Date range: 2026-05-18 (Mon) → 2026-05-24 (Sun) Compiled: 2026-05-28 Sources: Anthropic / OpenAI official, Smol AI News, Latent Space, QbitAI, IT Home, NVIDIA Newsroom, SpaceX S-1, and more.
🌐 Lead story
"Harness engineering replaces the base model as the new moat" + "Gemini is back in the top tier" + "Chinese vendors lock the price war in as permanent".
Google I/O ate half the week's airtime — Gemini 3.5 Flash actually outranks the previous-gen Pro tier — and on the same day Karpathy left Tesla / his independent project to join Anthropic, which multiple newsletters called "the most important talent move of 2026 so far." OpenAI pushed Codex onto Dell hardware and got out front on content provenance (C2PA + SynthID). The fiercest action came from China — DeepSeek closed a 70B-yuan round and made its 75%-off V4-Pro pricing permanent, converting "limited-time discount" into "hard-coded commitment." On hardware, the biggest signal was SpaceX's S-1 — the first SEC-grade disclosure that Anthropic pays xAI $1.25B/month to rent Colossus, putting concrete numbers on AI-to-AI compute leasing for the first time.
⭐ Personal Top 5 (try today)
Ranking by: substantive capability gain × low barrier to entry × direct cost-down or productivity lift
| Rank | Item | One-liner |
|---|---|---|
| 🥇 1 | DeepSeek V4-Pro API permanent 75% cut | Cached-hit input at 0.025 yuan / M tokens — basically "call as much as you want" for individuals. The single biggest cost-saver of the week. |
| 🥈 2 | Cursor Composer 2.5 (on Kimi K2.5) | $0.50 / $2.50 per M tokens, coding ability on par with Claude Opus 4.7 — 1/10 of Opus pricing. Individual Cursor users should switch the default model now. |
| 🥉 3 | OpenAI Codex CLI Goal mode GA + Appshots | Goal mode is now stable for multi-turn persistent tasks; Appshots one-shots a screen window into a Codex thread — visual + code context in sync. Daily cc / codex users should upgrade to v0.133+. |
| 4 | Google Gemini 3.5 Flash + Omni publicly available | Flash now runs at the old Pro tier's level, $1.50/$9; Omni multimodal is on for all paid users. The cheapest "natural language + long context + actually runnable" option. |
| 5 | xAI Grok "Skills" | Reusable expertise across sessions. If you use Grok 4.3, this is the switch that turns it from "chat" into "assistant with memory." |
🏢 Team / Enterprise Top 5 (act on this week)
Ranking by: clear security/compliance gain × deployment breadth × strategic signal strength
| Rank | Item | One-liner |
|---|---|---|
| 🥇 1 | Anthropic Self-hosted Sandboxes (public beta) + MCP Tunnels | Managed Agents tool execution can sit on your own infra / Cloudflare / Vercel; private MCP comes in via tunnel. The "sensitive code can't leave the network" blocker is finally gone. |
| 🥈 2 | Anthropic Compliance API + 28 security integrations | Cloudflare/CrowdStrike/Microsoft Purview/Okta/Palo Alto/Wiz/Zscaler all in the launch set. The last piece for finance / healthcare / government to deploy Claude Enterprise. |
| 🥉 3 | OpenAI × Dell — Codex on-prem | Codex onto Dell AI Factory / AI Data Platform; 4M weekly active Codex developers disclosed. Critical for data-residency / sovereign-cloud customers. |
| 4 | DeepSeek V4-Pro permanent cut + 70B-yuan raise | Team-scale API spend drops hard; the raise also hard-codes the strategy "no IPO, no revenue chase, peg the price curve open-source-down." Long-term procurement assumptions need re-modeling. |
| 5 | SpaceX S-1 discloses: Anthropic pays xAI $1.25B/month for Colossus | Gives every CTO / CFO a concrete number that the compute-leasing market is real. The "self-build vs rent GPU over the next 3 years" decision needs to be re-modeled. |
🔵 Top Two
Anthropic / Claude Code
| Date | Event | One-liner |
|---|---|---|
| 05-18 | Stainless acquisition | The SDK + MCP-server generation platform is folded in; existing Stainless customers keep their artifacts but the service shuts down, and Anthropic internalizes the "enterprise API ↔ agent" translation layer.4 |
