AI Product Updates Daily — May 29, 2026

The biggest thread running through May 29: frontier labs are simultaneously shipping better models and publicly retreating from the most extreme claims they made about those models. Claude Opus 4.8 posts real benchmark gains while Anthropic quietly walks back its "Mythos will be widely available soon" stance into a multi-week waiting room. Sam Altman tells an audience in Sydney he was "completely wrong" about AI displacing entry-level jobs quickly. And Microsoft — OpenAI's biggest enterprise ally — cancels most of its internal Claude Code licenses because token costs spiraled out of control before anyone checked the invoice.

OpenAI

GPT-5.5 Instant gets a readability upgrade — and loses Canvas

OpenAI updated GPT-5.5 Instant on May 28 to reduce what the company described as excessive bullet lists and improve overall response naturalness. The same update removed Canvas from GPT-5.5 Instant and GPT-5.5 Thinking entirely. Writing and coding tasks previously handled in the Canvas side panel now run inline in the chat. Paid users can still access Canvas through older models during a transition period.1 2

GPT-4.5 and o3 are headed for retirement

GPT-4.5 will leave ChatGPT on June 27, 2026, after a 30-day transition window. OpenAI o3 follows on August 26 after a 90-day window. Both remain available to paid users in the model selector during those periods; o3 also stays in the API for now.1

Codex comes to Windows — with computer use

ChatGPT for iOS and Android can now start Codex tasks on Windows, following Mac support added earlier this month. Windows Codex gains two capabilities from the Mac version: Computer Use (visual point-and-click control of Windows desktop apps) and remote-start from any iOS/Android or Mac device. The Codex app also adds a profile section showing usage data and token activity. iOS and Android picked up several updates alongside this: side-panel conversations, turn-end diff summaries, archived remote threads, one-tap model switching, and Spotlight/Shortcuts support on iOS.3

ChatGPT and Codex icons side by side on a blue background — ChatGPT + Codex for Windows launches May 29 3

Rosalind Biodefense launches

OpenAI opened the Rosalind Biodefense Program, giving vetted developers and U.S. government partners early access to GPT-Rosalind — a life-sciences-focused model — for biodefense and pandemic preparedness work. The program covers epidemiological modeling, early detection, non-pharmaceutical interventions, and medical countermeasure development. OpenAI says it briefed the White House and several federal agencies ahead of the launch.4 5

Anthropic

Claude Opus 4.8: real gains, same price

Anthropic released Claude Opus 4.8 on May 28, priced identically to Opus 4.7. The headline benchmark numbers against key competitors on release day:

Benchmark	Opus 4.8	Opus 4.7	GPT-5.5	Gemini 3.1 Pro
SWE-Bench Pro (agentic coding)	69.2%	64.3%	58.6%	54.2%
Terminal-Bench 2.1	74.6%	66.1%	78.2%	70.3%
Humanity's Last Exam (w/ tools)	57.9%	54.7%	52.2%	51.4%
OSWorld-Verified (computer use)	83.4%	82.8%	78.7%	76.2%
GDPval-AA (knowledge work)	1890	1753	1769	1314
Finance Agent v2	53.9%	51.5%	51.8%	43.0%

Anthropic says the model generates roughly 75% fewer undetected code defects than 4.7, and shows lower rates of deceptive or misaligned behavior in internal testing — comparable to Mythos Preview on that axis.6

The release also adds Dynamic Workflows in research preview, available to Enterprise, Teams, and Max plan users of Claude Code. Workflows let Claude plan, spin up hundreds of parallel sub-agents, and verify results before returning output — Anthropic positions this for large-scale codebase migrations measured in hundreds of thousands of lines. Users also get an adjustable effort dial to trade speed against reasoning depth and cost.6

Mythos-class models: weeks away, once safety work finishes

Anthropic confirmed that Mythos-class capability will come to all customers "in the coming weeks," but only after additional cybersecurity safeguards are ready. Currently it's available to a small set of organizations through the Project Glasswing access program.7

Anthropic's $965B valuation closes

The $65B Series H led by Altimeter, Dragoneer, Greenoaks, and Sequoia closed at a $965B post-money valuation, putting Anthropic ahead of OpenAI's last public valuation. The round was announced May 28 and widely covered on May 29.8

Claude Opus 4.8 benchmark comparison table — Claude Opus 4.8 benchmark results across key tasks 6

Google

Gemini usage limits get emergency fixes

After Google I/O last week introduced compute-based usage limits refreshing every 5 hours (replacing the old daily prompt cap), many users hit their quotas much faster than expected. Josh Woodward, VP of Google Labs, announced several fixes on May 29:

Prompts using the Flash-Lite model now count as free and won't draw against quota
Failed requests no longer consume quota ("Our system mistakes are on us, not you")
Deep Research mode now shows detailed usage breakdowns before you hit a cap
Single prompts using Gemini 3.1 Pro with large attached files now have a per-prompt quota ceiling
A bug causing Omni video generations to over-drain quota was fixed; Ultra subscribers get double the Omni video generations as compensation
Gemini will remember the last model you used across all sessions

