Today we're publishing LongExtractBench, a benchmark commissioned by @reductoai and independently validated by micro1. We evaluated seven production document extraction systems across the same 225 complex enterprise documents. The benchmark was intentionally difficult: documents averaged 358 pages and contained roughly 88,700 ground-truth fields each. Every system was evaluated using the configuration documented in the benchmark methodology. Key findings: • Reducto Deep Extract was the only system to successfully complete all 225 documents. • Direct frontier LLM baselines achieved substantially lower completion rates on long, complex documents. • In this benchmark, dedicated extraction platforms achieved higher completion rates than the direct frontier LLM baselines. • Recall was the clearest differentiator. Precision remained high across systems, but recall ranged from 33.8% to 99.6%, highlighting which systems consistently captured the information contained in long, complex documents. The full report includes the benchmark methodology, limitations, and reproducibility resources. Check out the report and results in the comments below.

micro1 x @cartesia: The Human Speech Data Project Every language, accent, and speaking style has its own rhythm, texture, and feeling. Building great voice AI means learning from the people who know those details best. That is why micro1 and Cartesia are launching The Human Speech Data Project, an initiative focused on improving the next generation of voice AI through high-quality speech data and human evaluation. The project brings together multilingual speakers, linguists, transcriptionists, native speakers, and voice actors to help: - Capture and evaluate natural conversational speech - Transcribe and annotate audio with a sharp ear for accuracy - Review AI-generated speech for naturalness, correctness, and edge cases - Bring expert judgment into the model improvement process Visit the link in the comments if you’d like to learn more about how you can contribute.

Two AI models can land the same pathology benchmark score for completely different reasons. Today we're introducing micro1's pathology-reasoning benchmark, the latest addition to our Realm Medical-Reasoning benchmark series. Built with practicing pathologists, it spans the range of cases a working service actually sees, from hematopathology and bone marrow workups to breast, thyroid, gastrointestinal, genitourinary, dermatopathology, and biomarker studies. Rather than testing medical knowledge in isolation, it measures something narrower and harder: whether a model can extract report facts exactly, preserve diagnostic limits, and avoid escalating to conclusions the specimen doesn't support. We scored three frontier models across the dataset: Claude Opus 4.8 - 82.6% GPT-5.5 - 76.3% Gemini 3.5 Flash - 75.7% However, the ranking was the least interesting part. The models were similarly strong at extracting facts and running calculations. Where they separated was judgment: knowing where the report ends. The most common way points were lost wasn't getting facts wrong. It was saying more than the report supports, resolving uncertainty the report left open, or naming a stage, biomarker, or treatment the specimen can't establish. We found that two models can land nearly the same score for completely opposite reasons. The aggregate numbers look close, but the trajectories tell a different story, and those differences get more interesting as the cases get harder. Our takeaway: what matters most isn't how often a model is right, but how it reasons when a report leaves room for doubt. Full report linked in the comments.

Earn up to $25k per successful company data partnership referral to micro1 Link in comments to get started

Introducing the Realm Financial Reasoning benchmark, our new evaluation of frontier AI on reasoning in finance and spreadsheet-grounded analysis. Tasks are built around the actual work product that practitioners deliver, from IFRS reconciliation workbooks and hedge-fund backtests to VC term sheet analyses and treasury cash-flow forecasts. Each task drops the model into a sandbox with the same source materials a human analyst would open: named-range Excel workbooks, broker PDFs, earnings call transcripts, monetary-policy decisions. Here's what the results showed (Pass@3): -GPT-5.5: 0.456 -Claude Opus 4.7: 0.398 -Gemini 3.1 Pro: 0.349 The three models score similarly, and none clears 50% on tasks that demand a judgment call. The back and middle office are defensible today, but on capital allocation questions current frontier models should be treated as research accelerators, not final decision-making support systems. Full report linked in the comments.