Coding agents need patching, frontier labs are inviting auditors inside, and model prices keep sliding while hardware gets denser. Here are seven stories worth your time, each summarized in our own words with a link to the original reporting.
StepFun opens Step 5 Preview API with 600B parameters at $1 in, $2.70 out
StepFun opened API access to Step 5 Preview, a 600-billion-parameter sparse mixture-of-experts that activates about 27 billion parameters per token and carries a 1-million-token context window for text and image input. Independent scoring puts it at 44 on the Artificial Analysis Intelligence Index, roughly matching much pricier flagships while undercutting them severalfold, with a 95 percent cache discount softening repeated-prompt costs further. Builders should note the weights story: only a placeholder ships today, with full open weights announced for October 15, so treat this as an API evaluation window first.
Read the full story at Artificial Analysis (September 18, 2026).
Plugin4Shell brings zero-click RCE to four AI coding agent plugin systems
Researchers at AIR disclosed a supply-chain flaw that lets a plugin repository owner replace pinned plugin code without any user click across Claude Code, OpenAI Codex, GitHub Copilot, and Gemini CLI. The root cause is a checkout that resolves a branch name matching the pinned commit hash instead of verifying the landed commit, so background auto-updates silently pull attacker-controlled code while the marketplace still displays the expected hash. Anthropic and OpenAI shipped fixes in Claude Code 2.1.179 and Codex 0.146.0, while Copilot had no patch at disclosure and Gemini CLI was deprecated without one, so audit installed plugins and pin only from repositories you control.
Read the full story at AI Weekly (September 19, 2026).
Anthropic names Accenture Faculty as first embedded frontier evaluator
Anthropic partnered with Accenture's Faculty unit to evaluate and red-team models from inside the company, with access comparable to employees across training and deployment decisions. Both companies expect to invest at least $1 billion each over five years in this capacity, covering alignment assessments and safeguard testing, while Anthropic says the arrangement is non-exclusive and similar pilots with nonprofit evaluators are in discussion. For builders the signal is governance as a feature: expect audit access, incident reporting, and published safety accounts to become part of enterprise procurement checklists.
Read the full story at Anthropic (September 18, 2026).
Anthropic hits $100B revenue pace and pushes IPO to November
Anthropic is now pacing above $100 billion in annualized revenue, up roughly half from the figure disclosed two months earlier and more than ten times end-of-2025 levels, driven by Claude Code and Cowork enterprise adoption. The company pushed its planned listing from October to November so third-quarter financials can be included, targeting what could become the largest public offering on record. The builder read is demand concentration: enterprise coding and cowork agents are currently funding the frontier, which shapes where APIs, tooling, and integrations get prioritized.
Read the full story at Yahoo Finance (September 18, 2026).
xAI ships Grok Voice Transcribe 2.0 with halved word error rates
xAI released Grok Voice Transcribe 2.0, claiming roughly double the accuracy of its predecessor at unchanged batch and streaming prices, with short-phrase error rates falling sharply across 19 languages and telephony English improving on first-final transcripts. The model adds speaker diarization, timestamps, and key-term biasing through the xAI speech API and currently tops a public streaming-accuracy leaderboard. For automation builders it is a practical drop-in: voice intake, meeting notes, and phone-call pipelines get cheaper errors without changing the billing line.
Read the full story at xAI (September 18, 2026).
Apple relaunches Siri on Foundation Models split between device and private cloud
Apple introduced a rebuilt Siri alongside its 2027 software updates, adding personal context from messages and photos, on-screen awareness, broader cross-app actions, and web-grounded answers. The work runs on next-generation Apple Foundation Models built with Google's Gemini collaboration, split between on-device processing with local Spotlight and App Toolbox indexes and server-side Private Cloud Compute, with camera-centered Visual Intelligence extended across iPhone, iPad, Mac, and Vision Pro. Builders should study the routing pattern: keep personal lookups local, escalate selectively to confidential cloud, and gate daily server-side usage behind limits.
Read the full story at eeNews Europe (September 18, 2026).
CXMT puts fifth-generation DRAM with 11.95nm half-pitch into mass production
ChangXin Memory Technologies announced mass production of its G5 DRAM platform at the World Manufacturing Convention in Hefei, launching two 24-gigabit LPDDR5X products for mid- to high-end smartphones and portable devices. The platform uses quadruple patterning to reach an 11.95-nanometer array half-pitch with a 45:1 capacitor aspect ratio and a high-k metal gate flow, lifting dies per wafer by at least half over the prior generation. With memory demand rising on AI infrastructure and DRAM revenue jumping sharply last quarter, local-AI builders get welcome news: more supply options beyond the three incumbent vendors.
Read the full story at Global Times (September 20, 2026).