DEV Community

C. Wheatley
C. Wheatley

Posted on Originally published at isymbolic-blog.vercel.app

AI Today: OpenAI Rents Out Its Harness

#ai #news #dailydigest

OpenAI spent yesterday selling infrastructure rather than intelligence: the harness that runs Codex is now an API anyone can call, and the voice model behind ChatGPT Voice is rented by the minute. Anthropic spent the same day publishing what people have been trying to do with its models, which is the less flattering half of the same business.

Agents & Models

OpenAI opened the Codex harness as the Agents API
Public beta since September 10, for all developers. MarkTechPost's writeup lays out the four objects: an Agent (model, instructions, tools, MCP servers), an optional Environment sandbox, a durable Session, and the events it streams. It compacts earlier context automatically as a session nears its limit. There's no platform fee — you pay for tokens, tools and container time — and nine sandbox partners have first-class integrations, including Cloudflare, Modal, Oracle, E2B and Vercel. Early numbers from named customers: SafetyKit 60% lower cost per case, Hypha 86% fewer failed responses, Ciridae a 4x latency cut on subagent flows. The strategic read is that OpenAI would rather sell the scaffolding than watch everyone rebuild it.

GPT-Live-1 is real, and it's $0.05 a minute
I owe this one a correction of sorts. On September 3 I threw out an aggregator's claim that OpenAI had launched a voice model called "GPT-Live," because no primary source existed. It exists now: a launch page, API docs, and a model card dated September 10. TestingCatalog reports full-duplex audio with 12 voices, telephony support, and roughly 80% fewer interruptions than turn-based systems. Paired with GPT-6 Astra at medium reasoning effort it gained 30 percentage points over GPT-Realtime-2.1 on Full Duplex Bench. Watch the pricing: $0.05/min is the voice layer only, backend model and tool calls bill separately, and the clock doesn't stop for silence — a caller on hold costs the same as one talking.

Policy & Regulation

Anthropic published eight months of disrupted misuse
The window is December 2025 through August 2026, across cyber operations, influence, surveillance, scams, biological misuse, conventional weapons and distillation. The cyber cases are the ones with teeth: a Russian espionage group tracked as GTG-20006 (Microsoft calls it Midnight Blizzard) engaged more than 20 organisations across Ukrainian ministries, defence bodies and drone supply-chain firms, per TechNode; a Chinese operator targeted 50 organisations and produced "more than a dozen possible zero day findings in a single month"; one influence network ran 70-plus fabricated news sites and 8,913 articles in 20 languages. On biology, FoneArena's summary is that safety systems blocked direct bioweapon prompts but researchers routed dual-use pathogen queries through proxy networks. Several outlets ran headlines saying Claude can no longer be assumed below the bioweapons threshold; I couldn't corroborate that wording in any source I could actually read, and it sits awkwardly next to "classifiers largely contained it," so I'm leaving it out.

California signed two bills creating an AI audit profession
Newsom signed SB 813 (McNerney), a framework letting independent organisations verify AI systems against state law, and AB 1405 (Bauer-Kahan), a state registry of AI auditors with independence and transparency standards. Together they answer a question the EU AI Act mostly left open: who does the auditing, and who certifies them. The announcement doesn't give implementation dates.

Infrastructure

Microsoft plans to more than triple data-centre capacity by 2032
Bloomberg reports a target of about 38 gigawatts, up from roughly 12 today, with the AI-chip share rising from about 2 GW to roughly a third of the total. Capex was $175 billion planned for calendar 2026 and $50 billion forecast for fiscal Q1 2027. One accounting detail worth noticing: Microsoft intends to spread long-term leases over 25 years instead of 15, which lowers reported annual capex without lowering the spend. Microsoft didn't comment to Reuters, so this is reporting, not an announcement.

Science

DeepMind published predicted effects for all 9 billion human DNA variants
AlphaGenome Atlas, out September 8, covers every possible single-letter change in the genome — 3 billion base pairs times three substitutions. It's about a petabyte, over 30 times the AlphaFold Database, with thousands of predictions per variant across hundreds of human and mouse cell types. The AVI score merges AlphaGenome and AlphaMissense so the 98% of the genome that doesn't code for protein gets scored too. Applied to UK Biobank data from 54,000+ participants it surfaced 22% more non-coding associations. Free for academic research, commercial access via Google Cloud.

Skipped as already covered: the CISA distillation advisory, DeepSeek V4.1-Flash, Positron's $875M, and Meta's Muse, all from yesterday. One trap worth flagging: a digest listed Anthropic's distillation-detection post — 16 million exchanges, 24,000 fraudulent accounts — as September 10 news. The post is dated February 23, 2026. Those numbers are seven months old and I've left them out.

Sources

Top comments (0)