ModelsMay 26, 2026·AI Weekly
Open-weights frontier model crosses MMLU 92
A new community release matches closed-frontier benchmarks while shipping under a permissive license, intensifying the open vs closed debate.
AI Industry News
A curated weekly digest of what's shipping and shifting.
PolicyMay 22, 2026·Policy Brief
EU finalizes AI transparency rules for foundation models
Providers must now disclose training data summaries and energy usage, with phased compliance starting Q4.
AgentsMay 18, 2026·Devtools Today
Coding agents hit 65% on SWE-bench Verified
Tool-using agents continue rapid gains on real-world software engineering tasks, narrowing the gap to junior engineers.
ModelsDecember 20, 2025·OpenAI
OpenAI ships GPT-5 with unified reasoning router
GPT-5 merges fast and deep reasoning behind a single endpoint, automatically allocating compute per query and setting new SOTA on AIME and GPQA.
ModelsOctober 14, 2025·Anthropic
Anthropic releases Claude 4 with day-long agent runs
Claude 4 Opus sustains multi-hour autonomous coding sessions and tops SWE-bench Verified at 72%, redefining expectations for agentic workloads.
HardwareJuly 9, 2025·NVIDIA
NVIDIA Blackwell Ultra ships, 1.5× training throughput
B200 Ultra and GB300 NVL72 racks begin volume shipments to hyperscalers, with FP4 training cutting frontier-model costs significantly.
Open SourceMarch 5, 2025·Meta AI
Meta releases Llama 4 with native multimodality
Llama 4 family launches with MoE variants up to 2T total parameters, native image and audio understanding, and a more permissive license.
Open SourceDecember 26, 2024·DeepSeek
DeepSeek V3 stuns industry at $5.6M training cost
A 671B-parameter MoE matches GPT-4o on many benchmarks while reportedly trained for under $6M, reshaping the cost narrative for frontier AI.
ResearchSeptember 12, 2024·OpenAI
OpenAI o1 introduces inference-time reasoning
o1 spends compute thinking before answering, dramatically improving math, science, and code reasoning and opening a new scaling axis beyond pretraining.
ModelsMay 13, 2024·OpenAI
GPT-4o launches with native multimodal voice
OpenAI's omni model handles text, vision, and realtime audio in a single network, with sub-300ms voice latency rivaling human conversation.
ModelsFebruary 15, 2024·Google DeepMind
Google unveils Gemini 1.5 Pro with 1M-token context
Gemini 1.5 Pro debuts a sparse MoE architecture and a 1M-token context window — later expanded to 2M — enabling whole-codebase and feature-film reasoning.
IndustryNovember 6, 2023·OpenAI
OpenAI DevDay introduces GPTs and the Assistants API
Custom GPTs, the Assistants API, and a 128K-context GPT-4 Turbo launch — marking the start of the mainstream agent platform era.
Open SourceJuly 18, 2023·Meta AI
Meta releases Llama 2 with commercial license
Llama 2 becomes the first frontier-class open-weights model usable commercially, igniting the open-source LLM ecosystem.
ModelsMarch 14, 2023·OpenAI
OpenAI launches GPT-4
GPT-4 debuts with multimodal input, sharply improved reasoning, and professional-exam-level performance, defining the modern LLM benchmark.
IndustryNovember 30, 2022·OpenAI
ChatGPT launches and reaches 100M users in two months
OpenAI's free chat interface to GPT-3.5 becomes the fastest-growing consumer app in history and kicks off the generative AI boom.
ResearchJune 11, 2020·arXiv
GPT-3 paper demonstrates few-shot learning at scale
A 175B-parameter model performs new tasks from prompts alone, validating scaling laws and triggering the LLM era.
ResearchJune 12, 2017·arXiv
'Attention Is All You Need' introduces the Transformer
Vaswani et al. publish the architecture that underpins every modern LLM, replacing recurrence with self-attention.