Anthropic ships Claude Sonnet 5.5 — 30% faster, up to 30% cheaper per task
Anthropic released Claude Sonnet 5.5, the second model in its Claude 5.5 family after last week’s Opus 5.5, positioning it as the everyday workhorse for coding, documents, slides and spreadsheets. List prices stay at $2/$10 per million input/output tokens, but Anthropic says the model burns fewer tokens and generates outputs more than 30% faster than Sonnet 5 — cutting per-task cost by up to 30%. On Terminal-Bench 4.0 it jumps to 70.6% from Sonnet 5’s 10.3%, and it is the first Sonnet tier under the same cyber safeguards as Opus and Fable. Haiku 5.5 is promised in the coming weeks. Practical implication: mid-tier model competition is now about latency and token efficiency at fixed list prices, not headline $/MTok cuts — and cyber capability is being treated as a first-class release gate even for “work partner” models.
What else matters
NVIDIA opens full-stack agent safety platform with OpenShell and Sentry
NVIDIA launched its Open Agent Safety Platform: open-source OpenShell for runtime policy enforcement on Vera CPUs (extensible to Arm/Intel), plus Sentry — an out-of-band BlueField-4 DPU watchdog that can quarantine agents crossing software boundaries in milliseconds. Partners named include Anthropic, Microsoft, Hugging Face, CrowdStrike, Red Hat, Cisco and others. Signal: after a summer of agent containment failures, the GPU vendor is productising hardware-rooted agent governance, not just model-level guardrails.
Manus 2.0 Cascade cuts agent token burn 23% as $4B raise looms
Singapore agent startup Manus shipped Manus 2.0 with a new Cascade harness that, in its tested config, used 23.2% fewer tokens, finished tasks 28.2% faster and cut run cost 32%. The release adds Cloud Computer, event-triggered Automations, and Manus Studio — weeks after resuming fully independent operations following a blocked Meta acquisition. Reports put a $500M raise at ~$4B valuation in flight. Efficiency at the harness layer is becoming as contested as model weights.
AutoTrust open-weights JEV-27B for self-hosted agent decisions
AutoTrust AI released JEV-27B under Apache-2.0: a 108.9M-parameter decision block (~0.4% of weights) on a frozen Qwen3.8-27B backbone that answers yes/no, multiple-choice and 0–5 ratings with calibrated probabilities in one forward pass, while System 2 generation stays byte-identical to the base when the block is off. Trained in ~9.2 B200-hours; ships with vLLM support on Hugging Face. Useful for shops that want structured agent decisions on-prem without shipping every choice to a third-party API.
Hugging Face → · PR Newswire →
Labs float self-regulatory “Standards Authority for Frontier AI”
Google, OpenAI and Anthropic are reportedly planning a FINRA-style Standards Authority for Frontier AI for late 2026 or early 2027, after a White House executive-order path stalled. Early scope would cover pre-release testing, incident reporting and auditor certification — initially among the three labs, not Microsoft, Meta or xAI. Treat as agenda-setting until a charter and membership rules land.
