On September 22, 2026, both OpenAI and Anthropic dropped new models at prices that would have been unthinkable 18 months ago. GPT-6 Sol launched at $2/$10 per million tokens, half the cost of GPT-5.6 Sol. GPT-6 Luna, the budget tier, shipped at $0.10/$0.50. Anthropic's Claude Opus 5.5 came in at $4/$20. BenchLM's frontier LLM token price index now sits at 15, 85% below its March 2023 baseline of 100.

I've led 50+ engagements. The most common excuse I hear for stalled AI projects is cost. "We can't afford to run this at scale." "The token budget kills our unit economics." "We're waiting for prices to drop."

Well, prices dropped. Dramatically. And I can tell you right now: most of those stalled projects are still going to fail. Because the cost of the model was never the real problem.

The Price Drop Is Real and Unprecedented

Let's be specific about what just happened. In a single week in September 2026, we got:

This isn't incremental. A frontier-quality call of a thousand tokens in and a thousand out cost about $0.04 on GPT-4 Turbo in early 2024 ($10/$30 per million). On GPT-6 Sol the same call costs just over a cent. For batch workloads using a Luna-tier model, you're looking at fractions of a penny per request. The compute barrier to running AI at enterprise scale is effectively gone.

Why Cheaper Models Won't Fix Your ROI Problem

Here's the uncomfortable truth that the pricing headlines bury. Enterprise AI adoption hit 88% in 2025, per McKinsey's State of AI survey: nearly nine out of ten organizations now use AI in at least one business function. But only 39% report any enterprise-level EBIT impact, and most of those put it below 5%.

The stat that should keep every CTO up at night: 95% of organizations running generative AI pilots saw no measurable profit-and-loss impact at all, according to MIT NANDA's GenAI Divide report.

Cheaper models don't fix this. The root cause of failed AI projects is almost never the cost of inference. It's poor data quality, weak system integration, and the absence of evaluation infrastructure. These are the same problems I've been diagnosing across healthcare, fintech, and SaaS for years, and they're model-price-independent.

What cheaper models actually do is make it cheaper to scale failure. An organization that deployed a poorly integrated RAG system at $0.04 per call can now run it at about a cent per call and still get the same garbage outputs, just more of them, faster.

The CTO Playbook: Redirect Savings Into What Actually Matters

If your AI inference costs just dropped by 50-80%, you didn't save money. You freed up budget. Here's where to put it:

1. Data Quality and Integration

Across the AI engagements I've led, most project failures trace back to the data layer, not the model. Connectors that drop permissions. Chunking strategies that destroy tables and charts. ETL pipelines that were "good enough" for dashboards but fall apart when an LLM tries to reason over the output.

The price collapse means you can now afford to run more sophisticated retrieval pipelines (multi-hop queries, re-ranking, hybrid search) without blowing your token budget. But only if you invest in the plumbing first. Take 30% of your inference savings and put it into data integration engineering. It's the highest-ROI investment you'll make this year.

2. Evaluation and Observability

If you can't measure whether your AI system is working, you can't improve it. And if you can't prove it's working, your CFO will kill the budget, regardless of how cheap the models are. Yet most enterprise AI deployments I audit have no evaluation pipeline beyond "someone from the product team checks the outputs occasionally."

With model costs down, you can now afford to run LLM-as-judge evaluations on a meaningful sample of production traffic. A frontier model evaluating your outputs costs 5-10% of your inference budget. That's a small price for knowing whether your system actually delivers value, and for having the data to prove it to leadership.

3. A Multi-Model Routing Strategy

This is the biggest tactical opportunity of the price collapse. With GPT-6 Luna at $0.10/$0.50 and frontier models at $4-10/$20-50, there's a 40-100x price spread between tiers. Smart teams are building routing layers that classify incoming requests and send them to the cheapest model that can handle them.

Here's the framework I use with clients:

Routing the majority of volume to a budget tier typically cuts effective per-request cost by well over half. Quality metrics usually hold, because most of those requests are simple classification tasks that a $0.10 model handles just as well as a $4 model.

Forrester's Warning: The Correction Is Coming

Forrester's 2026 predictions (published October 2025) say enterprises will defer 25% of planned 2026 AI spend into 2027. This isn't a vote against AI. It's a vote against undisciplined AI spending. Organizations are realizing that throwing models at problems without investing in integration, evaluation, and data quality produces impressive demos and zero ROI.

The price collapse accelerates this reckoning. When you can't blame model costs for your failures, the conversation shifts to execution quality. And that's exactly where it should be.

I've watched this pattern before, in cloud migration, in microservices adoption, in every technology wave. The hype cycle peaks, the tools get cheaper, and the organizations that succeed are the ones that focused on fundamentals while everyone else chased the next shiny thing.

What I'd Do Monday Morning

If I were walking into an enterprise as Fractional CTO next week, here's the sequence:

  1. Audit your current AI spend. Map every model call to a business outcome. If you can't trace a call to revenue or cost savings, it's waste.
  2. Implement model routing. Classify your request types and match them to the cheapest capable model. This is a 2-week engineering project that pays for itself immediately.
  3. Build an eval pipeline. If you don't have one, you're flying blind. Start with 50 golden examples and an LLM-as-judge setup. This is a 1-week project.
  4. Redirect 30% of inference savings to data quality. Fix your connectors, your chunking, your permission models. This is where the ROI actually lives.
  5. Set a 90-day ROI checkpoint. If you can't show measurable business impact in 90 days, restructure or kill the project. Cheap models make it tempting to let mediocre projects linger. Don't.

The LLM price collapse is real, and it's permanent. But cheaper models are a tool, not a strategy. The CTOs who win from here are the ones who treat the savings as investment capital, and put it into the boring, unglamorous work that actually makes AI deliver value.