This brief covers AI news from 2026-10-15 UTC.
Open weights, a sub-second voice model, and the first enforcement letters under Europe’s AI Act land on the same day.
StepFun publishes Step 5 Preview open weights
What happened: StepFun released Step 5 Preview weights on Hugging Face and ModelScope, hitting the October 15 date it set at launch. The 600B sparse MoE with 27B active parameters ships in FP8, BF16 and NVFP4 builds with up to 1M context and text, image and video input under the StepFun Model License 1.0.
Why it matters: A 600B-class multimodal model with open weights and quantized builds gives teams a self-hostable option for long-context work that previously required a closed API.
xAI ships Grok Voice 3.0 at the same price as 2.0
What happened: xAI released Grok Voice 3.0, a speech-to-speech model it says ranks first on Artificial Analysis’s Speech-to-Speech Index at 86.9 with under one second latency. It is available in the Voice Agent API at $0.08 per minute, and agents on grok-voice-latest were switched automatically.
Why it matters: Voice agents get a latency and quality upgrade without a price change, but anything pinned to grok-voice-latest already moved to the new model.
Anthropic reports $16.8B Q3 revenue ahead of IPO talk
What happened: Anthropic’s Q3 2026 revenue reportedly reached $16.8 billion, up about 46% quarter over quarter, with an adjusted profit, according to BBC News reporting on figures the company flagged to investors ahead of a stock market debut.
Why it matters: A frontier lab showing profit changes the pricing conversation: buyers can expect less discounting pressure and more enterprise packaging.
EU sends first AI Act information requests to OpenAI and Anthropic
What happened: The European Union issued its first AI Act information requests to OpenAI and Anthropic, the start of active enforcement of the law, according to AI Briefing’s daily roundup.
Why it matters: If you ship a model or agent into the EU, expect documentation and disclosure requests to reach your compliance team rather than staying theoretical.
Siemens EDA cuts agent turns 66% with NVIDIA NeMo Gym
What happened: Siemens EDA used NVIDIA’s NeMo Gym agentic evaluation environment to test its Fuse EDA agent against verified Liberty files, then changed the agent’s skill and MCP tool descriptions. NVIDIA reports the benchmark composite score rose from about 55% to roughly 95% while turns and output tokens fell.
Why it matters: Verifying final artifacts instead of intermediate steps is a concrete recipe for making tool-calling agents reliable enough to run in production.
Cohere: a strong lead agent plus cheap workers cuts cost 70%
What happened: Cohere published results on organizing teams of coding agents. A Sol lead coordinating 16 cheaper Luna planning workers reached 61.8% reward on ProgramBench versus 63.2% for an all-Sol team, cutting cost from $510 to $155.
Why it matters: Where you point extra agents matters more than how many you run, so budget for orchestration design before scaling up inference spend.
ChipAgents and NVIDIA launch Build-A-Chip-Bench
What happened: ChipAgents introduced Build-A-Chip-Bench with NVIDIA, a benchmark that runs AI agents through an end-to-end digital IC design flow from specification to physical implementation and verified GDS. Initial evaluations show a gap between functionally valid designs and physically viable ones.
Why it matters: It gives EDA and chip teams a shared scoreboard for agent reliability across timing, power, area and physical verification, not just RTL generation.