Four of today’s five stories are about money landing somewhere other than where you assumed it went. I run this site’s entire content operation on a coding agent, so the first one cost me an afternoon of recalculating.
The Harness, Not the Model, Is Setting Your Coding Agent Bill
Arena’s research team published a study on September 16 that does something almost nobody bothers to do: it holds the model constant and moves the software wrapper around it. Every writeup I found treated the result as a leaderboard question, which harness wins. That framing buries the actual finding, and the actual finding is the only number in the study that touches your invoice. On the two benchmarks tested, swapping harnesses barely changes whether the work gets done. It changes what the work costs by up to five times.
Arena’s own writeup covers 21 model and harness pairs, seven models run across Claude Code, Codex CLI and the open-source Pi, measured on SWE-bench Lite and Terminal-Bench 2.0. Thirty randomly sampled tasks per benchmark, three attempts each, capped at 100 agent turns, priced against a fixed direct-API list dated September 1.
Start with the pair that made me sit up. Claude Fable 5 solved 97.8% of attempts inside Claude Code, 96.7% inside Codex, and 96.7% inside Pi. Claude Code cost $1.33 per attempt. Pi cost $0.67. Same model, 1.1 percentage points of extra success, double the money. Across the models shared by all three harnesses, Arena measured Claude Code at roughly 2.0 times the cost of Pi and 1.6 times Codex on SWE-bench Lite, while the average effect of harness choice on success rate stayed inside two percentage points.
The mechanism is not mysterious, which is what makes it fixable. Claude Code’s mean initial context across all seven models runs more than ten times Pi’s, because of longer instructions and fatter tool schemas. You pay that on the first call of every task, before the agent has done anything. Pi ships four tools: read, write, edit, bash. On Fable 5 against SWE-bench Lite the two harnesses averaged 15.4 and 15.3 turns per attempt, near identical work, at roughly double the spend.
Then the part that should embarrass every vendor: models frequently perform better outside their own provider’s harness. Across six Anthropic and OpenAI models on both benchmarks, Arena found an alternative harness took the highest observed success rate in nine of twelve comparisons. Sonnet 4.6 hit 68.9% in Codex against 66.7% in Claude Code at comparable cost. GPT-5.6 Sol reached 83.3% in Pi on Terminal-Bench 2.0 against 78.9% in Codex, at $0.42 versus $0.76.
Harness choice is a procurement decision wearing a capability decision’s clothes. We have all been arguing about models while the wrapper quietly set the price.
Verdict: this matters, and it is the most actionable thing published this week. The industry has trained everyone to evaluate models and accept whatever harness the vendor bundles, and that default has a price attached that nobody put on a slide. If you run agents at volume on direct API billing, the size of the prize here is not a rounding error. It is potentially half the bill for output you would not be able to tell apart in a blind test.
Two honest caveats before anyone rips out their toolchain. Thirty tasks per benchmark with three runs each is a small sample, and Arena says so. And if you are on a Pro or Max subscription rather than metered API access, the harness tax lands as rate-limit pressure rather than a dollar figure, which is harder to see and just as real. What I am doing about it: pulling per-task token counts for my own pipeline before I change anything, because my workload looks nothing like SWE-bench and neither does yours.
Google Keeps Its Ad Exchange and Gets a Six-Year Babysitter
Judge Leonie Brinkema’s full remedies opinion in the Justice Department’s ad tech case came out from under seal this week, two weeks after she filed the sealed version on September 2. All 106 pages, unredacted. AdExchanger’s readthrough confirms the headline everyone already knew, that Google keeps AdX, and Brinkema calls structural remedies neither realistic nor needed.
The behavioral terms are where it gets interesting for anyone who buys or sells open-web inventory. As PPC Land documents, Google has to build functionally equivalent API integrations connecting AdX and DFP to Prebid, and submit AdX bids to rival publisher ad servers on the same terms DFP gets. It has to hand publishers bid data covering both wins and losses, and publish technical documentation explaining how DFP picks a winner. AdWords is barred from bidding directly into DFP. A court-appointed technical monitor gets full access to Google’s employees, systems and source code for six years, and the obligations apply globally, sixty days from judgment.