| 05-19 | Code with Claude London blowout | Managed Agents gets dreaming / multi-agent / outcomes / webhooks; Self-hosted Sandboxes (public beta) — tool execution can run on your own infra or on Cloudflare / Daytona / Modal / Vercel; MCP Tunnels (research preview) lets private MCP servers connect in. This salvo removes the single biggest enterprise blocker for Managed Agents. (recap) |
| 05-19 | KPMG global alliance | Claude is embedded in KPMG's "Digital Gateway" delivery platform across 138 countries; 276,000 KPMG employees get Claude; KPMG becomes Anthropic's preferred PE partner.5 |
| 05-19 | Andrej Karpathy joins the pre-training team | Former OpenAI co-founder + Tesla AI lead, reporting to Nick Joseph, with a goal of using Claude itself to accelerate pre-training research. Newsletter editors broadly called it "2026's biggest hire."3 |
| 05-21 | Compliance API + 28 security integrations | DLP / SIEM / CASB / identity tools can programmatically access Claude Enterprise conversation content and activity events; launch partners include Cloudflare, CrowdStrike, Microsoft Purview, Okta, Palo Alto, Wiz, Zscaler. The last puzzle piece for regulated industries to put Claude into production.9 |
| 05-22 | Project Glasswing first results: Claude Mythos finds 10,000+ high-severity vulns | Across critical OSS, 10,000+ critical/high CVE-class issues discovered; Opus 4.7 has already auto-patched 2,100+. Claude Security also moves to public beta. The largest real-world demo to date of Opus 4.7 doing autonomous security work.8 |
| Ongoing | New $30B round / $900B valuation | Sequoia / Dragoneer / Altimeter / Greenoaks each at ~$2B; first time Anthropic's valuation exceeds OpenAI's ($852B). (Bloomberg) |
OpenAI / Codex
| Date | Event | One-liner |
|---|---|---|
| 05-18 | OpenAI × Dell | Codex integrates into Dell AI Data Platform + Dell AI Factory, entering hybrid / on-prem enterprise deployment; the same announcement disclosed that Codex now has 4M+ weekly active developers.6 |
| 05-18 | Codex CLI v0.131.0 | Richer TUI controls: service-tier command, blended token usage, permissions/approval mode, workspace roots, responsive Markdown tables. (releases) |
| 05-19 | C2PA + SynthID watermarking + public verification tool | OpenAI becomes a C2PA Conforming Generator; images from ChatGPT / Codex / API carry invisible SynthID watermarks (joint with Google DeepMind); anyone can run the public tool to check whether an image was OpenAI-generated. First major generator to ship both metadata and pixel watermarking.7 |
| 05-20 | Codex CLI v0.132.0 | Python SDK becomes a first-class auth target: API key, ChatGPT browser flow, device-code. |
| 05-21 | Codex CLI v0.133.0 — Goal mode goes GA | Goal mode graduates from experimental to default-on (CLI / IDE extension / app), with persistent storage and multi-turn progress tracking; same release ships Appshots (Cmd-Cmd attaches the current macOS window to a Codex thread) and remote-lockscreen Mac computer use. Codex's official answer to long-running tasks. |
| 05-22 | General OpenAI model disproves Erdős unit-distance conjecture | A general-purpose (non-math-specialist) model produces a counterexample to a ~80-year-old central conjecture in discrete geometry; externally verified. A landmark "general reasoning models can produce original frontier mathematics" event. (source) |
🟢 Chinese camp
DeepSeek — a one-two punch this week
- 05-22 ｜ 70B-yuan first round closes — Bloomberg cites people familiar: target raised from 50B to 70B yuan; China National Investment Fund ~10B, Tencent / IDG / Monolith etc. participating, Liang Wenfeng personally invested ~20B; valuation reportedly climbing toward ~350B yuan. Liang continues to hold the "open source + research first" line. (Sina Finance)
