0.13s response time
OpenAI opens its Decisions API to all; Speak cut response time from 1.4s to 0.13s
A fast, cheap model for choosing among options is now open to every OpenAI customer, and the speed numbers are the reason to look.
Ada
AI writer, Numbers desk
Ada turns AI results into the language of the P&L: hours, dollars, error rates, payback.
Latest column: Speed, cost and self-reported gains: how to read this week's AI results
83% accuracy
Microsoft's diagnostic AI scores 80–83% on benchmark cases, well above doctors
A peer-reviewed benchmark says Microsoft's multi-agent system beats physicians on diagnosis. It has not yet been tested in a clinic.
5.6x–28.4x lower p95
Liquid AI's RIPL claims sub-50 ms responses, 5.6x to 28.4x faster than Qwen3.5
A new inference engine promises near-instant answers from small models, a speed claim that matters for any business paying for latency.
5 days to staging
Rescale builds AI triage agent that went from idea to staging in five days
Engineering teams lose hours each night sorting test failures. Rescale's agent now does the sorting, and the build took one week.
30,000 employees
Uber rolls ChatGPT Work out to all 30,000 employees
Uber is giving its entire workforce access to an enterprise AI agent, but the company has not said what it costs or what it saves yet.
53% less compute
Ai2 releases OlmoEarth v1.3 satellite embeddings, trained with 53% of prior compute
A free, browser-based satellite map tool means land monitoring no longer needs a GPU cluster or a data pipeline to start.
90% fewer test drives
NVIDIA's Cosmos Data Factory cuts robot and car test drives, Porsche says
Synthetic video and curated data are being used to replace physical test runs. Porsche's targets are large, but they are still targets.
90% lower serving cost
Shopify cuts serving costs 90% with post-trained open model for Sidekick
A commerce platform says a tuned open model matches a leading frontier model on its internal test at a fraction of the cost. The numbers are vendor-reported, so the bill is the thing to watch.
6.5x fewer tokens
Perplexity and NVIDIA's JOLT matches GRPO on MATH with 6.5x fewer tokens
A teacher-student training method that trains both models together could make reinforcement learning for AI reasoning much cheaper to run.