Thrive Holdings launches real-work AI evals from 75+ businesses, benchmarking 13 models
samaysham · x · 2026-10-16
Thrive Holdings released eval benchmarks built from real work across its 75+ portfolio businesses, measuring accuracy, execution time, and cost.
- Individual tax prep: models complete a Form 1040 from client documents. Claude Opus 5.5 (high) leads at 76.3% accuracy ($5.09/return, 6.1 min); Claude Fable 5.1 max hits 75.5% but costs $16.2; GPT-6.1 Sol (medium) scores 74.2% at just $1.4 and 2.7 minutes.
- Helpdesk ticket resolution: GPT-6 Astra (xhigh) tops pass rate at 61.5%, followed by Claude Opus 5.5 max at 59.6%.
The suite covers 13 models across multiple reasoning effort levels, framing model selection as an operating decision: identifying what work models can reliably take on and where humans must step in.
Related event: Thrive Holdings Launches Real-Work Benchmark Across 75 Companies(4 posts)→
More from Models
- NEURA launches Neuron AI, a multimodal PLA model giving robots touch, force feedback and memory — CyberRobooo · 2026-10-17
- Sentience claims per-user models: train, continually update and serve one model per person — soleio · 2026-10-17
- Four Gemini 4 Argon agents beat Minecraft in 27 hours, writing 2,783 programs — GoogleDeepMind · 2026-10-17
- Meta open-sources Muse Spark Open to mixed reception: 'trash' license, nerfed benchmark scores — gabriberton · 2026-10-17
- If new diffusion models are open-weight only, what's left for commercial use? — blackpanda8484 · 2026-10-17
- Reflection Beam OSS architecture: sliding window attention, gated attention, QK Norm — A_K_Nain · 2026-10-17