OpenAI cut its frontier model's price by 20%.
Claude Code now chains tool calls in JavaScript.
▲ AI DEV SIGNALS
by Marktechpost · November 18, 2026
Hi there,
Document extraction has a trust problem: a value can look right and still come from the wrong column. Today's featured story looks at an API that attaches a citation to every field. Also: OpenAI cut Astra's price by 20%, Claude Code can now chain tool calls in JavaScript, and a LoRA adapter lifted a 27B model from 39% to 65% on SWE-bench Pro.
01 · Featured: every extracted field now comes with a receipt
Sponsored by Nutrient
An LLM that reads a document can hand you a wrong value as fact. Nutrient's Data Extraction API ties each value back to where it sits on the page. It parses PDFs, scans, images and Office files into JSON that matches your schema.
(Sponsored) Get a free API key →
How it works: Each field returns a page, a bounding box, a match label and, when available, a confidence score. Labels run from id_match to fuzzy_match to not_found. Studio, a browser workspace, lets you test schemas on your own files before writing code. Saved setups become processors you call by ID.
The numbers: Agentic mode scores 86.2 on Nutrient's WholeDoc Bench at 1.8¢ per page, top of 19 settings. The best frontier-model setting, Claude Fable 5.1 (high effort), scores 67.1 at 17.8¢. Understand mode scores 80.8 at 0.9¢. The free plan includes $5 of monthly usage, about 330 understand-mode extractions.
The fine print: Nutrient built and ran WholeDoc Bench, and its main test set is private. The code is Apache-2.0, with a 10-document public sample. When no score exists, confidence is omitted, not low. Charts remain hard: 64.5 is the best chart score on the board.
Our take: Gate on the match label first, then on confidence. Send fuzzy_match and not_found to a human. That's how extraction becomes auditable instead of hopeful. Read the full breakdown.
02 · 4 more drops: a cheaper Astra, code-mode Claude, agent teams and a 27B coding boost
Astra drops to $8/$40 per 1M tokens and ranks #2 on the AA Intelligence Index at 62, behind Claude Fable 5.2 at 65. Sol lands within 0.5 points of Astra on FrontierCode 2.0 at about one-fifth the cost. The catch: coding agent scores didn't improve, and Astra posts a 59% AA-Omniscience hallucination rate.
In v2.1.323, Claude can write 1 program that chains Bash, Read and Edit calls. Background code runs for up to 2 hours while you keep chatting. The catch: it's experimental, terminal-only, and unavailable on Bedrock, Google Cloud and Foundry.
A lead agent runs a default team of Claude Opus 5.5, GPT 6.1 Sol and Gemini 4 Carbon. The agents share findings directly and edit shared files. The catch: it's Max-only, web-only, and ships with no published benchmarks.
Hugging Face trained the Apache-2.0, rank-32 adapter with RL inside the pi coding agent. Median turns on SWE-bench Pro fell from 59 to 23. The catch: results are self-reported on small sets, and it only targets pi.
03 · Signals: search swarms, faster browser agents, and context that costs more than it saves
SID AI ships SID-2 for agentic search. A 12-agent SID-2 Swarm scores 0.720 NDCG on SID's 506-question benchmark, above GPT-6 Astra at 0.704. Pricing starts at $0.25 per 1M input tokens.
Hark's Handoff² runs 2x faster than v1. Latency-penalized RL taught it to batch actions. It completes 82% of 478 internal web tasks, up from 73%.
Reson8 open-sources Rumble-2-Flash. It's a 0.5B Apache-2.0 speech recognition model for 40+ languages. It reports 6.6% average WER on the Open ASR English leaderboard.
Hugging Face now routes Inference Providers through OpenRouter. Each call picks a provider by price, speed and reliability, with automatic retries. Custom provider keys no longer work for language models.
More context can make coding agents pricier. Across 20,000+ runs, retrieved code cut search commands by 18–34% but raised cost 33–42% with a tenth-price cache. Resolution rates didn't improve.
Frozen LLM agents can teach themselves. Given a compute budget and tools to write text notes, capable models improved with no outside training loop. The notes transferred to new game mechanics and 3D navigation.
Tencent launches Hy4 Flash. It's a natively multimodal model with a 1M-token context. Tencent says it edges Hy4 preview on WorkIndex at 1/13 of the cost per task. Open weights are planned but not out yet.
Bland launches Tofu-1 for phone calls. It's a full-duplex speech-to-speech model built for noisy, high-stakes calls. It's in enterprise research preview, with no benchmarks, size or pricing yet.
1 question
If every extracted field came with a match label, how many would you still send to a human?
Have a great Wednesday.
until next week,
the marktechpost team
REACH 70K+ AI BUILDERS
Was this email forwarded to you? Sign up here

