Falling token prices won’t fix your AI budget

Claude Opus 5.5 was launched this week, and the announcement primarily focuses on pricing: it is about 40% cheaper to operate than Opus 5, costing $4 per million input tokens and $20 per million output tokens, with cache reads — which are noted to drive most of the costs for agentic and coding work — 60% cheaper at $0.20 per million. Additionally, output speed has improved, arriving more than 30% faster than before. Around the same time, a developer reported processing a billion tokens through DeepSeek 4.1-flash in just a week, costing approximately five dollars, compared to $25,000–$50,000 for the same workload using flagship models. This second figure should be considered as one user's report rather than an industry benchmark. However, the trend is clear: the cost of machine-generated intelligence is continually decreasing.

01What cheaper tokens actually change

• —Testing ideas: Trying out new ideas will no longer require a budget meeting. When prototypes cost mere pennies, the inquiry "Can we afford to test this?" becomes irrelevant, leading to more ideas getting the opportunity to be explored.
• —Extended agent work: Longer iterations involving review, retry, and refinement were previously curtailed by high token costs. With cheaper tokens, implementing patient automation becomes feasible.
• —Clarification of marginal work: Tasks like classification, extraction, and drafting are now more clearly defined. When the per-token cost approaches zero, the justification for automating these mechanical processes becomes self-evident.

02What cheaper tokens don’t address

The cost of generation has never been the primary expense in AI projects. The significant costs are found upstream and downstream of token usage: deciding what to build, verifying the returned results, integrating them into systems, overseeing data access, and maintaining outcomes. As these processes aren’t priced per token, they remain unaffected by reduced token costs.

An architecture based on faulty assumptions stays flawed, even at a 60% discount. In our cost-model analysis, we observed that teams could save substantially on generation expenses only to incur several times those savings in remediation costs — the generation phase was never the critical aspect.

03The bill that matters

As the cost of executing a task decreases, the remaining expenses often concentrate on judgment: determining which tasks require specific approaches, defining what constitutes "good" quality, and establishing who has the authority to approve work. This illustrates why asking "What does the model cost?" is the wrong initial question; instead, one should ask, "Which tasks, specifically, receive which treatment?" Cheaper tokens reduce the cost of machine operations but also make a lack of careful judgment easier to overlook — until the final product is launched.

So, take pleasure in the price reductions; more affordable intelligence is genuinely positive news. However, if the economics of your AI strategy rely solely on reduced token costs, there is a significant gap in the underlying thinking. It's crucial to evaluate the cost of judgment rather than just the cost of tokens.

Sources: Claude Opus 5.5 launch announcement by Addy Osmani and comparison of DeepSeek 4.1-flash costs by L. Charles Allard, both posted on LinkedIn on September 22, 2026.