Three threads run through today’s issue.
Running AI got cheaper in several places at once, yet much of the proof that these systems work is still missing or comes from the people selling them. Each cut in price is real money for someone, and each missing result is a reason to wait before believing the claim.
The cost story is the loudest: a voice model that holds its $0.08 price while topping an independent ranking, server software that answers up to three times sooner, a giant model that uses a small slice of itself, and a storm simulator that Ai2 says costs 1/300th as much as the physics it learned from.
The unchecked work is the quieter one. Google’s paper reviewer shows no accuracy results, a test of model-picking software has not been run yet, and Netflix lost days of chip checks because the check stopped without telling anyone.
And as software agents take on real jobs, the question of who answers for what they did gets more practical. Anthropic is making their step-by-step records easier to read, and one writer says piling more agents onto a late project repeats a mistake from 1975.
The Price Drops: The same work for a fraction of the bill
Five stories about doing the same job for less, and one investor who says the savings will make people spend more.
- xAI’s new voice AI tops an independent ranking and keeps its $0.08 price. Grok Voice 3.0 scores 86.9 on Artificial Analysis’s speech ranking, ahead of OpenAI and Google. The ranking is independent, but the reply time under one second and the call-center results come from xAI itself.
- Open-source AI server software now answers up to 3x sooner under load. A change to vLLM starts copying a prompt’s working notes to the machine that answers before reading has finished. In tests on AWS at peak load, the wait for the first word fell from about 1,744 to 540 milliseconds.
- A giant new AI model runs only a small slice of itself at a time. inclusionAI’s Ling-3.1-flash has 560 billion parameters but uses 25 billion for each word it writes. Its Hugging Face page leaves out the licence, the scores and the download size, so the savings are a claim with nothing under it yet.
- Ai2 says its storm simulator costs 1/300th as much as the physics model. HiRO-ACEv2 generates global extreme weather data at 3 km detail far more cheaply than the simulation that taught it. It only works for the one climate it was trained on, which is a big limit for anyone studying a warmer world.
- Investor Gavin Baker says cheaper AI models will mean more computing, not less. Baker, of Atreides Management, argues thinner profits for model makers will lift demand for chips and cloud capacity. He offers no data in his post, and the post discloses no holdings.
Nobody Checked: Results that are promised, missing or silent
A reviewer with no scorecard, a test that has not run, and a safety check that stopped without a word.
- Google releases an AI paper reviewer, but its post shows no accuracy results. Deep Review critiques the logic and maths of long documents, and Google predicts tools like it will shape peer review. The announcement quotes authors who liked it but never compares it with human reviewers.
- Can software that picks your AI model save money? A test is coming. Vals AI says a perfect picker would beat the best single model by 14 points at about 40% less cost. That is the ceiling for a flawless picker, and the results for real products are not out yet.
- Netflix’s chip health check quietly stopped, and bad machines shipped for days. A software upgrade removed a file the check depended on, and nothing flagged that it had stopped. Netflix says a check that fails silently is worse than having none, and its post makes two other admissions worth reading.
Who Answers for the Agents: Records, rewrites and old lessons
Software agents are being given real work, and the people running them need to see what they did and decide who owns the result.
- Claude’s AI agents can now show their work in your existing monitoring tools. Anthropic is sending step-by-step records of what its agents did in an open industry format. A team can then watch Claude beside other AI providers on one dashboard instead of learning a separate tool.
- Google Research code rewrites an AI agent’s instructions and tools per task. The open-source framework tries edits to an agent’s prompts and tools on each job and keeps what worked. The underlying model is left alone, so the agent you audited yesterday may be set up differently today.
- Adding AI agents to a late project may repeat a 1975 mistake. Tamas Mezei argues that cheap code moves the bottleneck to deciding, checking and coordinating. A hundred agents will not help a team that already struggles to agree on what to build.
- The test for AGI is a tenfold jump in the hourly rent for AI chips. Kyle Corbitt says agents that match human workers would push chip rental from $3.52 an hour to about $30. He shows the sum behind that figure, which makes its weak points easy to find.
Today’s Quick Hits
- Anthropic sold $16.8 billion in three months, but growth is cooling. Investors heard the figures in private ahead of a planned November listing. Sales rose about 45 percent in the quarter, down from more than 140 percent in the one before.
- Amazon’s Bee now sells a $75 clip-on gadget that listens to your day. The $74.99 wearable turns everyday conversations into notes and to-dos. The people most affected may be the ones standing nearby who never agreed to be heard.
- TypeSafe’s Jev can now look at pictures, and the maker says it is worse at it. The decision-making model gains picture input, a way to point at things in an image and longer memory. All of it sits behind opt-in switches that the maker says are not ready for production use.
- OpenAI’s free teacher course now comes in Brazilian Portuguese. The AI for Educators course on OpenAI Academy has been translated. The course and the ChatGPT study feature it points to already existed, and the post gives no figures on how many teachers use them.