AI Product Updates Daily — June 2, 2026

Microsoft Build opens today with three new in-house models and a Copilot super app roadmap, Anthropic files confidentially for an IPO at a $965B valuation, and JetBrains drops a 12B open-source model built for production inference. Here's everything that changed.

Microsoft Build 2026: MAI-Thinking-1, MAI-Image-2.5, and a Windows developer mode

Microsoft opened its Build 2026 developer conference in San Francisco this morning, with CEO Satya Nadella and Nvidia CEO Jensen Huang sharing the stage at Fort Mason Center. The event is smaller than past years — a deliberate pivot toward rebuilding developer trust — but the product announcements are concrete.1

MAI-Thinking-1 is Microsoft's first in-house reasoning model and the most strategically significant piece of the week. Microsoft AI chief Mustafa Suleiman confirmed that the model was not built through distillation — it wasn't trained on another model's outputs — which gives Microsoft a reasoning system whose lineage it fully owns. That's a meaningful departure from years of wrapping OpenAI weights in Copilot branding. The company also unveiled MAI-Image-2.5 and MAI-Image-2.5-Flash, image generation models Suleiman teased last week.2

Releasing entity: Microsoft
Release date: June 2, 2026 (Build keynote, 9:30 AM PT)
Availability: Enterprise preview; benchmark numbers not yet published
Previous state: Microsoft had no in-house reasoning model; relied entirely on OpenAI weights

On Windows, Microsoft is shipping a developer-optimized Windows 11 experience: a distraction-free environment pre-loaded with apps, tools, and scripts, responding to a long-standing developer complaint about setup overhead. Windows chief Pavan Davuluri had teased the announcement last week. The company is also continuing a broader effort to rewrite parts of Windows 11 for performance, already shipping in pieces since earlier this year.1

The Copilot super app bundles Microsoft's various Copilot assistants into a single interface. A leaked screenshot that circulated Friday was confirmed to be a Build mockup, not a finished product. The app also includes a first look at Microsoft Scout, an agent reportedly based on the internal OpenClaw project. Preview availability: late summer at the earliest.2

The Nvidia RTX Spark partnership also landed officially at Build. Microsoft and HP revealed RTX Spark PC form factors — miniature Windows on Arm machines with Nvidia silicon — that were absent from Computex, held for today's reveal. The hardware push ties directly to Microsoft's stated goal of letting developers run smaller models on-device rather than paying for cloud inference.

Microsoft Build 2026 opens today in San Francisco 1

Anthropic files confidentially for IPO at $965B valuation

anthropic.comhttps://www.anthropic.com/news/confidential-draft-s1-secExternal link

Loading content card…

Anthropic confirmed Monday it submitted a draft S-1 registration statement to the SEC, beginning the formal process for a public listing.3 The company does not have to proceed if market conditions shift, and no share count or price range has been disclosed.

The filing comes days after Anthropic closed its Series H at $65B in new funding, pushing its post-money valuation to $965B — overtaking OpenAI's last reported $730B valuation. The company's annualized revenue as of May ran to approximately $47B, driven primarily by AI coding tools.4

Anthropic is one of three high-profile AI companies racing to public markets. SpaceX is expected to list this month; OpenAI plans to file in coming weeks. If Anthropic completes its IPO near the $965B valuation, it would enter the S&P 500's top tier. Whether it is profitable is not yet public.4

Anthropic: Claude Agent SDK credit overhaul takes effect June 15

Announced on May 14 and effective in 13 days, Anthropic is separating Agent SDK and claude -p (headless) usage from Pro, Max, Team, and Enterprise subscription pools. Starting June 15, those workloads will draw from a separate monthly dollar credit — sized to match the subscription fee — billed at standard API rates with no rollover.5

Plan	Agent SDK credit / mo	Estimated tasks (Sonnet 4.6)
Pro ($20)	$20	~50 medium coding tasks
Max 5× ($100)	$100	~250 tasks
Max 20× ($200)	$200	~500 tasks
Enterprise Standard seats	$0	Not eligible

The change removes a structural subsidy: heavy Agent SDK users were effectively accessing $300–600 worth of API-equivalent compute on a $20 Pro plan. Interactive Claude Code (terminal TUI), Cowork, and claude.ai are explicitly unaffected.5 Teams running shared CI/CD pipelines can't pool credits across users; Anthropic's own documentation says production shared automation should migrate to direct API key billing.

