AI Product Updates Daily — May 28, 2026

Anthropic disclosed a $65 billion funding round at a near-$1 trillion valuation on the same day it shipped Claude Opus 4.8. Mistral renamed its assistant "Vibe" and staked a claim in industrial manufacturing. Apple's iOS 27 overhaul leaked renders of a standalone Siri app built on Gemini. A busy 24 hours across frontier models, developer tools, and platform features.

Anthropic: $65B Series H at $965B valuation

Anthropic raised $65 billion in a Series H round led by Altimeter Capital, Dragoneer, Greenoaks, and Sequoia Capital, valuing the company at $965 billion post-money. 1

Co-leads also included Capital Group, Coatue, D1 Capital Partners, GIC, ICONIQ, and XN. Strategic infrastructure partners — Micron, Samsung, and SK hynix — joined the round, with $15 billion of the total from previously committed investments by hyperscalers including $5 billion from Amazon. The company said its run-rate revenue crossed $47 billion earlier in May, up from its prior funding round in February.

CFO Krishna Rao said the funding would go toward safety and interpretability research, compute expansion, and scaling Claude-based products and partnerships. Anthropic has signed compute agreements with Amazon (up to 5 GW), Google and Broadcom (5 GW of next-generation TPU capacity), and SpaceX for access to Colossus 1 and 2 GPU clusters. Claude is now available on AWS, Google Cloud, and Azure.

anthropic.comhttps://www.anthropic.com/news/series-h外部链接

正在加载内容卡片…

Anthropic: Claude Opus 4.8 ships with dynamic workflows and effort control

Claude Opus 4.8 is available today at the same price as Opus 4.7: $5 per million input tokens and $25 per million output tokens. Fast mode — which runs at 2.5× speed — is now priced at $10/$50 per million tokens input/output, three times cheaper than fast mode was under prior Opus versions. 2

Benchmark notes from the release:

Online-Mind2Web: 84% (browser agent use), above both Opus 4.7 and GPT-5.5
Anthropic's alignment team rated Opus 4.8 roughly four times less likely than Opus 4.7 to let flaws in generated code pass unremarked
On Databricks' Genie agent, the model unlocks at 61% cheaper token cost than Opus 4.7

Three additions shipped alongside the model:

Dynamic workflows (Claude Code, research preview) — Claude can spin up hundreds of parallel subagents within a single session, verify outputs, and run codebase-scale migrations end-to-end. Available on Enterprise, Team, and Max plans.
Effort control (claude.ai and Cowork, all plans) — a new slider lets users choose between "high" (default), "extra," and "max" effort levels, trading response speed against quality.
Messages API mid-task system entries — developers can now update system-prompt instructions during an agent run without breaking prompt cache. Useful for updating permissions, token budgets, or environment context on the fly.

On alignment, Anthropic said Opus 4.8 "reaches new highs on our measures of prosocial traits like supporting user autonomy" and has lower rates of deceptive or misuse-enabling behavior than Opus 4.7, roughly matching Claude Mythos Preview. The company added it expects to make Mythos-class models generally available "in the coming weeks" once it completes the required cyber safeguards.

Opus 4.8 benchmark comparison table from Anthropic's launch post — Opus 4.8 benchmark comparisons against predecessor models 2

OpenAI: Frontier Governance Framework published

OpenAI published its Frontier Governance Framework — a public governance document explaining how its existing Preparedness Framework maps to California's Transparency in Frontier AI Act and the EU AI Act's Code of Practice for General Purpose AI. 3

Coverage areas in the document include cyber offense risk assessment, CBRN risk mitigation, harmful manipulation, loss-of-control scenarios, model reporting, security risk management, and incident response protocols. OpenAI characterized the Preparedness Framework as the internal foundation; this published document is the outward-facing governance layer for regulatory purposes. The company said it will update the framework as model capabilities and regulations evolve.

Mistral: "Vibe" replaces Le Chat; industrial AI and data center push

At its inaugural AI NOW Summit in Paris, Mistral renamed Le Chat to Vibe and positioned it as a unified agent platform for enterprise productivity and software development. 4

Vibe has two modes:

Vibe for Work: connects to Google Workspace, Outlook, SharePoint, Slack, and GitHub for multi-step tasks (email summaries, spreadsheet analysis, scheduled workflows).
Vibe for Code: a coding agent available on web, VS Code extension, and CLI — builds features, fixes bugs, ships pull requests.

Pricing: free tier, $14.99/month Pro, $24.99/user/month Teams, custom Enterprise.

Model consolidation: Mistral deprecated Pixtral, Magistrale, and DevStral as standalone products, folding their capabilities into Mistral Medium 3.5. Mistral Large 4, targeting industrial applications like fluid dynamics and computational chemistry, is due "during the summer." A 1-billion-parameter OCR model called Mr. Lossier can process thousands of pages per minute on a single GPU.

Industrial AI: Mistral announced a partnership with Airbus across its commercial, helicopter, defense, and space divisions, and is partnering with BMW Group on its "Large Industry Model" initiative for crash simulation. The underlying technology comes from Mistral's acquisition of Emmi AI in May 2026, which adds physics simulation capabilities trained to predict physical behavior in seconds rather than running full solvers.