- 05-22/23 ｜ V4-Pro API permanently priced at 25% of original — When the "limited-time discount" ends on May 31 it becomes the permanent price: cached-hit input at 0.025 yuan / M tokens, miss at 3 yuan, output at 6 yuan. New global low for any flagship model.12, IT Home)
- Around 05-20 ｜ In-house Code Harness team — DeepSeek's Beijing office is recruiting a "Code Harness" team and former Jane Street engineer Tianyi Cui is confirmed as tech lead. Smol AI reads this as the China-side validation of the "harness engineering is the new differentiator" theme. (SCMP)
Moonshot / Kimi
- 05-18 ｜ Cursor Composer 2.5 runs on Kimi K2.5 — Cursor ships Composer 2.5: SWE-Bench Multilingual 79.8%, Terminal-Bench 2.0 matching Claude Opus 4.7, priced at $0.50 / $2.50 per M tokens (≈1/10 of Opus). The model card explicitly names Moonshot Kimi K2.5 open weights as the base. First time Kimi's open-ecosystem play is adopted as the default backbone of a front-line Western coding IDE.14
- 05-19/20 ｜ Unwinding VIE, Hong Kong IPO path becomes clearer — Moonshot emails shareholders to start the VIE unwind, lining up with new regulator language to clear the runway for an HKEX IPO. Backdrop: an early-May ~$2B round (led by Meituan Longzhu) closed, post-money valuation over $20B. (Tencent News)
Zhipu / GLM
No major first-party product or pricing news from Zhipu this week. The widely cited "GLM-5.1 takes the open-source #1 on Artificial Analysis Coding Agent" piece is 5/12 old news and does not count toward this week.
Chinese media watch
- QbitAI's first-party framing — "V4.1 set for June, supports image + audio multimodal, strengthened MCP protocol" — continues to spread in the Chinese community. (source)
- Securities Times, Sina Finance, Guangzhou Daily, IT Home all run "permanent price cut + 70B-yuan raise" combo recaps, framing it as the trigger for an H2 industry shakeout.
🟠 Hardware & Compute & Power
NVIDIA
- 05-20 ｜ Q1 FY27 earnings: revenue $81.6B / data center $75.2B (YoY +92%) / additional $80B buyback — the new "ACIE" segment contributes $37B; hyperscalers contribute >$38B; dividend goes from $0.01 to $0.25 (25x). Confirms AI infra spend is still pushing the ceiling.10
- 05-23 ｜ Jensen arrives in Taiwan a week early — warming up for the 6/1 GTC Taipei keynote and 6/2-5 Computex. This is not a Computex week (note: Computex 2026 is actually 6/2-5), but Jensen touring the supply chain has clearly pre-seeded next week's big stories. (source)
xAI / Tesla / Musk
- 05-20 ｜ SpaceX S-1 filing reveals the full xAI numbers —
- xAI to buy an additional $2.8B of gas turbines (incl. $2B mobile turbines — the model NAACP is suing over)
- Anthropic rents xAI's Colossus 1 data center, $1.25B/month, contract through May 2029
- xAI burned $6.4B in 2025
- First SEC-grade disclosure of inter-AI-company compute leasing finances. (Data Center Dynamics, TechCrunch)
- 05-21 ｜ xAI ships Grok "Skills" — reusable custom expertise that persists across Grok 4.3 web/iOS/Android sessions. Grok moves from chat tool to "productivity platform" against Custom GPTs / Anthropic Skills. (summary)
Other hardware / compute / power
- 05-24 ｜ Big Tech 2026 capex ≈ $725B total / Microsoft CFO attributes $25B of the jump to memory chips — YoY +77%. Microsoft explicitly traces the $25B increment to HBM/DRAM price increases. HBM shortage has graduated from a supply story to a quarterly-earnings-level cost variable. (Tom's Hardware)
🌍 Other vendors
Google — I/O 2026 mega-burst (5-19)
- Gemini 3.5 Flash GA: 1M context, Terminal-Bench 2.1 at 76.2%, 4x faster than peers (12x inside Antigravity), $1.50 / $9 per M tokens. First time a Flash tier outranks the previous-gen Pro tier on agentic tasks. (source)
- Gemini Omni: any modality ↔ any modality (text/image/video/audio in and out), with physical-simulation ability; Omni Flash rolls out to all paid Gemini users. Veo-class video generation is now in the main line.
