Show HN: Don't Get Nerf'd by Document Processing

Pacific launched nerfguard-extract-1, a document extraction model that scored 96.13% Value F1 on Pacific's ExtractBench finance evaluation at $0.0024961 per page across 145 documents, which Pacific says is up to 99% cheaper and up to 3x faster than frontier alternatives. In the same benchmark, LlamaExtract 2.5 Agentic Plus scored 97.37% Value F1 at $0.11875 per page and Sol 6.1 evidence scored 97.82% at $0.23338 per page, while Google Gemini 3.8 Flash scored 90.88% at $0.011124 per page. Pacific said developers can call nerfguard-extract-1 through its native API or an OpenAI- or Anthropic-style structured outputs flow, with citations and confidence values returned per extraction.

Frontier doc extraction99% lower costup to 3x faster Complete with citations and confidence values per extraction Try it today https://console.pacific.app/playground Frontier quality. Dramatically lower cost. 96.13% Value F1 at $0.0024961 per page, measured on ExtractBench finance document dataset. View benchmark data https://extract.nerfguard.com/extractbench-finance.csv | ExtractBench Finance benchmark results. Public leaderboard rows use Finance accuracy and overall benchmark average costs. | | | | | | | | | |---|---|---|---|---|---|---|---|---| | System | Value F1 % | Page F1 % | Box F1 % | USD per page | Seconds per document | Documents | Source | Cost basis | |---|---|---|---|---|---|---|---|---| | nerfguard-extract-1 | 96.13 | 89.98 | 73.74 | 0.0024961 | 54.2 | 145 | Pacific evaluation | Finance evaluation | | Google Gemini 3.8 Flash | 90.88 | Not reported | Not reported | 0.011124 | 62.4 | 145 | Pacific single-call finance evaluation | Benchmark convention, same observed usage at non-promotional Standard rates | | LlamaExtract 2.5 Agentic Plus | 97.37 | 90.39 | 86.79 | 0.11875 | 179.4 | 145 | Pacific evaluation | Finance evaluation | | LlamaExtract 2.5 Cost-Effective | 95.50 | 66.75 | 54.48 | 0.01000 | 86.6 | 145 | Pacific evaluation | Finance evaluation | | GPT-6 Astra one-shot reasoning low | 93.80 | 0.00 | 0.00 | 0.06750 | 64.0 | 145 | Pacific evaluation | Finance evaluation | | Sol 6.1 evidence | 97.82 | 90.02 | 91.21 | 0.23338 | 105.1 | 145 | Pacific evaluation | Finance evaluation | | Sonnet 5.5 evidence | 91.22 | 57.06 | 26.42 | 0.02117 | 72.7 | 145 | Pacific evaluation | Finance evaluation | | Opus 5.5 evidence | 95.75 | 46.03 | 19.27 | 0.04089 | 85.1 | 140 | Pacific evaluation | Finance evaluation | | Reducto Deep Extract | 85.1 | Not reported | Not reported | 0.06350 | Not reported | 145 | ExtractBench public leaderboard | Overall benchmark average | | LlamaExtract Agentic | 91.3 | Not reported | Not reported | 0.03120 | Not reported | 145 | ExtractBench public leaderboard | Overall benchmark average | | Codex GPT-5.5 | 96.2 | Not reported | Not reported | 0.27830 | Not reported | 145 | ExtractBench public leaderboard | Overall benchmark average | | Extend Max Context | 79.2 | Not reported | Not reported | 0.10000 | Not reported | 145 | ExtractBench public leaderboard | Overall benchmark average | | Extend Review Agent | 79.2 | Not reported | Not reported | 0.12110 | Not reported | 145 | ExtractBench public leaderboard | Overall benchmark average | | Claude Code Opus 4.8 | 84.7 | Not reported | Not reported | 0.16170 | Not reported | 145 | ExtractBench public leaderboard | Overall benchmark average | | Datalab Accurate + Balanced | 62.1 | Not reported | Not reported | 0.03500 | Not reported | 145 | ExtractBench public leaderboard | Overall benchmark average | | Google Gemini 3.5 Flash | 75.4 | Not reported | Not reported | 0.01000 | Not reported | 145 | ExtractBench public leaderboard | Overall benchmark average | What $1 can extract. Compare how many one-page documents each model can process for the same $1. And their relative speeds. Easily swap it into any existing model call. Use Pacific’s native API, or keep your familiar OpenAI or Anthropic structured outputs flow. Just connect to Pacific and choose nerfguard-extract-1 . Read the API docs https://docs.pacific.app curl https://api.pacific.app/public/v1/extract \ -H "Authorization: Bearer $PACIFIC API KEY" \ -F "file=@invoice.pdf" \ -F 'schema={ "type": "object", "properties": { "invoice number": { "type": "string" }, "total": { "type": "number" } }, "required": "invoice number", "total" , "additionalProperties": false }' Questions, answered. What is nerfguard-extract-1? Document extraction that delivers frontier quality for up to 99% lower cost and at 3x the speed. This makes it practical to process entire document archives, automate invoice and contract workflows at scale, and build extraction into products where cost and speed previously got in the way. Give it a document and the fields you need, and get structured data with citations. Try it in the console playground https://console.pacific.app/playground and integrate it through Pacific’s API or as a model through your existing OpenAI or Anthropic client. Need even higher quality extractions, integrations with live syncing data sources, search and fine grain access controls? Check out Pacific Developer APIs https://pacific.app/ . What kinds of documents can I extract from? PDFs, spreadsheets, Office documents, CSVs, text, and Markdown. Use it for workflows such as pulling line items from invoices, terms from contracts, or figures from financial reports. The native Pacific endpoint accepts file uploads up to 200 MB; Chat Completions accepts base64 files up to 20 MB. See supported inputs https://docs.pacific.app/extract-openai multipart . How do I tell it what to extract? Define the fields in a JSON Schema—for example, an invoice number, a total, and a due date. Send that schema with the document. You can reuse the same schema across documents, or use a different one for each workflow. Can I use my existing OpenAI or Anthropic client? Yes. The compatible APIs make it easy to switch if you already use structured extraction through frontier LLM providers. Keep your existing client and request pattern, authenticate with a Pacific API key, and select nerfguard-extract-1 . Read the API docs https://docs.pacific.app for setup details. Can I check where an extracted value came from? Yes. Extracted fields include instances with their value and source citation, so you can trace an answer back to the document. Citations can include the page, reference text, and location on the page. You can also optionally utilize hosted citations to get permissioned links to show the highlighted line or table cell where extracted values came from. What happens to the documents I send? nerfguard-extract offers zero data retention. By default, citations are made available for you to easily retrieve later and saved for that purpose. To turn off, set hosted citations=False . Is Pacific SOC 2 Type II audited? Yes. Pacific is SOC 2 Type II compliant and goes through regular audits. How do I try it? Open the console playground https://console.pacific.app/playground , select nerfguard-extract-1 , and try it with your own document and schema. When you are ready to integrate it, create a Pacific API key and use one of the examples above. Time to scale More documents. Less waiting. A fraction of the cost. Try nerfguard-extract-1 with your own document and schema. Your first $1 in credits is on us. It goes further than you think.