コンテンツカードを読み込んでいます…

Microsoft

365 Copilot May 2026 bundle

Microsoft's official May 2026 Copilot update, published May 29, covers a wide range of products. Key items shipping or rolling out now:

Claude Opus 4.8 in Copilot: available in Copilot Cowork (Frontier) now, rolling to Copilot Chat, Excel, and PowerPoint for licensed M365 users. Suited for complex multi-step tasks and long workflows.
GPT-5.5 Instant in Copilot Chat and Studio: ships as "GPT-5.5 Quick response" in the model picker, replacing GPT-5.3 Instant with cleaner, less verbose answers.
Apple CarPlay: the M365 Copilot app now works hands-free in CarPlay — start or continue conversations, organize schedules, capture ideas.
Copilot Notebooks redesign (rolling June): notebooks now sync with OneNote and let users add Teams meeting transcripts as knowledge sources, generate Excel tables from notebook content, and auto-produce infographics.
Teams call delegation: Copilot can answer calls you can't take, capture the caller's intent, and schedule follow-up appointments via Bookings (Frontier, rolling June).
Word Agent writing suggestions: real-time clarity, tone, and structure suggestions inline in the writing flow (web, available now).
PowerPoint one-click skills: check presentation, visualize slides, and prep for Q&A, available now.
SharePoint custom skills: save multi-step workflows as reusable site-specific agents.
ISO 42001 certification extended to Copilot Studio, GitHub Copilot, Dragon Copilot, and Copilot Health.

Microsoft is building a Copilot super app

Fortune reported May 29 that Microsoft is consolidating GitHub Copilot, Copilot Chat, Copilot Cowork, and its internal "Autopilot" workflow tool into a single app, led by new Copilot chief Jacob Andreou. The project aims to unify the consumer and enterprise Copilot products that have historically run as separate teams. The app is not expected to debut at Build 2026 next week, though Build may reference the "one Copilot" direction. Microsoft CEO Satya Nadella's AI chief Mustafa Suleyman is expected to unveil a new Microsoft-built AI model at Build.11

Context for the stakes: fewer than 4.5% of Microsoft 365 paid users subscribe to Copilot, and GitHub Copilot's 4.7 million paid subscribers face growing pressure from Cursor and Claude Code.

Microsoft cancels most internal Claude Code licenses

Separately, Microsoft's Experiences and Devices division (Windows, M365, Outlook, Teams, Surface) started cutting most internal Claude Code subscriptions in mid-May, with a June 30 deadline. The division's 5,000 engineers were using Claude Code at an 84%–95% monthly adoption rate, running up individual API costs of $500–$2,000 per month and an annualized division-wide bill of $1.5M–$6M. Engineers were redirected to GitHub Copilot CLI. The cut applies only to direct subscriptions — Claude remains available to Azure customers through Microsoft Foundry.12

GitHub

GitHub Copilot moves to AI Credits billing

GitHub confirmed that all Copilot plans will shift to AI Credits consumption billing in the next billing cycle, replacing fixed seat subscriptions. Each tier includes a base Credits allocation; additional usage is billed by token volume. The change is designed to give enterprises per-team budget controls — a direct response to the cost dynamics that drove the Microsoft Claude Code cancellation.12

Nvidia

$6.5B into photonics since March

Nvidia has committed at least $6.5 billion to photonics companies since March 2026, including Lumentum, Coherent, Marvell, and Corning. The bet is on replacing copper-based data links inside AI data centers with light-based interconnects, which can move data faster at lower energy cost. Nvidia's total AI equity investment portfolio in 2026 now exceeds $40 billion.13

EU AI Act

High-risk AI compliance pushed to December 2027

The European Parliament and member states agreed to delay and weaken several EU AI Act provisions. High-risk AI system compliance (covering AI used in law enforcement, critical infrastructure, employment, immigration, and biometric ID) was pushed from August 2026 to December 2027. Developer sandbox requirements moved to August 2027; AI-in-product compliance for mechanical goods moved to August 2028; content watermarking requirements moved to roughly December 2026. SME compliance burdens were reduced, and industrial machinery that already falls under existing product safety laws was carved out.14

The jobs reversal

Both Sam Altman and Dario Amodei publicly stepped back from prior statements about AI-driven job displacement on or around May 29. Altman, speaking at a Commonwealth Bank event in Sydney, said he was "completely wrong" to predict that entry-level white-collar jobs would be disrupted as quickly as he thought; he no longer agrees with the job-apocalypse framing common in the industry. Amodei, who had forecast AI eliminating 50% of white-collar jobs, said he now believes AI is more likely to expand total employment.12

Both reversals land eight days after OpenAI filed its confidential IPO S-1 and two weeks before Anthropic's next major funding announcement cycle, which analysts have flagged as relevant context for the timing.

Yale Budget Lab research published in May showed no statistically significant increase in unemployment among high-AI-exposure roles through March 2026.