Everything that mattered in AI, one page a week.

Roughly twenty minutes after Anthropic's post, OpenAI released GPT-5.3-Codex, merging GPT-5.2-Codex's coding with GPT-5.2's reasoning. It runs 25% faster and uses less than half the tokens of its predecessor.

OpenAI says early versions debugged their own training runs, managed deployment and diagnosed evaluations — its first model instrumental in creating itself. Reported scores: Terminal-Bench 2.0 at 77.3% versus 64.0%, and OSWorld-Verified at 64.7% versus 38.2%. It is also the first OpenAI model rated High capability for cybersecurity under its Preparedness Framework, with API access held back pending safety review.

Also on February 5, OpenAI announced Frontier: a platform for building, deploying and managing agents across enterprise systems, framed around giving agents the same things people need at work — shared business context, onboarding, feedback-driven evaluation, and scoped identity and permissions.

It is positioned as vendor-neutral, supporting third-party agents, and runs locally, in a customer cloud or hosted, with SOC 2 Type II and ISO 27001/27017/27018/27701 compliance. Early adopters include HP, Intuit, Oracle, State Farm, Thermo Fisher and Uber.

HashiCorp co-founder Mitchell Hashimoto published “My AI Adoption Journey,” a six-step account of adopting coding agents. Step five is the one that spread: every time an agent makes a mistake, engineer a permanent fix into its environment — a documentation file or a programmatic tool — rather than re-prompting, so the harness gets stronger as the agent works.

He pairs it with practical habits and a caveat that skill formation in junior engineers deeply worries him. The framing — agent equals model plus harness — was picked up within weeks by Thoughtworks' Birgitta Böckeler and echoed in OpenAI's own harness-engineering write-up.

Mistral released two speech-to-text models. Voxtral Realtime is 4B parameters under Apache 2.0 with weights on Hugging Face, using a purpose-built streaming architecture rather than chunking an offline model, with configurable latency down to sub-200ms. At 480ms delay it stays within 1–2% WER of the batch model.

Voxtral Mini Transcribe V2 is the closed batch model at $0.003/min with speaker diarization, word-level timestamps, context biasing and 13 languages, claiming roughly 3x the speed of ElevenLabs Scribe v2 at about a fifth of the price.

Alphabet's Waymo closed a $16 billion round at $126 billion post-money — more than double its October 2024 valuation — led by Dragoneer, DST Global and Sequoia, with Alphabet remaining majority investor.

Waymo says it does more than 400,000 rides a week across six US metros, tripled annual volume to 15 million rides in 2025, and points to 127 million fully autonomous miles with a 90% reduction in serious-injury crashes. The capital funds groundwork for over 20 additional cities in 2026, including Tokyo and London.

Moonshot AI released Kimi K2.5, built from Kimi K2 by continued pretraining over roughly 15T mixed visual and text tokens. It is natively multimodal and targets coding, vision and deep-research workloads.

The new idea is Agent Swarm: for complex tasks the model self-directs up to 100 sub-agents across up to ~1,500 parallel tool calls, which Moonshot says cuts execution time up to 4.5x versus a single agent. Independent evaluators at Vals ranked it the new #1 open-weight model, first among open models on 13 of 17 benchmarks.

Arcee AI released Trinity Large, a 400B-total / 13B-active sparse MoE trained on 17T tokens in roughly 30 days on 2,048 NVIDIA B300 GPUs, under Apache 2.0 with GGUF quantizations and a full technical report.

The routing is extreme — 256 experts with 4 active per token, about a 1.56% routing fraction — plus sliding-window attention at a 3:1 local-to-global ratio. Arcee claims 2–3x faster inference than peers in its weight class, and shipped raw checkpoints including a 10T-token base with no instruction data.

The Washington Post published the first detailed account of more than 4,000 pages of court filings unsealed in Bartz v. Anthropic, the class action brought by novelist Andrea Bartz and the nonfiction writers Charles Graeber and Kirk Wallace Johnson. Among them was an internal planning document dated 13 April 2024: “Project Panama is our effort to destructively scan all the books in the world.” A second line in the same document reads: “We don't want it to be known that we are working on this.”

The mechanics were industrial. Anthropic began buying in early 2024 and hired Tom Turvey, who had spent years building the partnerships behind Google Books, to source stock — bulk lots from second-hand dealers, often in batches of tens of thousands, including Better World Books, World of Books and Wonder Book. A vendor proposal in the filings sought capacity to convert between 500,000 and two million books over six months. A hydraulic-powered cutter took the spine off each volume, high-speed production scanners captured the pages, and recycling companies collected the remaining paper. The filings put the spend at tens of millions of dollars within about a year.

Destroying the books is what made it lawful. Judge William Alsup held in June 2025 that scanning a book Anthropic had bought was fair use specifically because the print original was destroyed: “one replaced the other.” Keeping the paper alongside the scan would have meant two copies. The same order found that Anthropic's earlier download of more than seven million pirated books — from Books3, LibGen and the Pirate Library Mirror — was not fair use regardless of what the books were later used for, and that is the part the $1.5 billion settlement covered.

Microsoft announced Maia 200, its second-generation inference accelerator: TSMC 3nm, over 100 billion transistors, 216GB of HBM3e at 7 TB/s, 272MB of on-chip SRAM, native FP8/FP4, and an on-die NIC over standard Ethernet supporting clusters up to 6,144 accelerators.

Microsoft claims 30% better performance per dollar than the latest hardware in its fleet, 3x the FP4 performance of AWS's third-gen Trainium, and FP8 above Google's seventh-gen TPU. Deployment began in Azure US Central with a Maia SDK preview open to developers.

At the American Meteorological Society meeting Nvidia launched Earth-2: open models, libraries and frameworks across the whole weather pipeline. Earth-2 Medium Range handles 15-day forecasts across 70+ variables on a new Atlas architecture; Nowcasting produces kilometre-resolution 0–6 hour storm forecasts; Global Data Assimilation builds initial atmospheric states in seconds on GPUs instead of hours on supercomputers.

Weights and code are on Hugging Face and GitHub via Earth2Studio. Nvidia says trained models run roughly 1,000x faster than conventional methods, and Israel's meteorological service reported a ~90% cut in compute time.

The European Commission opened a new formal Digital Services Act investigation into X and extended its existing proceedings on X's recommender systems. The new probe assesses whether X properly evaluated and mitigated systemic risks from deploying Grok's functionality in the EU, and whether a mandated ad-hoc risk assessment happened before deployment.

The Commission said the risks appear to have materialized, and extended the probe to cover X's switch to a Grok-based recommender. Potential breaches cited include Articles 34, 35 and 42.