DeAI Daily Brief — 8 October 2026

Today in DeAI: Nous Research confirms a $90 million Series B at a $1.5 billion valuation, Anthropic cuts Claude Haiku 5.5 prices by up to 90 percent, Liquid AI ships open weights for two edge decision models, and OpenAI and Google both expand content-provenance tooling.

Nous Research raises $90M at a $1.5B valuation for a Hermes enterprise push

Nous Research, the lab behind the open-source Hermes Agent, confirmed a $90 million Series B led by Robot Ventures, with The Wall Street Journal reporting the valuation at $1.5 billion (TechCrunch). The money funds Hermes for Businesses, a paid enterprise version of the agent; no price, feature list, or launch date is published. Its 24-million-clone and ~2.5%-of-global-tokens traction figures are company self-reports. Why it matters: an open-weights-native lab is now selling ownership-of-the-stack to enterprises, a direct pressure point on hosted-API vendors. Our full coverage

Claude Haiku 5.5 arrives with a 90% price cut under 100K tokens

Anthropic priced Claude Haiku 5.5 at $0.10/$0.50 per million input/output tokens for prompts up to 100,000 — 90% below Haiku 4.5's $1.00/$5.00 — rising to $0.50/$2.50 beyond that boundary. It exactly matches OpenAI's GPT-6 Luna under 100K tokens, per Simon Willison's independent read, which also flags a hidden cost: the new tokenizer packs the same prompts at roughly 1.25x the token count. Anthropic also halved Sonnet 5.5 cache-read pricing to $0.10 per million. Why it matters: the cheap-fast tier keeps repricing downward, and every third-party host of open small models now has to answer for its own margins against these list prices. Our migrate-from-Anthropic guide

Liquid AI opens up d1-3B and d1-omni-600M edge decision models

Liquid AI released open weights for two d1 decision models that answer in a single forward pass with zero generated tokens: d1-3B (text and image) and an experimental 600M-parameter d1-omni handling audio too. The company reports d1-3B scoring 48.57 on its own Decision Index v0.2.1 benchmark and answering a question in 16 ms on a Jetson AGX Thor and 8 ms on an RTX 4090 — company-measured figures pending independent runs, though the weights themselves are verifiable on Hugging Face. Why it matters: classification, routing, and moderation workloads just moved from an API line item to downloadable weights that run on edge hardware. Our earlier decision-model coverage

OpenAI and Google both push content provenance forward

OpenAI laid out its EU AI Act text-provenance approach: opt-in textGrain watermarking for API customers, invisible watermarks on eligible EU ChatGPT and Codex output, and detector access gated to approved researchers. Its own evaluations show the limits — about 80% detection on 200-token passages at a 1% false-positive target, dropping to 17% when a quarter of words are swapped. The same week, Google opened its SynthID detector to the public and claims more than 180 billion watermarked images and videos — a self-report. Why it matters: provenance is becoming a default build consideration in the EU, and the detection-accuracy caveats are the part most coverage skips. Our EU enforcement coverage

AkashML named launch GPU provider for Architect's prompt auctions

AkashML says it is the launch GPU provider for Architect's Liquid Inference, a routing system where every inference prompt is auctioned and routed to the lowest qualifying offer, with OpenAI- and Anthropic-compatible SDKs. The evidence so far is provider posts echoed by calendar aggregators; usage volume and token-economics impact are unverified claims. Why it matters: per-prompt auctions are a second routing-design signal for decentralized inference after fixed marketplace pricing — worth watching for actual throughput data. Our decentralized networks overview

Watching tomorrow

Whether third-party hosts and aggregators reprice small open models to match Haiku 5.5's $0.10/$0.50 tier, and whether Nous Research publishes any concrete Hermes for Businesses product surface.