Welcome to another edition of Software Talks AI, Monterro's monthly read on AI for B2B SaaS leaders. On the agenda this month: the token cost transition hitting AI tools across the board, the agent platform race converging on a recognisable harness shape, an EU AI Act update, and Kaisa stealing the show at Autosbuzz by landing Mobile.de as its first AI customer.
But first, let's talk token costs. Tokens are the world as AI models see it. They don't see words the way we do, they see tokens. Tokens are what we pay for, in energy or money, when we use AI. We've been living in a golden era of AI usage. AI now has immense agentic capabilities which we've been able to use at a pretty steep discount. That discount is starting to disappear, as AI tools are transitioning their payment models.
GitHub Copilot, for instance, transitioned from feature-based to token-based AI credits on June 1st, which will increase costs by as much as 5x for some teams. Cursor made the same move in June 2025, Windsurf in March 2026, Replit through effort-based credits, and OpenAI Codex on April 2, 2026. Headline subscription prices look stable at $10, $20, and $39, but those numbers now buy a pool of model spend rather than a fixed number of requests. As GitHub itself acknowledged, a handful of requests can now incur costs that exceed the entire monthly plan price.
Frontier model rates are not following the cheap-and-abundant narrative either. OpenAI doubled GPT-5.5 to $5 input and $30 output per million tokens in April. Anthropic kept Opus 4.7 at $5 and $25 but shipped a new tokenizer that produces up to 35 percent more tokens for the same input text, raising effective costs 20 to 35 percent with no change to the rate card. Google's Gemini 3.1 Pro is priced 60 percent higher than 2.5 Pro, and while the recently announced Gemini 3.5 Flash is more affordable than 3.1 Pro, it is significantly more expensive than past models in the series.
A clear $100 to $200 per-seat ceiling for power users has formed across the industry. Anthropic Claude Max, OpenAI ChatGPT Pro, Cursor Ultra, Windsurf Max, and Claude Code Max all converge at the same two price points.
In the short term, we need to be ready for a price hike in AI services, but there are many small adjustments we can make to improve token efficiency. When using AI services, we can save tokens by spending more time planning before building, and when implementing AI into products, we can use techniques such as prompt caching and batch APIs to reduce token costs by as much as 90%. In both cases, we can also consider using cheaper models. Regardless, we must be model fluid.
In three-to-five years, we believe inference will be drastically cheaper and abundant. Capacity gets built, open-source catches up on the second tier, and models become optimized. Plan for cheap and abundant. Operate as if it is not. Do not lock product pricing to today's inference COGS. And do not invest in defensibility that depends on which model you call rather than what you can do with cheap intelligence that competitors cannot. Tokens get cheaper. Proprietary data, customer workflows, integration depth, and trust do not. That is where capital should compound.