Token spend is now an engineering constraint with a budget line — compute strategy is product strategy. The complication: the price of any given capability keeps collapsing while total spend keeps rising. Both are true at once.
Two keynotes and talks in three tracks — with the same cautionary tale surfacing independently in two rooms. As far as I can tell, none of the speakers compared notes.
¶1Anannya Roy Chowdhury opened with the most relatable sentence of the conference: “I'm going to tell you how just running two agents cost me the rent of a month in one single day.” Chowdhury ▸ 00:00 A weekend demo game, two agents, $1,847 — because multi-agent costs compound: context replayed every turn, retries on failed validation, six hundred operations per game. Her fixes are the theme in practice: replace token-burn with actual math (a Bayesian belief map took a 15-thousandths-of-a-dollar decision to zero), heuristics for the obvious, the model only for ambiguity. Same system, same agents, 82% cheaper — and it played better. slide
¶2Sarah Sachs' keynote said the quiet part about the market: it's structured against buyers. A model gets upgraded at identical per-token pricing and quietly uses “three times as many output tokens” on the same task. ▶ spoken, ≈01:22 A $200-a-month coding plan can consume $5,000 of compute — a subsidy with strings. Her conclusion is close to a company policy: “You win on the product. You can't be winning on tokens.” ▶ spoken, ≈04:56 The playbook that follows: build multi-model, evaluate cost on the whole task rather than the API call, re-pick the default model every few weeks, keep open-weight models as leverage — and forgo discounts for optionality, because if you can't walk away you have no leverage. One more number worth the ticket price: “harness engineering and architecture decisions can account for about three-x the change in price as model selection.” Sachs ▸ ≈14:54
¶3Geoff Huntley zoomed all the way out: “software development now costs less than minimum wage” ▶ spoken, ≈00:17 — and when execution is that cheap, the unit economics of firms move: “it's not necessarily that there's AI layoffs as such. Just backfills have stopped.” Huntley ▸ ≈11:44 Meanwhile the Leadership room got the organisational ledger from Krishna kanth Mundada's “AI Tax”: across 22,000 developers, throughput up 66% — and incidents per PR up 242%, review time up 441%, churn up 861%. slide His survey headline: “89% of managers saw no measurable productivity improvement… they were paying the full rate but not filing any deductions.” Mundada ▸ ≈06:22 His history lesson — factories took forty years to see gains from electric dynamos — lands as reassurance and warning at once: the bill always arrives before the refund.
¶4One story surfaced in two rooms, independently: Uber burned its entire 2026 AI budget in four months. Sachs put the headline on a keynote slide; Mundada had the sourced version in the Leadership room — three agents, ten times the cost, no guardrails. slide, twice When two speakers who never met cite the same wreck, it's probably the wreck to study.
¶5The tension in this theme is genuine, and George Cameron's market data holds both ends: “you can get intelligence cheaper than ever — but we're also spending more than ever within our companies” ▶ spoken, ≈08:00 — because agentic loops run twenty to a hundred turns, and the turn count “acts as a multiplier on the cost.” ▶ spoken, ≈11:10 Jeremy Howard, off stage, named the incentive underneath: the vendors make more the more tokens you spend. The price of capability falls; the bill rises anyway. Which is exactly why the spend is engineering now.
“I'm going to tell you how just running two agents cost me the rent of a month in one single day.” The slide: “THE WEEKEND BILL — $1,847. One weekend. One event. One ‘simple’ game… 67% reasoning cost.”
“It might be the same price per token, but you run it on the same exact task — and what happens? It's three times as many output tokens.”
“You win on the product. You can't be winning on tokens.” And from the playbook slide: “Forgo discounts for optionality — if you can't walk away, you have no leverage.”
“Harness engineering and architecture decisions can account for about three-x the change in price as model selection.” And: “you're still paying token rates on errors and retries.”
“I'm going to say some pretty provocative things — like software development now costs less than minimum wage.”
“It's not necessarily that there's AI layoffs as such. Just backfills have stopped.” And near the close: “Removing waste from your systems and processes is a bigger accelerator than AI itself.”
“89% of the people believed they have not gained any marginal gain out of using AI… They were paying the full rate but they were not filing any deductions.”
“You can get intelligence cheaper than ever — but we're also spending more than ever within our companies.”
“Twenty to a hundred turns is quite sensible for many knowledge-work tasks — and it acts as a multiplier on the cost.”
“AI companies make more money the more tokens you spend, so they want to keep the conversation going… it's been trained to be agreeable and to keep you talking, because that's where the profits come from.”
“It's like everything is free now, because we can say go and do it. But it's not free — AI tokens are not cheap. Even if work's paying for them… someone somewhere in our organisation is paying for it.”
“OpenAI has a service tier — flex pricing… in off-peak hours it feels almost real-time, so you can basically get a 50% discount. We're not using agents at the same time as everyone else.”
“Two years ago, a majority of our contributors were humans… today a huge, growing majority of our customers are actually agents. They're not using the UI at all.”
“It might still make sense, probably for a lot in this room, for your regular coding activities: use the best model achievable and pay the price.”
Ride the collapse, or discipline the spend? The data says both at once: per-capability prices fall relentlessly, and companies spend more than ever, because agentic loops multiply every price fall away. Huntley celebrates the abundance — token-maxing is how a tour-guide operator becomes a developer. Sachs inverts the same phrase: the labs want you to token-max; your users want you to outcome-max. Nobody defended the one indefensible position: not knowing your unit costs.
| Engineers | Bound the problem before the model sees it; heuristics for the obvious, the LLM only for ambiguity — 60% of turns at zero cost in Chowdhury's rebuild. Count retries: you pay full token rates on errors (Sachs). Off-peak flex tiers are a real ~50% lever from this timezone (Hart). |
|---|---|
| Team leaders | Cost per completed task, not per API call — “that's where they get you” (Sachs). Re-select the default model on a calendar (Notion: every few weeks), not on vibes. Instrument spend so one runaway workload shows up in days, not at the end of the quarter. |
| Org leaders | Keep optionality: build multi-model, treat open-weight models as negotiating leverage, and forgo discounts that buy lock-in (Sachs). Guardrails before the mandate — that's the whole difference between the Uber story and the success stories (Mundada). Expect the J-curve: the bill arrives before the refund. |
| Key takeaway | Stop routing everything to the frontier model by default. Mundada: “you don't want to throw everything at super-high-cost Opus 4.8 or Codex 5.5 models — size it right; a lot of the cheaper models do really well, and that protects your scale.” |
Huntley says abundance: execution is nearly free, so spend it. Sachs says discipline: token costs compound and the incentives aren't your friend. Which mindset actually fits your product's economics — and where would each one hurt you?
For now these are placeholders. Once the individual sessions are up on Conffab, each title here — and every quote above — will link straight to the talk: video, transcript and all.