cryptobriefing.com
AI Product Updates Daily — May 22, 2026
OpenAI files a confidential IPO S-1 targeting a September debut at up to $1 trillion. An internal OpenAI reasoning model disproves Erdős' 80-year-old unit distance conjecture. Gemini Omni launches with conversational video generation. Claude Compliance API draws Cloudflare, Netskope, and Relativity integrations. Trump pulls a landmark AI executive order hours before signing.
Three headline-sized events landed on the same morning: OpenAI quietly filed a confidential IPO prospectus, SpaceX's freshly public S-1 confirmed the full scope of Anthropic's compute deal, and an internal OpenAI reasoning model settled an 80-year-old open problem in mathematics. Below is the full sweep.
OpenAI files confidential IPO S-1
OpenAI submitted a confidential draft registration statement to the SEC as early as Friday, May 22, according to reporting from multiple outlets 1 2. Goldman Sachs, Morgan Stanley, and JPMorgan Chase are serving as underwriters. The target window for a public debut runs from Labor Day to Thanksgiving 2026 — most likely September. The S-1 itself will not be publicly visible until approximately 15 days before the roadshow begins.
Current private-market valuation sits at roughly $852 billion, with some analysts projecting the IPO could price near $1 trillion — which would make it the largest technology IPO on record 3. OpenAI's annualized revenue is $25 billion. Forecasters currently put the probability of OpenAI listing before Anthropic at about 85%, with the two companies' windows staggered roughly a month apart (September vs. October).
リンクプレビューを読み込んでいます…
The filing coincides with SpaceX's S-1 hitting SEC EDGAR two days earlier, on May 20. That document confirmed that Anthropic is paying SpaceX $1.25 billion per month — $15 billion per year — for compute capacity through May 2029, leasing Colossus infrastructure in Memphis 4. The contract effectively subsidizes SpaceX's AI division and pushes the unit toward breakeven; SpaceX's AI segment ran an operating loss of $2.47 billion in Q1 2026 on $8.18 billion in revenue.
OpenAI reasoning model cracks 80-year-old math problem
On May 22, OpenAI published a proof that a general-purpose internal reasoning model had resolved the planar unit distance problem — a conjecture Paul Erdős first posed in 1946 5.
The problem asks: among n points placed in a plane, what is the maximum number of pairs that can be exactly distance 1 apart? Decades of work had led most mathematicians to believe the answer scaled only slightly faster than linear — Erdős' own conjecture set an upper bound of n^{1+o(1)}. The OpenAI model disproved that bound. It constructed an infinite family of point configurations achieving at least n^{1+δ} unit-distance pairs for a fixed positive δ. A Princeton mathematician subsequently refined the constant to δ = 0.014.
The proof was checked by a group of external mathematicians, and companion remarks were published by Fields medalist Tim Gowers, who called it "a milestone in AI mathematics." Number theorist Arul Shankar wrote: "current AI models go beyond just helpers to human mathematicians — they are capable of having original ingenious ideas, and then carrying them out to fruition."
What makes the result technically surprising is where the key ideas came from: algebraic number theory — specifically infinite class field towers and Golod–Shafarevich theory — applied to an elementary geometry question. The model was not trained for mathematics specifically, was not given scaffolding targeted at this problem, and produced the proof as part of a broader evaluation of Erdős problems.
リンクプレビューを読み込んでいます…
Google launches Gemini Omni
Google published the official Gemini Omni announcement on May 22, detailing capabilities that go beyond the I/O keynote teaser 6. The first Gemini Omni model — Gemini Omni Flash — handles generation and editing from any combination of text, image, audio, and video input.
Key capabilities:
- Conversational video editing: natural language edits that preserve character consistency, physical plausibility, and scene continuity across iterations
- Physics-grounded generation: the model's world knowledge informs outputs — motion, fluidity, and spatial relationships are more accurate than in earlier video models
- Avatars: users can create a digital likeness to generate videos in their own voice and appearance; audio-driven editing is still in testing
- SynthID watermarking: every Gemini Omni video carries an imperceptible watermark verifiable through the Gemini app, Chrome, or Google Search
Availability: live now for Google AI Plus, Pro, and Ultra subscribers via the Gemini app and Google Flow. YouTube Shorts and YouTube Create users get free access starting this week. API access for developers and enterprise customers follows in the coming weeks.
Gemini 3.5 Flash in the real world
PCMag put Gemini 3.5 Flash through hands-on testing on May 22 and found a model that genuinely outpaces GPT-5.5 on coding speed — completing tasks "far faster" — but at the cost of reliability: instruction-following errors were noticeably more frequent 7. The Pro plan's usage quota ran out within 15 minutes of intensive use, requiring a 6-hour reset.
