The week left one data point worth more than ten announcements: changing the scaffolding of a coding agent moves the success rate by a few points, within the noise of a rerun, and can triple the cost per task. Around that number revolve the other news items, from the flat margins of the S&P 493 to Meta's request for 35 billion from the market, up to two model releases in two weeks. The thread is one: sophistication has a price, and it must be justified.
Image generated entirely with AI.Does the harness really matter?
A controlled study compared the harnesses of Claude Code, OpenCode, and mini-SWE-agent, measuring two separate things: tasks solved and cost per task. The first remains practically flat, the second can triple. The reason is mechanical: the more elaborate harness drags along a long system prompt, with instructions, tool definitions, and context, resent to the model at every single step. On a twenty-step task, that is twenty times the same input tokens, paid at full price.
The operational point is the prompt cache. If the prefix remains identical between one turn and the next and the provider bills it at a reduced price, the distance between a fat harness and a lean one narrows. If instead each step injects something dynamic at the top of the context, the cache breaks and you pay for everything. The triple cost comes almost entirely from there.
Before changing framework, it is worth looking at three things:
- how stable the system prompt prefix is;
- how many tools are declared, because every definition weighs on every call;
- whether the routing phases can run on a small model instead of the strong model.
The scaffolding that looks most sophisticated often adds ceremony, not capability.
The reasonable expectation is that the harness counts when the model is weak and must be held by the hand. With current models, most of the work is done by the model, and the scaffolding is often theater. It makes sense, then, to measure the ratio between tasks solved and dollars spent, as one does when tracking the real cost of an agent. A harness that doubles spending for two points of pass rate has already lost.
Where does the AI payoff end?
Torsten Slok of Apollo looks at the S&P 493, that is, the index without the seven big names, and finds flat margins and revenue per employee. No jump attributable to AI outside the tech sector. His explanation: the bottleneck is in the craft, not in the model. Turning low-cost intelligence into margins requires accountability, trust, unwritten rules, judgment.
Gartner adds the bill side. Of 272 executives interviewed, 58% consider it a priority to bring AI projects into production at large scale within twelve months. But those who use variable pricing without consumption limits will exceed initial cost estimates by more than 50% by 2027. Hence the "no telemetry, no deployment" rule and the role of FinOps. An agent makes dozens of calls for a single goal, and the bill grows faster than the work it removes.
Then there are Meta's numbers, which are the other side of the same coin. Nearly 400 billion in market capitalization lost since the end of September, a 35 billion share sale, free cash flow at negative 2.12 billion against positive 10.6 a year earlier, long-term debt risen to 80.9 billion. Revenue grows by 26% and spending guidance talks about 135-145 billion: control, at that scale, has already broken down. The line to read twice is another one: Nvidia, AMD, and Broadcom have signaled non-binding interest for up to 5 billion of those new shares. The suppliers financing the buyer of their own chips.
Meanwhile the price of a token keeps falling, with cuts of up to 90% for large customers and competitive models at one fifth the cost of leading models. When the price collapses, the competitive advantage shifts from the model to the process around it.
What to watch: whether in the next four quarters the S&P 493's revenue per employee moves while AI costs grow, and whether one of the three chipmakers turns interest into a binding purchase. Meanwhile the conclusion does not change: the tool is installed in days, the process takes months. It is the terrain of training and support more than of model choice.
Image generated entirely with AI.
Is the model still the advantage?
Google DeepMind published Gemini 4 Carbon exactly two weeks after Gemini 4 Argon. Sixty-two... no, 59 points on Artificial Analysis's Intelligence Index v5, first place, gap widened over open models. The two characteristics that matter for those building products: optimization for long agentic workflows, which remain open for many iterations without losing context and state, and coverage of the entire cyber defense cycle, with progressive access precisely for security capabilities.
The cadence is the real news. Two weeks between one generation and the next mean monthly regression tests for those with a rigid architecture on a single model, and a configuration change for those with an abstraction layer. The question to hold onto is already written: what remains of a product when the model underneath changes at this frequency.
On the other side, Mistral has released for download the weights of Mistral Large 4, a trillion-parameter multimodal model trained in Europe, open license on the weights, positioned as the best open-weight model for coding. The size moves the bar for self-hosting: that tier was almost exclusively the preserve of closed models via API. For those working with data that cannot leave the infrastructure, the question becomes whether to pay per token or load the weights onto their own hardware.
The limit is the size itself. A trillion parameters requires hardware and serving engineering, and nobody gives that part away. For a nightly cron that reads feeds and summarizes, it is disproportionate. The practical value, if it arrives, is measured on two things: real inference cost at third-party providers and code quality in independent tests. And on a third, which would be the real news: a small variant under the same license.
Image generated entirely with AI.
SundAI comes out once a week, but AI news doesn't wait for Sunday. Every morning I pick the best one and comment on it.
TodayNovember 1, 2026
Google burns 51 billion on AI and the market wipes out 400 billion
What remains to govern?
The smallest news of the week is perhaps the most useful: OpenAI, Google, Microsoft, and Anthropic have agreed on a single signal in robots.txt to say whether a site can be used for training. One rule written in one place, instead of chasing the user agent that changes every month. The limit is that it remains voluntary: no regulatory obligation, no public log, no guarantee about who participates today and changes their mind tomorrow. A voluntary opt-out rewards those with a reputational interest in respecting it.
On the opposite front, access governance is shifting from users to agents. IBM has announced verified identities and expiring credentials for AI agents, while Microsoft warns that agents are now a new class of insider risk. An agent with the wrong credentials on critical infrastructure multiplies an incident in seconds. The case reported by Wired, a Claude agent that ran for two months on a Pentagon network, says the question is no longer theoretical. The indicator to follow in 2027 is simple: how many companies will have a spending cap for agents and a readable log of their actions. If there are few, the rest is slides.
On tools, the week left more concrete things than announcements:
- Kimi K3.1 Pro: a one-million-token window and competitive coding results at one fifth the cost of leading models. The weights arrive on November 4, so for now it is judged by the endpoint.
- Dessa (Block): a voice model that processes audio without going through text and runs locally on a MacBook M4 Max. Latency and privacy change together.
- Flora (Nvidia): an open-source data engine that labels an hour of robot video for 72 cents. It is the kind of infrastructure that makes a robotics dataset sustainable.
- Claude Code hooks: they stop untested code before commit and protect API keys. Small, but it is control over the perimeter, not the demo.
- OpenMobile-2: an open environment with over 110 Android apps for training agents that operate smartphones. Here the bottleneck shifts from benchmarks to interface variety.
The criterion for screening all this remains that of the first news item: how many tasks solved per dollar spent, how stable the prefix is, how much an extra step costs. The rest is background.
