AI weekly digest: The race breaks into a sprint
GPT-5.5, Claude Mythos withheld for safety, Gemini 3.1 Ultra with 2M context, Gemma 4 open-source, Google's new TPUs, Arcee's 400B model, EU targets ChatGPT under DSA, 73K tech layoffs, NY RAISE Act, and Hinton's UN warning.
Three frontier models dropped in a single week, a 10-trillion-parameter model was deemed too dangerous to release, and Google fired a shot at Nvidia’s chip monopoly. Meanwhile, 73,000 tech workers lost their jobs in four months as companies redirect billions toward AI infrastructure.
1. OpenAI releases GPT-5.5 — weeks after GPT-5.4
OpenAI launched GPT-5.5 on April 23, calling it their “smartest and most intuitive” model yet. It scores 82.7% on Terminal-Bench 2.0 (vs. Claude Opus 4.7 at 69.4% and Gemini 3.1 Pro at 68.5%), with major gains in multi-step agentic tasks: planning, tool use, and self-checking. The model matches GPT-5.4 per-token latency but costs 2× more via API. Available to Plus, Pro, Business, and Enterprise users in ChatGPT and Codex.
Source: Introducing GPT-5.5 — OpenAI
2. Anthropic withholds Claude Mythos Preview, launches Project Glasswing
On April 7, Anthropic announced Claude Mythos Preview — and simultaneously said it would not release it publicly. During internal red-teaming, Mythos autonomously discovered and exploited zero-day vulnerabilities across every major OS and browser, including a 27-year-old OpenBSD TCP SACK RCE. Where Opus 4.6 succeeded twice out of hundreds of attempts on Firefox exploits, Mythos succeeded 181 times. Instead of a public launch, Anthropic created Project Glasswing: a $100M-credit coalition with AWS, Apple, Cisco, CrowdStrike, Google, JPMorgan Chase, Microsoft, NVIDIA, and Palo Alto Networks to deploy Mythos for defensive cybersecurity only.
Source: Claude Mythos Preview — red.anthropic.com
3. Google launches Gemini 3.1 Ultra with 2M-token context
Google released Gemini 3.1 Ultra, featuring a 2-million-token context window that processes 1,500+ pages of text or hours of video natively — no transcription step needed. On the Artificial Analysis Intelligence Index, Gemini 3.1 Ultra and GPT-5.4 Pro are tied at 57 points as the top publicly accessible models. The model shows strongest performance on ARC-AGI-3, GPQA Diamond, and long-context reasoning tasks, while GPT-5.4 still leads on SWE-Bench Pro coding.
Source: Gemini 3.1 Pro — Google Blog
4. Google drops Gemma 4 — open source’s best pound-for-pound models
Google released the Gemma 4 family on April 2 under Apache 2.0: four model sizes (E2B, E4B, 26B MoE, 31B Dense) purpose-built for edge and on-device deployment. The 31B Dense variant outperforms models 20× its size on standard benchmarks. All models support 256K context, native vision and audio, and 140+ languages. Since the first Gemma generation, the family has been downloaded over 400 million times, with 100,000+ community variants on Hugging Face.
Source: Gemma 4 — Google Blog
5. Google unveils TPU 8t and TPU 8i, challenging Nvidia
At Cloud Next on April 22, Google announced two custom AI chips: TPU 8t for training and TPU 8i for inference. The TPU 8t delivers 2.8× the performance of the Ironwood TPU at the same price, with 3× compute, 10× faster storage access, and 2× chip data transfer rates. Google isn’t replacing Nvidia — these supplement the Nvidia-based systems in Google Cloud — but the performance-per-dollar gap is narrowing. Both chips ship later this year.
Source: Google Cloud launches two new AI chips — TechCrunch
6. Arcee AI ships Trinity-Large-Thinking — 400B open-source reasoning model
The 24-person startup spent $20M (nearly half its total funding) on a 33-day training run using 2,048 NVIDIA B300 GPUs to produce Trinity-Large-Thinking: a 398B-parameter sparse MoE model with 256 experts, only 4 active per token. The result is 2–3× faster inference than dense models of similar size, released under Apache 2.0. Trinity-Large-Preview became the #1 most-used open model in the U.S. on OpenRouter, serving 80.6B tokens on peak days.
Source: Arcee releases Trinity-Large-Thinking — TechCrunch
7. EU prepares to classify ChatGPT under Digital Services Act
The European Commission is preparing to designate ChatGPT as a “very large online search engine” under the DSA, after its search feature crossed 120 million monthly EU users — well above the 45-million threshold. If confirmed, OpenAI would have four months to comply with obligations around advertising transparency, recommendation algorithm disclosure, content moderation reporting, and suspected criminal activity flagging. ChatGPT would join Amazon, Google, Meta, and X under the strictest tier of EU digital platform regulation.
Source: EU set to classify ChatGPT under strict online platform rules — Computing
8. Tech sector cuts 73,000 jobs in four months as AI spending surges
Oracle laid off up to 30,000 employees — 18% of its workforce — to redirect $8–10B toward AI data centers, despite reporting record revenue and backlog. Snap cut 1,000 roles (16% of staff) after AI reached 65% of its new code output and projected $500M+ in annualized cost savings. Across the industry, 95 companies cut more than 73,000 positions in the first four months of 2026. The pattern is consistent: not performance-driven reductions, but deliberate reallocation of headcount budgets to AI infrastructure.
Source: Oracle cutting thousands as company ramps AI spending — CNBC
9. New York finalizes the RAISE Act for frontier AI models
Governor Hochul signed the final version of the Responsible AI Safety and Education Act on March 27, substantially revising the December 2025 original. The law targets “frontier models” (trained on >1026 FLOPs) and requires large developers ($500M+ revenue) to publish a “frontier AI framework” detailing safety protocols and report incidents to the state within 72 hours. It takes effect January 1, 2027, making New York the first U.S. state with enforceable frontier model governance — a template likely to influence pending federal legislation.
Source: Governor Hochul Signs RAISE Act — NY Governor
10. Geoffrey Hinton calls for AI governance at UN, says models need “a steering wheel”
At the UN AI For Social Development Conference in April, Nobel laureate Geoffrey Hinton called AI “a very fast car with no steering wheel” and urged governments to act before capabilities outrun oversight. The conference underscored the need for transparent, accountable, rights-based AI governance — specifically addressing bias, opaque algorithms, and the concentration of training data in a handful of corporations. The global AI market is projected to grow from $189B (2023) to $4.8T by 2033, lending urgency to governance frameworks.
