What the Latest AI Industry Updates Reveal: GPT‑4.5, GLM‑5.1, Optimus, Nvidia B200 and More

1. OpenAI GPT‑4.5 Official Release

OpenAI announced GPT‑4.5 with a context window increased from 2 million to 5 million tokens (+150%). Multimodal capability now includes image + video up to 10 minutes, whereas GPT‑4o only supported image + text. Benchmarks show a 15% rise in reasoning (100 → 115) and a 1.9% improvement in HumanEval coding (90.2% → 92.1%). Pricing remains unchanged at $5 per million input tokens and $15 per million output tokens. Developer sentiment on X/Twitter and Reddit is mixed: some praise the ability to process an entire novel like *The Three‑Body Problem* in one go, while others note a 40% slowdown in inference speed and modest code‑ability gains compared with DeepSeek V4‑V4.5. A neutral view sees GPT‑4.5 as a transitional step toward GPT‑5.

Market data from OpenRouter indicates a 20% rise in API calls on day 1, far below the 150% surge seen after GPT‑4o, and many enterprise customers remain cautious.

2. Zhipu GLM‑5.1 Open‑Source Ecosystem Weekly Data

Within a week, the GLM‑5.1 ecosystem recorded over 500 fine‑tuned models, 35 k+ GitHub stars, 800 k+ Hugging Face downloads, and 15 k+ active developers. The top five models by download count are:

GLM‑5.1‑Law (80 k downloads) – developed by Peking University Law Lab

GLM‑5.1‑Med (65 k) – by Union Hospital

GLM‑5.1‑Code (52 k) – community‑driven

GLM‑5.1‑Fin (48 k) – by a leading securities firm

GLM‑5.1‑Edu (35 k) – by Beijing Normal University

Zhipu’s official support includes a free 1 million‑token fine‑tuning cloud, a 50 million‑yuan GLM ecosystem fund, and a 10 million‑yuan application contest prize pool.

3. Tesla Optimus First Batch Delivered to BMW Factory

Tesla released a video showing Optimus performing tasks at a BMW plant. Measured success rates and cycle times are:

5 kg part handling – 98% success, 45 s vs. 30 s for a human

Screw tightening – 95% success, 2 min vs. 1.5 min for a human

Visual quality inspection – 92% success, 30 s vs. 45 s for a human

Equipment巡检 – 88% success, 10 min vs. 15 min for a human

BMW’s official statement confirms the robot is “acceptable for repetitive, hazardous tasks” and plans to purchase another 100 units for pilot testing.

Production data from Musk’s X account shows 1 000 units delivered initially, a Q2 2026 target of 5 000 units, and a unit price of $20 k.

4. Nvidia B200 Production Delayed to Q3 2026

Nvidia CFO Colette Kress confirmed the B200 accelerator will ship in Q3 2026 due to severe HBM3E supply shortages (high impact), limited CoWoS packaging capacity (medium‑high impact), and broader supply‑chain adjustments (moderate impact). Mitigation measures include priority HBM3E agreements with SK Hynix and Samsung, investment in TSMC CoWoS capacity, and a B100 “emergency” version delivering 70% of B200 performance.

Market reaction: Nvidia shares fell 3% after hours, and some data‑center customers are shifting to Huawei’s Ascend 910D.

5. Google DeepMind AlphaEvolve 2.0 Release

DeepMind announced AlphaEvolve 2.0, an AI system that autonomously designs chips. Compared with human‑engineered TPU v5, AlphaEvolve‑designed TPU v6 achieves:

Energy‑efficiency: 135 vs. 100 (+35%)

Die‑area utilization: 88% vs. 75% (+13%)

Design cycle: 4 months vs. 18 months (‑78%)

Design cost: $50 M vs. $300 M (‑83%)

Industry response (IEEE Spectrum) notes chip‑design engineers fear job loss, while Google stresses a “human + AI collaboration” model without layoffs. Competitors such as MediaTek and Qualcomm are evaluating similar tools.

6. Microsoft Windows 11 AI Features Updated (24H2)

Microsoft rolled out the Windows 11 24H2 update, extending Copilot+ PC capabilities (local AI search of PC history, real‑time subtitles in 44 languages, built‑in Paint Co‑creator, automatic video upscaling). The update is not Windows 12.

Official Q&A clarifies that Microsoft never announced a “Windows 12 in 2026” roadmap; such rumors are media fabrications.

7. ByteDance “Jimo” Video Generation Hits 10 M Daily Active Users

Operational data shows Jimo reaches 10 M DAU, 5 M videos generated per day, and an average session length of 25 minutes—25% higher than competitor Kuaishou’s “KeLing AI”. Growth drivers include an in‑app one‑click publish button, >100 k video templates, and a referral‑for‑free‑quota mechanism.

Kuaishou counters with a creator‑revenue‑share program for high‑quality videos.

8. SenseTime “DailyNew” 7.0 Medical Version Clinical Validation

Clinical trials across three‑grade hospitals report AI accuracy versus human experts:

Lung nodule malignancy – 96.8% vs. 93.5% (3.3% gain)

Diabetic retinopathy – 95.2% vs. 91.8% (3.4% gain)

Breast cancer lymph‑node metastasis – 94.5% vs. 92.1% (2.4% gain)

Coronary artery stenosis – 93.7% vs. 90.5% (3.2% gain)

The devices have submitted Class‑III medical‑device applications to the NMPA, with approval expected in June.

9. China AI Overseas Expansion to Middle East and Southeast Asia

Major Chinese AI firms announced market‑specific launches:

Alibaba partners with Saudi PIF and G42 to launch an Arabic‑language “Tongyi Qianwen”.

Tencent collaborates with GoTo and VNG for an Indonesian/Vietnamese “HunYun” version.

Baidu works with local governments on a multilingual “Wenxin Yiyan”.

ByteDance releases a multilingual “Jimo”.

Huawei teams with regional operators to provide Ascend compute centers.

IDC forecasts the Middle‑East AI market to reach $8 B and Southeast Asia $12 B by 2026.

10. China Ministry of Industry and Information Technology AI Compute‑Voucher Policy

The ministry issued the “AI Compute‑Voucher Management Measures (Trial)”. Subsidy caps and ratios are:

Micro‑enterprise (< ¥5 M revenue): ¥100 k max, 50% subsidy.

Small enterprise (¥5‑50 M): ¥300 k max, 40% subsidy.

Medium enterprise (¥50 M‑400 M): ¥1 M max, 30% subsidy.

Specialized‑new‑technology firms: ¥2 M max, 50% subsidy.

The goal is to subsidize 100 k firms and stimulate ¥50 B of AI compute consumption in 2026.

11. Estonia AI Judge Pilot – First Digital Court

The pilot processed 120 small‑claims cases (< €10 k) on day 1, with an average handling time of 2 hours and 85% user satisfaction. Human judges performed a mandatory 100% review, and the system is limited to low‑value disputes.

12. United Nations AI Governance Fund – Mid‑Term Evaluation

Projects in Kenya (AI agricultural assistant, 80% progress, 90% pest‑identification accuracy), Nigeria (AI medical diagnosis, 70% progress, 95% malaria‑diagnosis accuracy), Vietnam (AI education platform, 85% progress, 20% student‑math‑score improvement), and Indonesia (AI disaster warning, 75% progress, tsunami alerts 30 min early) show measurable impact.

China contributed 50 technical experts, trained 500 local engineers, and donated $20 M of compute equipment.