Verdict: this matters more than the breakup headline suggested, and I think the coverage got the emphasis backwards. A breakup would have taken years of appeals to produce anything. Forced bid-loss transparency and a Prebid integration change auction mechanics on a sixty-day clock, on every continent, with an engineer in the building who can read the source. If you have spent years unable to explain why your open-web CPMs behave the way they do, the data that answers that is now a compliance obligation rather than a favor.
Anthropic Deleted the Word “Cowork”
As of September 16, Claude Cowork stops being a place you go and becomes something Claude does. Anthropic’s own announcement frames the merge as routing, where Claude works out how much machinery a request needs instead of asking you to pick a product first. Claude Docs and Claude Slides launched the same day in beta on paid plans, with export to Google Docs, Word, PowerPoint and PDF, and Claude Design now runs inside conversations too. Rollout hits Pro and Max on web, desktop and mobile over the coming weeks, then Team and Free, with Enterprise admins getting at least thirty days of notice.
Verdict: mostly marketing, with one genuinely load-bearing part. Docs and Slides are a checkbox response to Google Workspace and Microsoft 365, and Fortune is right to read the whole move as a superapp play. But the routing is not cosmetic. Making the model decide how much compute a task deserves is the same bet that sits underneath every agent product shipping this year, and Anthropic just made it the default surface for its consumer tiers. Watch what that does to your rate limits before you celebrate the slide deck feature. Anthropic’s line is that you “don’t have to do anything different,” which is true right up until the router disagrees with you about how much work your prompt needed.
A One-Month-Old Company Is Raising $700 Million
Bloomberg reports that Emulate, a UK startup founded roughly a month ago by former DeepMind world-model researchers Jack Parker-Holder, Matthew McGill and Philip Ball, is in advanced talks for a $700 million round led by Index Ventures and Lightspeed at about $3.7 billion. The company is building systems that simulate and predict how the physical world behaves.
Verdict: marketing, and I mean the round itself. Advanced negotiations are not a signed term sheet, the number can move, and reporting a deal that has not closed as a fact is how valuations get manufactured. The structural point is more useful than the outrage: this is the second ex-DeepMind team to price a seed round like a Series C this year, after Ineffable Intelligence took $1.1 billion at $5.1 billion in April. What is being bought at formation is not traction, it is a specific twenty-person research team and the certainty that somebody else will buy it if you pass. That is a labor market clearing at venture scale, and it tells you nothing about whether world models will work.
Spain Logged the First Breach an Agent Ran End to End
Spain’s data protection agency published a case on September 14 describing what it says is the first notified personal data breach carried out by design through an AI agent. Per SecurityWeek’s writeup, the chain ran login, then vulnerability probing across the victim’s applications, then modification of personal data, then access to invoices. The agency says a well known large language model was involved and has named neither the model nor the organisation, and it has not yet investigated or verified the report.
Verdict: this breaks your stack, though not for the reason it is being shared. Security consultant Simon Phillips put the honest version plainly, that there is not enough information yet to understand what happened or how. Treat the specifics as unconfirmed. What is confirmed is the shape: a human pointed an agent at a target and the agent chained reconnaissance through to data modification without needing a person at each step. Every one of us has now granted some agent a credential with write access to something that matters. The defensive work here is boring and overdue, which is scoping agent credentials to the narrowest possible permission and logging agent actions somewhere a human actually reads. Nobody will fund that this quarter, and it is the only item on this list you could start today.
What I Am Watching
The harness study is the one to act on, because it is the rare piece of research where the finding converts directly into a line item you control. The Brinkema remedies are the one to calendar, since sixty days from judgment arrives faster than any adtech team will be ready for. And Emulate is the one to ignore until money actually moves.