90% Discounts, Zero Payoff: Both Ends of the AI Bill
90% Discounts, Zero Payoff: Both Ends of the AI Bill
Read together, this week's AI signals say more than any single one of them: vendors are discounting harder than ever to win accounts, an economist says most companies have yet to see the money, and in between sit architecture consolidation and a collapse in labeling costs. The common thread is that value is shifting from model capability to unit economics.
Price war: discounts as an acquisition tool
According to The Decoder, OpenAI is offering discounts of up to 90% to Anthropic's largest customers. Companies including Rippling, Replit and DataCamp say their AI costs fell by as much as 90% after switching. The report also notes that Anthropic's growth has slowed, that OpenAI is trading its own margins for market share, and that Chinese rivals such as DeepSeek are piling on.
The implication for buyers is plain: the model layer is turning from a technical choice into a commercial negotiation. The same inference budget can carry a very different unit price depending on the vendor — and the quote depends heavily on your call volume and contract terms.
Architecture consolidation: is one agent enough?
Anthropic's public position has shifted: it now recommends that companies evaluate whether sub-agents can be consolidated into a single agent that loads context on demand. In one architecture comparison, the single-agent setup ran about twice as fast. Anthropic has also published a migration kit on GitHub.
The value of this signal is corrective. Over the past year, multi-agent orchestration has been treated as the default mark of a sophisticated system, but it also brings harder debugging, duplicated context passing and stacked latency. If models can now handle complexity alone, the orchestration layer has to justify its weight again — not every system needs five roles talking to each other.
A cost cliff: 72 cents versus $88
Nvidia has open-sourced Flora, a data engine for robot video labeling: one hour of robot video costs $0.72 to label, versus $88 with GPT-6 Astra — roughly 125x cheaper, per the report, with slightly higher quality scores. The system is built on three fine-tuned Qwen3.5 models, and the report notes it still falls short in some cases.
In the same week, Anthropic co-founder Jack Clark moved his prediction for self-improving AI forward to 2027, citing a new benchmark from P-Zero Research that measures how well models design and evaluate experiments on their own; results are improving faster than expected.
The payoff gap
TechCrunch reports that an Apollo economist argues AI has delivered "zero financial payoff" to corporate America outside of tech, despite massive investment and widespread hype. In roughly the same window, Foreign Affairs published an essay by historian Niall Ferguson, researcher Lisa Klaassen and former Google CEO Eric Schmidt casting the US and China as AI empires and every other country as a dependency — and arguing the race will be decided not by breakthroughs but by who can bankroll it.
Put "who can bankroll it" next to "zero financial payoff" and the contradiction is clear: the capex race and the company P&L do not run on the same timeline.
Conclusion: turn AI from a story into a ledger
Four actionable steps for readers:
- Renegotiate. If your primary model contract is more than six months old, re-quote at current market rates — and put migration costs (evaluation, compliance, engineering) into the comparison. A discount is not the same as savings.
- Audit your architecture. Check each sub-agent for necessity and whether a single agent loading context on demand can replace it. Measure latency and failure rates before refactoring.
- Try open-source data pipelines first. For high-volume, measurable labeling work (video, actions, logs), evaluate a fine-tuned small-model pipeline; the cost gap is usually larger than the capability gap.
- Define ROI with metrics. Before launch, write down measurable targets and an observation window, so investment scale does not substitute for evidence of return.
The technology is still moving fast, but procurement, architecture and accounting are three things you can start on now.
Sources
- OpenAI offers discounts up to 90 percent to poach Anthropic's biggest customers — The Decoder
- Anthropic says one AI agent beats many subagents now that models can handle complexity alone — The Decoder
- Nvidia open-sources Flora to make robot video labeling 125x cheaper than GPT-6 Astra — The Decoder
- Anthropic co-founder Jack Clark moves his prediction for self-improving AI forward to 2027 — The Decoder
- Ex-Google CEO Eric Schmidt calls OpenAI and Anthropic modern colonial trading companies — The Decoder
- AI delivers 'zero financial payoff' to most companies, Apollo economist says — TechCrunch
Subscribe to weekly briefing