AI
Inference costs fell again this quarter — but not for everyone
Headline per-token pricing keeps dropping. Long-context and tool-heavy workloads are moving the other way, and that gap is reshaping product roadmaps.
Key takeaways
- Average per-token inference pricing has fallen materially compared with a year ago
- Long-context and tool-heavy workloads have not seen the same cost declines
- Teams are now budgeting by task completion rather than by token price
The cheap-tokens narrative is real but incomplete. Averaged across providers, simple text generation is materially cheaper than a year ago. Workloads that hold large contexts open, call tools repeatedly, or run reasoning at depth have not tracked that curve.
Why the averages mislead
Teams building assistants pay for retries, verification passes and orchestration overhead that never appear in a price-per-million-token comparison. Several product leaders now budget by task completion rather than tokens.
The practical consequence: architecture is a cost decision. Caching, smaller routing models, and strict output bounds are doing more for unit economics than any vendor discount.
Why it matters
AI product teams face rising costs for complex workloads even as headline pricing falls, forcing architecture decisions to control spend. Vendors offering flat token discounts may understate real costs for reasoning-heavy applications.
Sources and references
- Vendor price sheets
- Updated 16 Sept 2026 · 16:20 UTC
- AI-assisted draft, edited and fact-checked before publication · Reviewed by Advisable Desk.
In this story
- Inference pricing
Related stories
Small models on device start winning real workloads
Latency, privacy and cost are pushing classification, extraction and routing off the cloud — and product teams are noticing users prefer it.
Daniel Okafor · 3 days ago · 5 min read
Agentic checkout is quietly rewiring the front door of online retail
Assistants that browse, compare and buy on a shopper's behalf are moving from demo to production. Retailers now face a harder question than adoption: who owns the customer relationship when software does the clicking?
Daniel Okafor · 6 hr ago · 7 min read
Acquiring margins compress as payment orchestration spreads
Merchants routing the same transaction across several acquirers are turning processing into a spot market. Pricing teams are discovering how little of their spread was ever defensible.
Inês Ferreira · 5 hr ago · 6 min read
Daily Brief
One email each morning. The stories that actually change decisions.
Five minutes of e-commerce, AI, payments, business and startup news, edited by humans and sent before markets open.
No spam. Unsubscribe in one click.