AI DIGEST: 14–20 SEPTEMBER 2026

Here are our picks for the middle of September, the week was about measurement rather than capability. A controlled comparison finally asked whether the scaffolding around a model earns its keep, subject experts re-marked six benchmarks and found the fault in the benchmarks rather than the models, and two labs published figures for how much of their own work their models now do. Underneath all of it sits the cost nobody budgets for, which is reading what the machines wrote.

Two open-weight releases deserve a mention. DeepSeek released V4.1-Flash, a 552-billion-parameter MoE model that scores 74.2 on the DeepSWE v1.1 coding benchmark, just ahead of Opus 5 at 74.0. Bonsai 2 27B takes the other route, using ternary weights to reach an effective 1.76 bits per weight and a 5.9GB footprint while still accepting text and images across a 262,000-token context, and it runs under CUDA or Apple's MLX.

Writing Code Got Cheap, Reading It Did Not

A Sonar survey reported by IEEE Spectrum puts 42% of the code in shared codebases at machine authorship, while 96% of developers say they do not fully trust it, and the distance between those two numbers lands on whoever reviews the changes. A weekend experiment with GPT-6 Astra shows what that review costs: 35 hours of unsupervised agents and roughly four billion tokens produced 75,000 lines of code across 79 commits and nothing of lasting value, written in a style that is quick for the model and punishing for a human to check, such as rewriting whole C files by splicing strings in Python instead of editing them. Mistral's account of moving 40,000 lines of Fortran 77 to C++ points the other way: an automatic check that the new code returned the same numbers as the old, combined with splitting the job into separate coder, tester and reviewer roles, beat letting one agent run unsupervised. Spotify's engineering team published its own list of what broke and what it fixed once AI raised the rate of change across its systems. If you are buying agent throughput this year, budget the review capacity to absorb it, and put an automatic check in front of the people doing that review.

The Harness Claim Got a Controlled Test

The popular claim that the software wrapped around a model matters more than the model itself now has a measured answer. Arena.ai ran 21 combinations of seven models and three harnesses and found that swapping the harness barely changed how often a task was completed, while it changed the cost of completing it substantially, with the simplest harness often as good as the elaborate ones.

That sits awkwardly beside the demonstrations that keep circulating, including a talk showing one set of model weights going from 30% to 95% on the ARC-AGI benchmark on the strength of a better harness. A paper on training weaker models from stronger ones suggests how both can be true, because what matters is the fit between a particular model and a particular harness: copying a strong model's full working transcripts into a system that was tuned around a weaker model made the weaker model worse, and correcting only the steps it actually got wrong did not. Read harness work as a cost reduction project rather than a capability project, and measure it in spend per completed task.

Benchmark Scores Were Wrong in Both Directions

A re-grading of six physics benchmarks by subject experts found that many of the answers recorded as wrong were not the model's fault: the question was flawed, the stored solution was itself incorrect, or the grader rejected a correct answer. Once corrected, frontier models score considerably higher than published, and several of these benchmarks are close to saturation, meaning almost no headroom is left to measure anything with. Pulling the other way, Vals AI reports that models can exploit the same guardrails used to train them, which makes a self-reported score hard for an outsider to verify, while Goodfire found an internal signal inside a model that lights up when it is gaming its reward, which is a way to catch that behaviour at scale. IBM Research tackled a third problem, that an agent which passes a task once often fails the same task on a later attempt, by adding rules that make repeated runs agree with each other. Treat any single benchmark number, including one you produced yourself, as a measurement with error bars: run it more than once, check what the grader actually accepted, and keep a set of problems nobody trained on.

Speech Models Improved Faster Than Conversation Did

Voice and audio had a crowded week. Google introduced two live models on 15 September, saying the heavier of them takes first place on Artificial Analysis' speech-to-speech quality index at 82.6 and scores 97.7% on Big Bench Audio, while the cheaper one detects and switches between 97 languages mid-conversation and can run a tool call in the background while it carries on talking. Qwen3.8-Omni-Flash handles text, images, audio and video in a one-million-token context, Tencent published the weights for AuK, a 1.5-billion-parameter model that generates and edits speech from written instructions, and a technical report describes a single system covering speech, designed voices, vocals, sound effects and music. None of them addresses the part a caller actually notices: a new benchmark for turn-taking ran 14 systems across six conversation styles and found that none matches a person at judging when to speak, when to wait and when to interrupt, with error rates that swing sharply by the type of conversation. Voice agents are ready for scripted, one-request-at-a-time work and still awkward in open conversation, which is exactly where your customers will judge them.

A Quarter of the Research, and a Serving Stack in Two Weeks

Two labs put figures on how much of their own work their models are doing. Anthropic says Claude now accounts for about a quarter of its research work, a figure that covers supervising tens of thousands of internal agents and the computing behind them. Z.ai describes pointing a GLM-powered agent at the production serving stack for its own GLM-5.3-Flash model, built across more than 100,000 domestic accelerators in under two weeks, with kernel fixes and system-level tuning that tripled throughput, while people kept hold of the objectives and the risk decisions. A paper on recursive self-improvement sets out what would have to hold for this to keep compounding rather than flatten out. The transferable detail is the shape of the work rather than the headline.