Flash | 9.70 / 10 | 9.64 | $42.96 | 107x more expensive |
| Kimi K2.6 | 9.71 / 10 | 9.61 | $55.73 | 139x more expensive |
| Gemini 3.1 Pro Preview | 9.84 / 10 | 9.74 | $62.16 | 155x more expensive |
| GPT-5.6 Sol | 9.56 / 10 | 9.32 | $62.47 | 156x more expensive |
| GPT-5.5 | 8.96 / 10 | 8.55 | $205.70 | 513x more expensive |
| Claude Haiku 4.5 | 8.28 / 10 | 7.90 | $30.67 | 77x more expensive |
| Claude Sonnet 4.6 | 8.66 / 10 | 8.26 | $105.33 | 263x more expensive |
| GPT-5.4 Mini | 7.94 / 10 | 7.52 | $18.48 | 46x more expensive |
| Meta Muse Spark 1.1 | 8.80 / 10 | 8.33 | $38.98 | 97x more expensive |
| GPT-5.6 Terra | 8.60 / 10 | 8.11 | $33.58 | 84x more expensive |
Cost breakdown
| Model | Quality | Confidence | Cost / 1k runs | Overpay | Mode |
|---|---|---|---|---|---|
| DeepSeek V4 Flash ★ DeepSeek | 9.54 / 10 CI [9.40, 9.68] | RANKED | $0.40 | best value | batch |
| MiniMax M3 MiniMax | 9.46 / 10 CI [9.30, 9.61] | RANKED | $5.29 | 13x | batch |
| GPT-5.4 Nano OpenAI | 9.46 / 10 CI [9.22, 9.71] | HIGH | $7.61 | 19x | batch |
| DeepSeek V4 Pro DeepSeek | 9.54 / 10 CI [9.40, 9.69] | RANKED | $11.52 | 29x | batch |
| GPT-5.6 Luna OpenAI | 9.51 / 10 CI [9.27, 9.75] | HIGH | $12.15 | 30x | batch |
| Qwen 3.7 Plus Alibaba Cloud (DashScope) | 9.31 / 10 CI [9.17, 9.44] | RANKED | $22.47 | 56x | batch |
| Qwen 3.6 Flash Alibaba Cloud (DashScope) | 8.99 / 10 CI [8.49, 9.48] | MEDIUM | $23.13 | 58x | batch |
| Grok 4.5 xAI | 9.23 / 10 CI [9.04, 9.42] | RANKED | $29.30 | 73x | batch |
| Qwen 3.6 Plus Alibaba Cloud (DashScope) | 9.65 / 10 CI [9.47, 9.84] | RANKED | $37.25 | 93x | batch |
| Gemini 3.5 Flash Gemini | 9.70 / 10 CI [9.64, 9.77] | RANKED | $42.96 | 107x | batch |
| Kimi K2.6 Moonshot AI | 9.71 / 10 CI [9.61, 9.81] | RANKED | $55.73 | 139x | batch |
| Gemini 3.1 Pro Preview best Gemini | 9.84 / 10 CI [9.74, 9.94] | RANKED | $62.16 | 155x | batch |
| GPT-5.6 Sol OpenAI | 9.56 / 10 CI [9.32, 9.79] | HIGH | $62.47 | 156x | batch |
| GPT-5.5 OpenAI | 8.96 / 10 CI [8.55, 9.36] | MEDIUM | $205.70 | 513x | batch |
Overpay shows how much more you pay than the best-value model that clears the quality bar (marked ★) — the best-value good-enough option. "16x" means you overpay 16× — 16× that reference for no quality benefit above the bar. Typical call shape for this task: 4084 input tokens → 5339 output tokens, EMA-tracked from production traffic. Cost is the observed, all-in $ per 1,000 task runs: each model's own measured usage on this task — output verbosity, thinking/reasoning tokens, cache reads and writes, and the spend on its billed failures — priced at current list rates and adjusted by the billing overhead we actually reconcile against provider invoices. Models that answer tersely cost what they actually cost; models that think at length pay for it. Not comparable to providers' advertised $/1M list rates — this is what running the task costs, not a per-token price.
Prompt templates
The system + user template pair used for this task.
STRUCTURED_OUTPUT_EXTRACTION_SYSTEM_PROMPT +
STRUCTURED_OUTPUT_EXTRACTION_USER_PROMPT
(1565 calls in window)
System prompt
You are a JSON extraction assistant. Extract structured data from the provided text that matches the given JSON schema. Output ONLY the structured data matching the schema.
User prompt
Extract the structured data from the following text that matches this JSON schema:
## JSON Schema:
```json
{schema_json}
```
## Text to extract from:
{raw_text}