Infrastructure: Mistral is investing €4 billion in French and Swedish data centers. A new 10 MW inference facility at Les Ulis (south of Paris) opens Q3 2026. A Sweden site hosting NVIDIA Vera Rubin GPUs is planned for 2027. The company is also targeting €1 billion in revenue for 2026.

Mistral: model consolidation details

Alongside the Vibe announcement, Mistral published its new model overview:

Model	Status	Key use
Mistral Medium 3.5	Current flagship	Natively multimodal, absorbs Pixtral / Magistrale / DevStral
Mistral Large 4	Due summer 2026	Industrial / physics AI
Mr. Lossier (1B params)	Available	OCR, thousands of pages/minute on single GPU
Voxtral	Available	STT + TTS with voice cloning; real-time "duplex" model coming

Figma Make: two-way GitHub sync, production code from canvas

Figma added two-way repository sync to Figma Make, its AI design-to-code product. Designers can now import an existing Git repository into the Figma desktop app, edit the app's code visually, and push changes back through standard GitHub pull requests — subject to the same CI pipelines and security checks as any other commit. 5

Previously, Figma Make could export to a new GitHub repository but could not sync with an existing codebase. The update is available to Full seats on paid plans ($16–$90/month). The model stack behind the feature toggles between Anthropic's Claude 3.6 Sonnet, Claude Opus, and Google Gemini depending on the task. Figma stock (NYSE: FIG) has fallen roughly 81% from its July 2025 IPO-day high to around $21–$22 by May 2026.

Apple: standalone Siri app with Gemini inside, heading to iOS 27

Bloomberg published leaked renders of Apple's planned iOS 27 AI overhaul ahead of WWDC in June. 6

Key details:

A new standalone Siri app — designed to compete with ChatGPT, Claude, and Gemini — will surface chat history and allow document and photo uploads.
The standard Siri voice trigger will respond via the Dynamic Island (the pill-shaped cutout) for quick queries.
Swiping down to Spotlight Search will use an AI-powered Siri backed by Google Gemini under the hood, per Apple's agreement announced in January 2026.
Results appear in card-style formatted text.

ChatGPT had 900 million weekly active users as of February 2026; Apple's installed base across all devices is 2.5 billion.

Asana acquires StackAI for $75M

Asana acquired no-code agent-builder StackAI for $75 million. 7

StackAI, a YC Winter '23 company, builds AI workflow automation that connects to Salesforce, Slack, and Google Workspace. It had raised just under $20 million total, most recently a $16 million Series A. Founders Tony Rosinol and Bernard Aceituno will join Asana.

Asana CEO Dan Rogers framed the acquisition as part of the company's "human-agent teams" platform strategy: "StackAI lets them go further, agentifying the most complex business processes end-to-end." Asana's stock has lost more than half its value since ChatGPT launched in late 2022.

Sesame launches iOS app: four conversational AI agents, free in 39 countries

Sesame, the conversational AI startup co-founded by Oculus founders (including Palmer Luckey), launched its iOS app in 39 countries. 8

The app features four AI agents — Maya, Miles, Simone, and Charlie — each with distinct voices, personalities, and persistent memory. Sesame's differentiator is mid-sentence course correction: the AI can run multiple parallel searches while speaking and weave in new results without stopping, designed to feel more like conversation than query-response. Maya and Miles had over one million users during the beta period, per Sequoia.

Features include search cards with images, notes, texting mode, deep-dive mode, and incognito mode (no memory saved). The full experience is free for now; there may be a waitlist. Android preview is coming. The company is targeting intelligent eyewear for 2027.

Sesame conversational AI agent app screenshot on iPhone — Sesame's iOS app puts four AI agents — Maya, Miles, Simone, and Charlie — in a natural conversational interface 8

YouTube adds AI podcast recommendations and Auto speed for Premium users

YouTube announced new podcast features for Premium subscribers: 9

AI podcast recommendations: personalized suggestions based on genre, mood, or shows you already listen to (similar to the existing "Ask Music" feature for playlists).
Auto speed: dynamically adjusts playback speed during slower-speech or information-dense segments without jarring the listener.
On-the-go mode: optimized controls for background listening — skip forward/back, jump to next episode.

Auto speed and on-the-go mode are live on Android for Premium users; iOS rollout is coming. YouTube says Premium users watched 800 million hours of podcasts in April 2026, with YouTube Podcasts at over 1 billion monthly active users.

DeepSeek: 75% price cut made permanent

DeepSeek made its earlier temporary 75% price reduction on V4 Pro permanent. 10

Updated pricing:

DeepSeek V4 Pro: $0.435/million input tokens, $0.87/million output tokens; prefix cache reads at $0.003625/million tokens (87× cheaper than Western cloud cache reads)
DeepSeek V4 Flash: 10–25× cheaper than comparable entry-level models like Claude Haiku

On OpenRouter, DeepSeek V4 Flash rose to the top spot for token usage over the past week (+48%). DeepSeek's three models collectively processed nearly 6 trillion tokens in a week. The company said the price reduction is sustainable due to architectural efficiency gains and caching optimizations, with all models released under MIT license for self-hosted deployment.

Sources: 1 2 3 4 5 6 7 8 9 10