A frequently missed detail: Enterprise Standard-seat members get $0 in new credit. Only Enterprise usage-based ($20) and Premium ($200) seats carry the credit.

OpenAI responded on the same day Anthropic made the announcement (May 14) by offering new business customers two months of free Codex at the $200/month Pro tier.

JetBrains open-sources Mellum2 — a 12B model for production inference

blog.jetbrains.comhttps://blog.jetbrains.com/ai/2026/06/mellum2-goes-open-source-a-fast-model-for-ai-workflows/External link

Loading content card…

JetBrains published Mellum2 today under Apache 2.0, targeting the latency and throughput bottlenecks that frontier models handle poorly in production pipelines.6

The model uses a Mixture-of-Experts architecture: 12B total parameters, 2.5B active per token. It's text-only (no multimodal features) and trained specifically on natural language and code data. According to JetBrains' technical report, inference time is less than half that of comparable-sized models, which is the primary design target.6

Releasing entity: JetBrains
Release date: June 2, 2026
Availability: HuggingFace, Apache 2.0 license
Intended use cases: Agent routing, low-latency RAG pipelines, sub-agent tasks in multi-step workflows, on-device / private deployment

The original Mellum was a code-completion model; Mellum2 expands to general natural language and code reasoning. JetBrains frames it as a "focal model" — not a frontier replacement, but a fast component for the high-frequency repetitive tasks in modern agent stacks.

OpenAI: ChatGPT conversion ads launch June 5

OpenAI notified advertisers in the ChatGPT Ads Manager beta that conversion-optimized campaigns will go live on June 5.7 Early access requires at least one active conversion event (via JavaScript Pixel or Conversions API) set up before June 1.

The advertising product launched to all US businesses on May 5 with no minimum spend. The platform has since added custom audience targeting (May 14), daily budgets and ZIP-code-level geo (May 22), and now conversion optimization — six product updates in 26 days. According to Criteo data cited in the announcement, ChatGPT's retail conversion rates are running close to double those of traditional search, though that figure comes from Criteo's own customer reporting and hasn't been independently audited.7

The ad business generated approximately $100M in annualized revenue in its first six weeks.

Florida sues OpenAI over ChatGPT safety

Florida became the first U.S. state to file a lawsuit against OpenAI, naming CEO Sam Altman as a co-defendant. The complaint, filed Monday, alleges that ChatGPT poses a safety risk to minors and that OpenAI has knowingly marketed the product as safe while aware of those risks.8

Plaintiff: State of Florida
Defendants: OpenAI, Sam Altman
Core allegation: Failure to warn users of safety risks; misrepresentation of product safety
Status: Lawsuit filed; no court date set

Intel Crescent Island: air-cooled inference GPU shipping in limited quantities this year

Intel announced that its Crescent Island inference GPU will begin limited shipments before year-end. The chip targets AI inference workloads specifically — not the training market Nvidia dominates. Key design choices: LPDDR5 memory instead of HBM, air cooling instead of liquid cooling. Both choices reduce infrastructure costs for buyers and cut Intel's own bill of materials.9

Intel plans to manufacture the chip in its own fabs. If yields hold, the cost advantage compounds. The company also said Crescent Island will have a variant compliant with U.S. export controls for China — a market Nvidia has effectively ceded through successive rounds of regulatory restrictions.

Releasing entity: Intel
Release date: Limited shipments by end of 2026
Pricing: Below Nvidia/AMD data center AI GPU pricing (specific numbers not disclosed)
Previous state: Intel had no commercially competitive inference-focused AI GPU

Strava puts its developer API behind an $11.99/month paywall

Fitness platform Strava has begun requiring a paid subscription for all developer API access, citing zero-code AI applications that were scraping fitness data at scale.10 Public profiles and fitness club listings are also moving behind authentication. The $11.99/month fee applies to all developers regardless of use case.

This follows a pattern seen at Twitter/X, Reddit, and other platforms that have monetized previously free API tiers in response to AI-driven usage spikes. Strava's reasoning attributes the change directly to AI app automation rather than human developers.