analysis
Mercury 2.5 gets an independent quality and speed evaluation
Artificial Analysis's new Mercury 2.5 evaluation separates fast text output from answer delay and quality. Its cost figures also use rates that differ from Inception's promotion.
The BIG CHANGE archive
News, analysis and practical guides.
Explore our reporting on how AI and robotics change work, business and the world.
Explore the stories
42 stories
analysis
OpenAI agent reached non-public Medicare statistics files. What did it access?
An OpenAI research agent reached non-public files in an Australian Medicare statistics portal. Officials report no evidence of patient-record access; the method and full scope remain under investigation.
analysis
Anthropic says its Claude apps got 3x faster. Here is what changed
Anthropic reports a threefold average speedup across 13 Claude app measurements after an August sprint. The engineering account shows the user gains, measurement loop and human release controls behind the claim.
analysis
Claude found a CRISPR-like pattern. Its biological function is still unknown
Claude flagged an unusual repeat array beside a phage enzyme. Anthropic measured short RNAs, but has yet to establish the enzyme's activity or any editing capability.
analysis
Gemini 3.8 TTS is cheaper now. Its price doubles January 1
Google's stable Gemini 3.8 TTS models lower generated-audio costs, but prices double January 1 and migration changes prompts and WAV handling.
analysis
Did OpenAI halve GPT-6 Sol's Codex quota? The math does not prove it
A Codex subscriber reports lower API-equivalent value after GPT-6 Sol's price cut. That figure alone cannot establish fewer tokens or less work per subscription.
analysis
Stripe put domain owners at the center of its internal AI
Stripe separates agent maintenance from infrastructure. Its design raises a practical question: who owns instructions, permissions and evidence that the work is correct?
analysis
Basecamp raises $140M to take EDEN's lab results toward therapies
Basecamp Research's $140 million round backs its biological models and therapy pipeline. We examine the dataset, selected experiments and current development stages.
analysis
Qwen Audio 3.1 widens its voice stack. The 95% price cut needs context
Qwen Audio 3.1 adds sound creation and understanding. Its API documentation is uneven, and changing billing units complicate the advertised ASR savings.
analysis
OpenAI is marketing ChatGPT through creators. Here’s what the labels mean
OpenAI influencer marketing is bringing ChatGPT into everyday creator posts. This disclosure map shows what the labels establish and what they do not prove.
analysis
Amodei's AI slowdown ties safety to preserving a U.S. lead
China's official records recognize frontier-model risks. Its diplomats reject competitive framing, a distinction that matters when assessing calls to slow AI.
analysis
AT&T's workforce is shrinking. The AI explanation is incomplete
AT&T's shrinking workforce sits alongside copper retirement, network automation and broader cost reductions. The reviewed records do not separate their employment effects.
analysis
Greece wants AI in schools. The harder question is what role it plays
Greece's teacher-first AI pilot separates questions about workload from pupil learning. The public plan makes expansion conditional on success.
analysis
Senators seek military AI inquiry after reported errors
A Senate request raises questions about military intelligence. The White House's June policy sets current testing requirements; its implementation is a separate oversight question.
analysis
Britain's information defence centre needs a clear chain of command
Britain proposes a centre for hostile-state information attacks, though NSOIT already covers similar risks and Parliament has questioned ownership and scrutiny.
analysis
An AI agent can skip a required skill and leave no score behind
New skill evaluators separate loading, selection and instruction following. Their missing-result behavior shows why teams must treat evaluation coverage as evidence of its own.
analysis
Rabbit OS3 works without an r1, but remains a technical preview
Rabbit OS3 works in a browser without an r1. Model billing, cloud processing and technical-preview terms shape what connecting your computer means.
analysis
Nscale’s $103.4 billion AI contract figure includes capacity still to be built
Nscale discloses a wide gap between active and contracted compute, with financing and delivery separating signed demand from service.