OCR comparison: Unlimited-OCR 2, Docling and MinerU on academic PDFs
Docling took 3.5 minutes for all 10 papers, Unlimited-OCR 2 on SGLang 31 minutes and MinerU 62 minutes, with no accuracy advantage for either slower tool on born-digital papers.
Computer use: Claude Haiku 5.5, GPT-6.1 Terra and GPT-6 Luna
On the web-page check Haiku scored 0.95 in 23.4 s, Luna 0.9 in 43.6 s and Terra 0.9 in 46.7 s; on the desktop-app check all three scored 1.0 in 35.1 s, 47.9 s and 40.0 s.
Eight small local models on transcript-cataloguing tasks
Gemma 4 31B and 12B passed 5 of 6 checks and LFM 2.5 8B passed 0 of 6; median generation speed ranged from 10.3 to 153.7 tokens per second.
A 17-hour local run: Gemma 4 31B over 508 documents
508 documents took about 17 h 19 min on a DGX Spark at a median of about 32 tokens per second, using 8,571,832 input tokens and 1,589,326 generated tokens.
Draft-token setting and speed on Gemma 4 31B
With the draft-token setting at 0, 3 and 5, median speed was 10.6, 32.6 and 41.3 tokens per second; each setting passed 4 of 6 checks.
Builder pilot: Opus 5.5 rebuilds five app tasks first built by Sonnet 5.5
At list prices Opus cost $9.02 against $3.78 for Sonnet (2.4 times); blind review scores were close (3.30 and 3.40 against 3.27 and 3.23). One build per cell.
Reasoning tokens used up an output cap on GPT-5.5 Pro
With max_output_tokens at 6,000, all 6,000 tokens went to reasoning and no text came back; an uncapped run took 542 s and 22,828 output tokens.
Local image generation on an M4 Max: FLUX.2 Klein 4B and Qwen-Image
At 1024 x 1024, FLUX Klein took about 18 s; Qwen-Image at 20 steps took about 8 minutes.
Models tested
Green is open weights, amber is closed. Each model lists its entries.