April 2026
New Zealand Threat Landscape 2026
We’ve been pleased to see more of our customers embracing AI. Some are using it to improve internal productivity, some are exploring automation, and others are starting to think about how it could fit into the services they deliver. That momentum is good to see. But there is also a practical issue that needs to be on the table early: AI pricing is becoming more complex, more variable, and in some cases more expensive once real usage starts to scale.
For a while, many teams treated AI as if it would behave like ordinary SaaS: pick a seat price, roll it out, and assume the economics would stay broadly stable. That assumption is breaking down. What the last 12 months show is not a universal, across-the-board AI price hike. It’s something more important for builders: AI pricing is becoming more variable, more usage-based, and more sensitive to which model, feature, or workflow you choose.
GitHub Copilot is the clearest example. In May 2025, GitHub delayed enforcement of premium request limits so users could continue using premium models while it improved usage visibility. By June 2025, those limits became real. Today, Copilot Pro includes 300 premium requests, Pro+ includes 1,500, and overages cost $0.04 (USD) each. More importantly, different models consume that allowance at different rates. On April 16, 2026, GitHub launched Claude Opus 4.7 in Copilot at a 7.5x multiplier, versus 3x for earlier Opus models. Same subscription, same developer, very different burn rate. Not only that, they are going to retire the previous model early.
Microsoft Copilot points in the same direction, but through a different pricing design. Microsoft 365 Copilot Chat is now available at no extra cost for eligible Microsoft 365 users, but agents are pay-as-you-go. Full Microsoft 365 Copilot remains a paid per-user product, and Microsoft is also adjusting some surrounding bundle prices in 2026. In other words, the cost model is shifting away from “one predictable AI seat” and toward a mix of included chat, premium licenses, and metered automation.
Claude makes the story even more nuanced. Anthropic’s current API pricing actually shows Opus cheaper than it was at the Claude 4 launch, which proves this is not a simple “AI only gets more expensive” narrative. But lower nominal price does not always mean lower real-world cost. Anthropic also says Opus 4.7 can use up to 35% more tokens for the same text, some endpoints cost 10% more than global routing, and heavier users are pushed toward extra-usage billing or premium Max tiers. Effective cost can rise even when the headline price looks flat – or lower.
Do these pricing moves prove that every AI provider is losing money? No. Public data doesn’t support that blanket claim. But they do prove something builders should take seriously: AI providers are managing scarce compute, protecting reliability, and building stronger monetisation around high-end usage. GitHub explicitly says limited computing power and heavy concurrency require rate limits. Microsoft says it is expanding datacenters and server capacity for growing AI demand. Anthropic is investing tens of billions in compute infrastructure. The era of “unlimited frontier AI at a simple flat rate” looks increasingly temporary.
The practical lesson is simple: don’t scope AI implementations as if your costs are fixed. Budget for model routing, overages, fallback models, and repricing. If a feature depends on premium reasoning models or agentic workflows, a 50-100% cost swing is a real risk scenario – not because every vendor will definitely double prices next year, but because effective usage costs can already move that fast today.
