Price pressure: big customers are pushing OpenAI and Anthropic toward open models

By Zendoric · October 7, 2026.

Bloomberg reports, in its daily Tech In Depth newsletter written by journalist Rebecca Torrence (October 6), that OpenAI and Anthropic are starting to face pricing pressure from their highest-volume customers. The reason, according to the outlet: startups building software products on top of AI are increasingly turning to open-weight models — those whose parameters are published and can be run outside the vendor's cloud, on their own or third-party infrastructure — because they come out cheaper than the proprietary APIs of the sector's leaders.

The original article is behind a paywall and the available excerpt does not spell out which customers, how large the savings are or which specific contracts are at stake; the headline and lede do make the direction of travel clear, and that direction matches what we have been observing in the industry for months. It is worth saying so plainly: there is a headline and a thesis, but no verifiable figures to repeat here without making them up.

As industry context, the underlying reason is no secret. Open models such as Zhipu's GLM, Alibaba's Qwen, DeepSeek and Moonshot's Kimi have closed much of the quality gap with OpenAI's and Anthropic's closed models over the past year, and they run at a fraction of the cost per token because they do not carry the margin a frontier lab charges for a managed API. For a startup that bills by usage and competes on razor-thin margins, swapping a closed model for an open one on tasks that do not demand top-tier reasoning — classification, summaries, simple autocomplete — is a spreadsheet decision, not an article of technological faith.

The same Bloomberg newsletter notes, in a separate item, that Hon Hai (Foxconn), Nvidia's manufacturing partner, reported a 47% jump in quarterly sales thanks to spending on AI hardware. The contrast is telling: infrastructure demand is still surging upstream — chips, data centers, energy — while the price of AI software is being compressed downstream. It is the classic squeeze of a maturing market: the fixed costs of building the frontier keep rising just as the distribution layer starts fighting on price.

Our reading is that this does not herald the end of OpenAI's and Anthropic's business, but its segmentation. The market is splitting in two: everyday tasks are migrating toward cheap open models, while complex reasoning, mission-critical agents and environments that require certified security will keep paying the frontier-model premium, provided those models demonstrate with benchmarks — not with marketing — that the premium is worth it. This is exactly the mechanism by which AI gets cheaper and more democratic: good for the abundance we aspire to, uncomfortable in the short term for labs that need to finance multibillion-dollar rounds and giant data centers with customers increasingly willing to switch providers over a percentage point of margin. Whoever wins this phase will not be the biggest model, but the one that best knows how to say for exactly which task it is worth paying more.