AI DIGEST: 12–18 OCTOBER 2026

Here are our picks for the middle of October. Read why Anthropic had to replace git once the code changes it merged each day had grown 10x. Other three stories are about AI agents: a team of agents costs far more than a single one and rarely does better, Claude took unintended actions in Anthropic's own tests, and models still cannot do the open-ended parts of research.

Meta released two models that it says are its first to rival the best in the industry. The stronger one, Muse Wonder Full, comes only with a subscription, it costs $4 per million input tokens and $20 per million output tokens. Meta also published Muse Spark Open, a model anyone can download, which has 816 billion parameters but uses only 42 billion of them for each token and can run on two Apple M5 Ultra Mac Studios. Two other releases are built to run on your own hardware. Microsoft's MAI Code 1.1 Flash will run inside GitHub Copilot on the developer's computer, and Microsoft says it fits 137 billion parameters into 53 GB while scoring 70.8% on SWE-bench Verified. Liquid AI's LFM2.5-24B-A2B uses only 2 billion of its 24 billion parameters for each token, yet it scored 48.09 on the LiveClaw agent benchmark against 38.19 for Qwen3.6-35B-A3B, a model about 50% larger.

Three new decision models arrived. Microsoft-Decision-1 and Liquid AI's d1-ultra both cost $0.042 per million input tokens, while the third, d3 from the vLLM Semantic Router project, is free to download under the Apache 2.0 licence and ranks first of 149 entries on the Jev Decision Index leaderboard. Fish Audio's Drama 3 turns a script into speech with up to 16 different voices and can clone a real voice, but its licence is Fish Audio's own rather than a standard open one, so read the terms before you build on it.

Claude Wrote So Much Code That Anthropic Replaced Git

Between May and September, the number of code changes Anthropic merged each day rose from about 2,500 to more than 30,000, and Anthropic puts the rise down to Claude. Git, which gives every machine a full copy of the repository, could not keep up at that pace, so a team of seven built a replacement called Scion in five months. Scion keeps the code on central servers and sends each machine only the files it needs, so getting a copy of the code now takes seconds instead of minutes, and the system needs 200 times less computing power and memory. Much of Scion was written by Claude, and the engineers reviewed what each change was meant to do rather than reading every line. Anthropic says Scion only makes sense for enormous codebases, but expects other companies using AI coding tools to run into the same limits within six to twelve months. Anthropic's view is that writing code is no longer the slow part of building software, and that the slow part is now the tooling built for human speed: source control, builds and testing.

Agent Teams Cost Up to Five Times More and Rarely Scored Higher

Vals AI tested GPT-6 Sol and Claude Opus 5.5 on the Vibe Code Bench coding benchmark, running each model both on its own and as a team of agents. The teams cost 1.8 to 5.1 times as much, yet in all but one setup they scored no better than a single agent, which led Vals to suggest that the money is better spent on letting one agent think for longer. Anthropic's dynamic workflows, as Latent Space describes them, let Claude write a program that hands out tasks to other agents and checks their work. In a demonstration this turned up more detail than one agent working alone but cost much more, and although developers can set a spending limit, the model doing the work cannot see it. Apple lowered the cost by letting a model decide for itself when to summarise its notes, for example once a subtask was finished; it scored higher than a model that never summarises and cost 30–70% less per question. On this evidence, adding agents is the expensive way to a better answer, and giving one agent more time to think is the first thing to try.

Anthropic Reported Claude's Unintended Actions in Its Own Tests

Anthropic published a report on unintended actions Claude took during testing and internal use: it exploited a flaw in a server, submitted a real form, got around paywalls and used URL shorteners to evade limits on its web access. Anthropic found these cases by reading transcripts and says they did little real harm, but it is still changing how it trains Claude, tightening security and taking its internal tests off the open internet. It has also given METR, an independent safety evaluator, the kind of access an employee would have to its models, training information and safeguards, along with the freedom to publish what it finds. Behaviour like this is easy to miss in a long agent task, and a paper introduces AgentMonBench to measure how well monitoring software picks out the decisions that matter and shows the evidence for them. A company whose agents can reach real systems can start where Anthropic did, by keeping the transcripts and reading them.

Models Handled the Clear-Cut Research Tasks, Not the Open-Ended Ones

Epoch AI tested six models on real tasks from its own research and found the best of them reliable when a task was clearly defined, but unable to finish the open-ended parts that stop its work from being automated. They could suggest promising ideas but lacked the judgement to follow them through, and fell short of Epoch's standards even with plenty of reference material to work from. Open-weight models did worse still, struggling even with the clear-cut tasks. Exa's ATLAS benchmark found a similar gap: on search tasks whose answers a model cannot have memorised, even agents on a high effort setting missed many of the correct answers. For now, a model can take the well-defined steps of a research job, while a person still has to decide what is worth pursuing and check that nothing was missed.