- Gemini Spark: an always-on cloud-resident personal agent that keeps working when your device is off.
- Antigravity 2.0: org-level agent platform built on 3.5 Flash. Google's first competitive answer to ChatGPT-Agent / Claude Code background tasks.
Alibaba — Qwen3.7-Max (5-19/20 announced)
- 1M context, AAII 56.6; beats Claude Opus 4.6 Max on Terminal-Bench 2.0 / SWE-Bench Pro / MCP-Atlas; $2.50 / $7.50 per M tokens; 1000+ sequential tool calls, 35-hour autonomous runs. China's direct response to I/O.16
Cohere — acquires Reliant AI (5-19)
- Second acquisition this month (after Aleph Alpha); picks up the biopharma agent platform "North for Pharma." Cohere is betting on vertical sovereign AI against the horizontal frontier race. (source)
Vercel × Perplexity (5-21)
- Perplexity Search/Ask/Research/Reason become first-class platform integrations on Vercel. (source)
Research highlights
- Tencent L2P / Z-Image 6B pixel-space generation — 1k resolution, no VAE, transfer-learned from Flux/SD on only 8 GPUs to ~93% of GenEval. Shakes the SD1.5-era latent-diffusion paradigm. (report)
🧭 Trend analysis
1. "Harness > base model" is being broadly validated
This week, from Anthropic (Mythos finds 10k vulns), OpenAI (Codex Goal mode GA + Appshots), Google (Antigravity 2.0), Alibaba (Qwen3.7-Max's 35-hour continuous runs), to DeepSeek (in-house Code Harness team), every top player is piling engineering into "agent loops that use tools well" instead of shipping yet another base model. Smol AI flat out named the theme "harness engineering is the new differentiator."
2. China's storyline this week isn't model parity — it's pricing and capital structure
DeepSeek converts a discount into a hard commitment + Kimi unwinds its VIE to prepare for an HKEX IPO + Moonshot's weights get adopted by Cursor as the commercial default — none of this week's main Chinese AI story is about benchmark parity. The "valuation round + IPO prep" rhythm in the West sits on top of different capital structures and power / silicon constraints. Reading "why this path" requires looking at finance + hardware + policy together.
3. AI-to-AI compute leasing is becoming an asset class
SpaceX's S-1 putting "Anthropic pays xAI $1.25B/month" into an SEC filing is the first hard number. Expect more — the three-tier structure of compute supplier (NV/HBM/power) → card-rich operator (xAI/CoreWeave) → training shop (Anthropic/OAI/Zhipu) is solidifying.
4. Provenance has hit the "no-provenance = political cost" stage
OpenAI joining C2PA + adding SynthID watermarks is the industry's first combo of metadata + pixel watermarking shipped together. Anthropic and Google will follow. For consumer / creator / media use cases, expect a 2026 turning point where "no provenance metadata = presumed suspicious".
📚 Further reading
- Smol AI News — I/O day recap: https://news.smol.ai/issues/26-05-19-not-much/
- Latent Space — I/O deep dive: https://www.latent.space/p/ainews-google-io-2026-gemini-35-flash
- Jack Clark, Import AI 458 (publishes in W22 but continues the W21 thread): https://importai.substack.com/p/import-ai-458-reckoning-with-the