AI · Wednesday, 2 September 2026 · Issue 99 · 8 stories

Thomson Reuters builds its own $40M model, stops renting AI from Anthropic

The cost to own a frontier model just dropped below the cost to rent one. Thomson Reuters spent $40M building its own LLM and is already pulling work off Claude. That math changes the vendor calculus for every company with proprietary data.

Models & Releases

Thomson Reuters built its own model for $40M to stop renting AI from Anthropic

Thomson Reuters unveiled Thomson, its first proprietary LLM, trained on decades of Westlaw, Practical Law, and Reuters content starting from an open-source foundation. The whole build cost $40 million over two years, with the final training run around $450,000, a rounding error next to frontier-lab budgets. It is already taking document review work off Anthropic's Claude inside CoCounsel Legal, a bet that owning the model beats renting it.

thomsonreuters.com ↗6 min read at source

Claude Fable 5.1 doubles its science score, then nails the pelican test

Anthropic's Fable 5.1 claims a new standard for coding and long-running tasks, backed by a jump to 52.6% on Terminal-Bench-Science, up from 24.7% for Fable 5. Simon Willison's first test is more telling: at low and medium effort the model skipped reasoning entirely, while max effort burned 65,927 tokens, $1.83 and 13 minutes producing the best pelican he's seen from any Anthropic model. The cost cliff between those extremes is the real takeaway.

simonwillison.net ↗5 min read at source

TontaubeV1: an open-weight TTS for voice cloning and long-form audio

TontaubeV1 is a 2.9B-parameter, character-level TTS release built for long-form narration and low-latency streaming, with zero-shot voice cloning from as little as 5 seconds of reference audio. English and German are primary, with six more languages supported. A four-codebook design splits semantic from acoustic prediction, and audiobook, conversational, and agentic styles ship as compact BPE-encoded controls. Weights are open, with a matching vLLM adapter on GitHub.

huggingface.co ↗6 min read at source

Research

The cheap fix that beats linear attention: sliding windows win long-context tests

Sliding-window attention with attention sinks, one of the oldest and cheapest fixes for quadratic cost, matches or beats post-trained linear-attention models across several LLMs. On long-context reasoning (Needle-in-a-Haystack, BABILong) it scores 2 to 10 times higher, with zero post-training. The authors' blunt takeaway: labs spending compute on linearization may be solving a problem a rolling window already handles.

arxiv.org ↗2 min read at source

Fused-kernel attention breaks the memory wall: 11.6x faster long-context decoding

Long-context decoding stalls on memory bandwidth, not compute. Faster Flash Decoding fuses the attention selector and compute into one kernel, replacing metadata indices with low-bit scanning plus a top-delta filter that drops blocks adaptively, no retraining required. The payoff: up to 11.6x kernel speedups, 2.37x end-to-end throughput, and 256K context without losing accuracy on RULER or LongBench. Attention sparsity just became practical for production inference.

arxiv.org ↗2 min read at source

Products & Industry

OpenAI quietly tests pay-per-completed-job pricing for big accounts

OpenAI is reportedly letting a handful of major customers skip token-based billing and pay only when its AI actually finishes the job, like a support ticket handled end-to-end. Terms, prices, and clients are undisclosed, and OpenAI hasn't commented. It's a notable posture change: the vendor, not the buyer, now eats failure. OpenAI follows Intercom, Zendesk, and Salesforce down the outcome-based path, so expect the per-success-rate math to get very interesting.

thenextweb.com ↗3 min read at source

Tools & Open Source

Stream a 104GB Qwen off your SSD: ~12 tok/s on a 48GB Mac

slotstream is a single Swift binary that streams a 125B-parameter Qwen MoE (104GB at 4-bit) off your SSD, so far-less-RAM Macs can still run it. On a 48GB Mac it hit ~12 tok/s warm decode at 32GB peak, with a ~2s start: only the 3.8GB trunk loads. It speaks the Ollama and OpenAI APIs, and doctor simulates your machine's plan before you download. Reality check: only the 48GB row is measured; the rest are estimates.

github.com ↗9 min read at source

Agent Memory Should Be a File Format, Not a RAG Pipeline

Cal Paterson argues agent memory should be data, not a pipeline: portable Markdown pages with optional YAML frontmatter and a SQLite vector index, packed as a .memoryfield file. He skewers lab-tied harnesses, pgvector-plus-Neo4j systems, and "high modernist" fact graphs for treating memory as a process. Plain files stay legible to humans, agents, and future models alike. A proposal, not a shipped tool, but the critique lands.

calpaterson.com ↗11 min read at source

You’re caught up.

That was all 8 stories in the 2 September issue.