On the API side, confirmed pricing is $1.50 / $9 per million input/output tokens — cheaper than Claude Sonnet 4.6 ($3 / $15) but 3× more expensive than Gemini 3 Flash's preview pricing ($0.50 / $3) 3. Developer benchmarks on five coding and agent tasks put it above both Claude Opus 4.7 and GPT-5.5, though it trails flagship models on deep knowledge and long-context tasks.
Meanwhile, data from OpenRouter — a third-party model routing platform — shows Chinese AI models now account for over 60% of developer usage, up from roughly 1% in 2024. Running the same 10-benchmark evaluation suite costs $544 with GLM-5.1, $948 with Kimi K2.6, and $4,811 with Anthropic Claude. Anthropic publicly acknowledged in a policy paper released this month that US frontier models lead Chinese counterparts by "a matter of months" — a framing aimed at prompting additional government investment in compute and policy.
Claude Compliance API: Cloudflare, Netskope, Relativity all integrate
Three enterprise security and legal-tech vendors announced integrations with Anthropic's Claude Compliance API on May 22, suggesting a coordinated enterprise rollout.
Cloudflare CASB: Cloudflare extended its Cloud Access Security Broker to scan Claude Enterprise and Claude Platform activity without requiring endpoint agents or inline traffic inspection 8. Security teams can detect risky file uploads, chat messages that violate DLP policies, shared projects, and API key creation events, then trigger remediation directly from existing Gateway policies. The first two integrations are free for new Cloudflare customers.
Netskope: Netskope One gained Claude Enterprise coverage through the same API, extending DLP, UEBA, and compliance-mapping (GDPR, HIPAA, SOC 2, PCI-DSS, NIST 800-53) to Claude activity 9. Netskope's own data shows enterprise Claude usage rose from 56.2% to 94.9% of organizations between April 2025 and April 2026. The integration enters private preview in June.
Relativity: The legal data intelligence platform added native Claude Enterprise data collection to RelativityOne, joining ChatGPT Enterprise and Gemini Enterprise as supported sources 10. Collected conversations and files can be analyzed alongside other enterprise data using the company's Relativity aiR legal AI tools, supporting legal holds, e-discovery, and governance workflows.
The wave of integrations follows Anthropic's May 21 announcement of acquisitions, compute deals, and first-profit projections — and positions Claude Enterprise as an auditable, compliance-ready option for regulated industries.
リンクプレビューを読み込んでいます…
Trump pulls AI executive order hours before signing
President Trump abruptly canceled a planned White House signing ceremony for an AI executive order on the afternoon of May 21, hours after AI executives had been briefed on the policy and invitations were sent 11 12. Trump told reporters: "I didn't like certain aspects."
The order would have created a voluntary 90-day government review window for frontier AI models before release, with NSA participation in classified testing. The internal split was between officials who wanted the framework to remain entirely voluntary — to avoid impeding US competitiveness — and national security officials who wanted mandatory NSA access. No revised signing date has been announced.
Anthropic Q2 financials: first quarterly profit confirmed
Anthropic disclosed Q2 2026 projections to investors this week, showing revenue of $10.9 billion (up 130% from Q1's $4.8 billion) and operating profit of approximately $559 million — the company's first quarterly profit, two years ahead of its original internal timeline 3.
Compute cost efficiency improved: from $0.71 per dollar of revenue in Q1 to $0.56 in Q2. The number of enterprise customers spending over $1 million annually doubled from 500 to more than 1,000 between February and April 2026. Claude Code and its developer ecosystem are the primary growth driver, generating recurring, high-margin API revenue from engineering teams.
Anthropic's private-market valuation sits at $90 billion ahead of an expected Q4 2026 IPO. The company confirmed it will expand GPU capacity at Colossus 2 (with NVIDIA GB200 / Blackwell Ultra) starting in June, which it expects to enable further relaxation of Claude Code's rate limits.
参考ソース
- 1OpenAI prepares confidential IPO filing with Goldman Sachs and Morgan Stanley
- 2OpenAI set to file confidential IPO prospectus within weeks
- 3AI News Today — May 22, 2026
- 4SpaceX IPO live updates — Anthropic compute deal disclosed
- 5An OpenAI model has disproved a central conjecture in discrete geometry
- 6Introducing Gemini Omni
- 7Do Google's I/O AI Announcements Live Up to the Hype? I Put Them to the Test
- 8Announcing Claude Compliance API support with Cloudflare CASB
- 9Netskope announces integration with Claude Compliance API
- 10Relativity adds collection of Claude Enterprise data with Claude Compliance API integration
- 11Trump postpones AI executive order signing
- 12Trump delays AI security executive order
このコンテンツについて、さらに観点や背景を補足しましょう。