What Anthropic shipping Sonnet 4.6 means for Canadian operators

Sonnet 4.6 landed in late 2025 and quietly redrew the cost map for production AI in Canada. The headline is straightforward: Sonnet 4.6 reaches roughly ninety percent of Opus 4 capability on the workloads that matter to operators (code generation, structured tool use, long context comprehension) at about one fifth the cost. For Canadian businesses paying in US dollars on top of a soft loonie, that is the difference between AI being a line item and AI being a budget category.

We have already migrated four production agents from Opus to Sonnet 4.6 with no measurable drop in output quality. The retention monitoring agent we run for one client (which reads roughly 200,000 tokens of operational logs nightly) dropped from $14 a night to $2.80 a night after the swap. Annualized, that is a five thousand dollar savings on a single agent.

The implications for the firms we advise are concrete. First, the threshold for putting an AI agent on a workflow drops sharply. Workflows that did not pencil out at Opus pricing pencil now. Second, you should be auditing your existing prompts and chains for opportunities to downgrade. We are seeing many engineers default to the most expensive model out of caution; that caution costs real money in production. Third, the gap between hyperscaler AI (OpenAI, Anthropic, Google) and self-hosted open models continues to narrow on cost per task, but the speed of iteration on the closed side means the rational play for most operators is still hybrid: open models on the bulk paths, closed models on the paths that need their accuracy.

If you are operating any meaningful volume of agent traffic and have not done a model selection review since Sonnet 4.6 shipped, it is overdue. We do that review as part of our advisory retainer, or as a one-off engagement billed at our standard consulting rate.

The trend line here is more important than the specific number. Each major model release in 2025 reduced cost per quality unit by roughly half. If 2026 sustains that pace, the AI infrastructure decisions you make this quarter will be obsolete by Q3. Build for portability, not for any specific provider.