Latest in AI
OpenAI’s Cheaper Sol Puts Autonomous AI’s Safety Economics to the Test
OpenAI’s GPT-6.1 Sol cuts flagship-model pricing by roughly 80%, promising wider autonomous AI adoption while raising questions about permission-seeking, oversight, safety evidence and the risks of deploying agents at scale.
Hirundo’s Westernized Qwen Raises a Harder Question Than Censorship
Hirundo says its Westernized Qwen model removes politically aligned behavior while preserving capabilities, but questions about benchmark design, neutrality, provenance and independent validation complicate the release’s significance.
Claude Haiku 4.5 Sent a False Murder Tip During Anthropic Testing
Anthropic says Claude Haiku 4.5 accidentally submitted a fabricated murder tip to Philadelphia police during testing, highlighting risks when AI agents interact with live government websites.
Anthropic’s Haiku 5.5 Targets the High-Volume AI Workload
Anthropic’s Claude Haiku 5.5 cuts API pricing to target high-volume enterprise AI workloads, but higher token use, context economics and reliability concerns mean businesses must measure complete tasks, not token rates alone.
OpenAI’s Fine-Tuning Exit Turns Custom Models Into a Managed Stack
OpenAI’s fine-tuning wind-down is pushing teams toward managed customization stacks built from prompts, retrieval, schemas, tools, evaluations, and selective provider-led training, and exposing new portability risks for AI products.
Blueprints AI wants to turn a description into a permit package
Blueprints AI says its new V2 system can turn descriptions, sketches and PDFs into coordinated construction documents, while licensed professionals remain responsible for review and permit submission.
Meta’s Muse Wonder tests whether memory can win the frontier AI race
Meta’s latest frontier AI release, Muse Wonder, bets persistent memory can make assistants more valuable than rivals. The models face scrutiny over proprietary benchmarks, API access and whether users can trust Meta’s long-term memory systems.
Mistral Large 4’s Open Weight Bet Faces Its First Real Test
Mistral Large 4’s promised open-weight release will test whether trillion-parameter scale, sparse inference and multimodal capability can deliver reliable performance and manageable deployment costs against Reflection’s Beam and closed models.
AlphaEvolve’s Real Breakthrough Is Teaching AI Not to Trust Itself
Google DeepMind’s AlphaEvolve uses Gemini models, automated evaluators and evolutionary search to discover algorithms, revealing why measurable objectives, independent validation and human release gates matter.