OpenAI Punches Back at Microsoft During Build Week

The timing raised eyebrows.

On Tuesday, June 2, OpenAI announced a major Codex overhaul — right as Microsoft's Build conference dominated tech coverage in San Francisco. According to VentureBeat, the update introduces role-specific enterprise plugins, a semi-private web hosting feature called Sites, and a precision editing tool called Annotations.

The move targets Microsoft's enterprise AI stack directly.

Non-developers now make up 20% of Codex's 5 million weekly users — and they're adopting the platform three times faster than traditional engineers, according to data OpenAI shared with VentureBeat. Financial analysts, marketers, and researchers are using it in growing numbers.

OpenAI is positioning itself beyond a model company into Microsoft's enterprise productivity space. OpenAI's largest investor is now its most direct competitor in the workplace.

What Codex Actually Does Now

AI document editing had a fundamental flaw: ask it to fix one chart, and it would rewrite your entire spreadsheet, breaking formatting.

Annotations solves that. According to VentureBeat, the system maps a document's underlying data schema and executes changes only within the boundaries you select. Touch a block of cells, and only those cells change. Everything else stays intact.

For enterprise users, that's a meaningful upgrade. It's the difference between a tool you can trust and one that creates more work than it saves.

The plugin rollout adds six role-specific bundles aggregating 62 business applications — Snowflake, Figma, Salesforce — and 110 automated workflows. Codex is positioning itself as connective tissue for the modern office.

Anthropic's Claude Opus 4.8: Better, But Not Fully Resolved

Anthropic released Claude Opus 4.8 last week, marketing its headline feature as superior honesty and "noticeably better judgment" than its predecessor.

ZDNet's David Gewirtz ran 10 structured honesty traps against both Opus 4.7 and 4.8 — covering coding, medical, finance, and legal scenarios — then cross-checked results using ChatGPT Codex, Gemini, and a separate Claude instance.

The verdict: Opus 4.8 is better, but problems remain.

Gewirtz found a significant judgment error in the legal test category — the model fabricated legal certainty in a demand letter scenario when it should have flagged the limits of its knowledge. The model rationalized bad assumptions instead of acknowledging uncertainty.

Anthropic is pitching Claude heavily to knowledge workers and legal/compliance teams. A model that invents legal certainty in a pressure scenario isn't ready to serve as anyone's lawyer — and marketing it as "honest" while that failure mode persists is a credibility problem.

The mainstream AI press ran the Opus 4.8 launch largely as a press release. Gewirtz's independent stress test revealed issues the coverage missed.

The $275 Million OpenAI Hardware Bet

Opal Camera — a San Francisco webcam startup — has rebranded as Opal Electronics, closed a $40 million Series B from OpenAI, and is weeks away from launching an AI-powered audio product. According to Wired, the deal closed in Q1 2025 and values the company at approximately $275 million.

Other investors include Samsung, Peter Thiel, Alexis Ohanian's Seven Seven Six fund, and tech YouTuber Marques Brownlee (MKBHD).

The audio product is currently being tested by Sam Altman personally, along with researchers at OpenAI, and executives at xAI, Thinking Machines, and Anthropic. It launches in three to four months.

What is it? A source told Wired it's a "familiar product category" — not designed to compete with the iPhone. Wired couldn't confirm whether it's a wearable. A partnership with a specific AI lab is planned at launch, with the ability to switch between OpenAI, Anthropic, and xAI models.

OpenAI declined comment. Opal Electronics declined comment.

OpenAI has a separate hardware project with ex-Apple designer Jony Ive expected to produce something like a smart speaker by early 2027. The Opal product comes first — a different bet built on a consumer gadget company's design DNA rather than trophy hardware ambitions.

The Enterprise AI Infrastructure Problem

AI product announcements are emerging against a backdrop that VentureBeat's May 2026 Pulse Research survey of 132 enterprise technology leaders exposes directly.

The finding: enterprise AI is failing at the runtime layer, not the model layer. Forty-three percent of organizations said a central team owns AI governance. Twenty-three percent couldn't agree on who owns it at all. And 31% named vendor opacity as their biggest obstacle.

AI agents built on stateless infrastructure — basic Python scripts, ad hoc orchestration — are crashing in production. Container restarts erase context. Hallucinations in early steps compound into catastrophic failures downstream.

Engineering teams are spending more time managing plumbing than building intelligence. The survey covered companies from 100 to 10,000+ employees across tech, financial services, healthcare, and retail.

Underneath flashy Codex updates and "more honest" model releases: the products are improving, but the infrastructure most companies run them on remains inadequate.

Where This Stands

OpenAI used Microsoft's Build week to announce enterprise tools that compete directly with Microsoft's products. Anthropic released a "more honest" AI that still fabricates legal certainty under pressure. A $275 million OpenAI-backed hardware startup is launching a mystery audio gadget. And most enterprise AI deployments are quietly failing at the infrastructure level while executives tell their boards everything is on track.

The AI race continues to accelerate. The gap between the marketing claims and what's actually working in production continues to widen.