1. Anthropic Ships Claude Opus 5 at Half the Frontier Price
Anthropic released Claude Opus 5 on July 24, positioning it just below its frontier Fable 5 system while costing roughly half as much to run. It posts 43.3% on Frontier-Bench (agentic coding), 30.2% on ARC-AGI-3 — about three times the next-best model on that novel-reasoning test — and a 1,861 GDPval-AA v2 Elo for knowledge work. Pricing holds at $5 per million input tokens and $25 per million output, the same as Opus 4.8. It is live on Claude.ai, the API, Claude Code and Cowork, and is now the default on Claude Max.
2. Kimi K3 Open Weights Land July 27 — 2.8T Parameters, 1.4TB of Memory
Moonshot AI's Kimi K3 topped a major coding leaderboard after its July 17 debut, beating Claude Fable 5 on coding tasks and ranking #3 overall on the Artificial Analysis Intelligence Index. The full weights are due on Hugging Face July 27. It is a 2.8-trillion-parameter mixture-of-experts model that fires only 16 of 896 experts per token (~50B active), so per-token compute resembles a mid-size model. The catch is memory: roughly 1.4TB even at four-bit MXFP4 precision, before any context loads. Moonshot has not yet published license terms, so commercial usability stays unconfirmed until the weights drop.
3. DeepSeek V4 Hits Stable Release, Undercutting Closed Models by ~30x
DeepSeek's V4 family reached its stable release on July 24, moving out of the preview it started in April. V4-Pro runs 1.6T total parameters with 49B active; V4-Flash is 284B total / 13B active. Both default to 1M-token context with 384K max output, under an MIT license with weights on Hugging Face and ModelScope. DeepSeek-V4-Pro-Max scores 80.6% on SWE-bench Verified — the top open-weights entry — and per output token runs roughly 28x cheaper than Claude Opus 4.8. A NIST evaluation put its overall capabilities about eight months behind the frontier.
4. EU Orders Google to Open Android and Search Data to AI Rivals
The European Commission issued binding Digital Markets Act decisions on July 16 requiring Google to give rival AI assistants access to 11 Android feature groups — voice activation and cross-app actions like booking a taxi or replying to messages — and to share anonymized search data with third-party engines and chatbots. Search data sharing starts January 2027; the Android changes must ship with Android 18 by August 1, 2027, with concurrent wake-word support by August 2028. No fines were attached, but non-compliance carries penalties of up to 10% of global annual turnover. Google's Android president Sameer Samat said the Commission is "on the wrong track" and flagged privacy and security concerns.
5. Gemini 3.5 Pro Slips a Third Time; Alphabet Shares Drop 4%
Google missed its July 17 target for Gemini 3.5 Pro, the third delay for the flagship model, with no model card or benchmarks published. The company is reportedly weighing a stopgap Gemini 3.6 Flash release to fill the gap. Alphabet shares fell about 4% on the news, in the same week as the EU's Android decision. Together the two hit both ends of Google's business — model development and distribution — as enterprise contracts come up for renewal.
6. Oracle Cuts 30,000 Jobs to Bankroll Stargate
Oracle is eliminating around 30,000 roles — roughly 18% of its workforce and the largest restructuring in company history — to free an estimated $8–10 billion a year for AI data center construction. The money backs a reported $300 billion five-year cloud contract with OpenAI covering some 4.5 gigawatts of additional Stargate capacity. AI data center, security and database engineering teams were explicitly protected; sales and other divisions absorbed the bulk of the cuts. The restructuring charge is expected to run about $1.6 billion this fiscal year.
7. China Launches WAICO with 29 Founding Members
The World AI Conference in Shanghai closed on July 20 with the formation of the World Artificial Intelligence Cooperation Organization, a new governance body with 29 founding members. Xi Jinping delivered the conference's first-ever keynote address. The event ran 140+ forums with more than 1,100 exhibitors, and Huawei used it to demonstrate its Atlas 950 SuperPoD domestic computing system. It lands as the US moves the opposite direction — a March federal framework that recommends against creating a new AI regulator and seeks to preempt state AI laws.
8. SAP Closes Prior Labs Deal, Betting Over €1B on Tabular Foundation Models
SAP completed its acquisition of Prior Labs, committing more than €1 billion over four years to foundation models built for structured business data rather than text. Prior Labs is the team behind TabPFN, published in Nature, which handles forecasting and anomaly detection on spreadsheets and database records without hand-built feature engineering. The deal marks tabular foundation models as a distinct enterprise category — one that has drawn far less attention than chatbots despite matching where most corporate data actually lives. It also gives Europe a frontier lab focused on something other than assistants.
9. AI Text Detectors Miss Up to 18% of Generated Writing
Epoch AI tested Pangram, GPTZero and Originality.ai against styled AI output and found that as much as 18% of generated text slipped through undetected. Scientific writing proved especially hard to flag, where the formal register masks the tells detectors rely on. The results undercut the reliability of tools now widely deployed for academic integrity checks and content screening. It arrives alongside a separate RadLE 2.0 benchmark showing radiology models delivering wrong diagnoses with full confidence — no uncertainty signal for a human reviewer to catch.
10. AI Funding Keeps Compounding: Meshy, Helsing, Candid Health
Capital kept flowing through mid-July. Meshy raised roughly $400 million in Series B at a $1.5 billion valuation for text-to-3D asset generation, and Candid Health took $120 million in Series D for automated medical billing. Defense AI firm Helsing announced $1.8 billion on July 13, with JPMorgan Chase, Lightspeed and Iconiq leading. China's PsiBot is raising close to $100 million at a $1.48 billion valuation, one of nearly 40 AI unicorns minted in 2026 so far. AI agent startups alone pulled in about $1.8 billion across 12+ deals this month, with enterprise automation taking 58% of it.
// KEY TAKEAWAYS
This was the densest open-weight week the industry has seen: DeepSeek V4 went stable on the 24th, Kimi K3's weights drop on the 27th, and both now sit close enough to the closed frontier that the gap is measured in months rather than generations — while costing 20–30x less per token. Anthropic answered with Opus 5 at half of Fable 5's price, which suggests the real competition has shifted from raw capability to cost per unit of work. Meanwhile the bill is coming due elsewhere: Oracle is cutting 30,000 jobs to pay for data centers, Google took hits from both a missed model target and an EU order to open Android, and regulators moved in opposite directions — Brussels tightening interoperability rules, Washington pushing to preempt state law, and Beijing standing up its own governance bloc.