The Future of AI Pricing: Will Models Get Cheaper or More Expensive?

AI Costs Are Dropping. AI Prices Aren't. What Gives?

There's a disconnect in AI pricing that deserves attention. The cost to run AI models, measured in cost per token or cost per query, has fallen roughly 90% since GPT-4 launched in 2023. Training costs are dropping too, thanks to better hardware and more efficient training methods.

But consumer subscription prices haven't budged. ChatGPT Plus launched at $20/month in February 2023. It's still $20/month in March 2026. Claude Pro is $20/month. Gemini Advanced is $20/month. Three years of cost reduction, zero price reduction.

So what's going on, and where is pricing headed?

Why Costs Are Falling

Several forces are driving down the cost of running AI models:

Hardware improvements. NVIDIA's latest GPUs are significantly more efficient per dollar than the H100s that powered AI in 2023-2024. Custom chips from Google (TPUs) and Amazon (Trainium) are even cheaper for specific workloads. More compute per dollar means lower cost per query.

Model efficiency. Newer models achieve similar quality with fewer parameters and less compute. Techniques like mixture-of-experts, quantization, and distillation let companies run capable models on less hardware.

Scale economies. As AI providers process more queries, they can optimize infrastructure more aggressively. Fixed costs (data centers, research teams) are spread across more users.

Competition from open source. Models like Llama and Mistral's open weights models can be run on relatively inexpensive hardware. This puts a ceiling on what commercial providers can charge because users have a free alternative, even if it's slightly lower quality.

Why Prices Haven't Dropped

If costs are falling, why aren't prices? A few reasons:

R&D spending is massive. OpenAI reportedly spent over $5 billion on research in 2025. Anthropic and Google are spending in the same range. The cost savings from cheaper inference are being redirected to fund more expensive research. Subscription revenue subsidizes the next generation of models.

Price anchoring. The $20/month price point is established. Consumers are used to it. Dropping the price would cannibalize revenue from existing subscribers without necessarily attracting proportionally more new users. From a business perspective, it's rational to keep prices stable until competition forces a change.

Feature additions instead of price cuts. Instead of lowering prices, providers add more features at the same price. GPT-5 is better than GPT-4 was, but it costs the same. Consumers feel like they're getting more value without the price actually changing. This is a common strategy in tech: improve the product rather than lower the price.

No price competition yet. All three major providers charge $20/month. None of them has undercut the others. As long as they maintain this price equilibrium, there's no pressure to drop. It's not explicit collusion, but it's a stable Nash equilibrium that benefits all three companies.

What Could Change Prices

Aggregators and alternative pricing models. Companies like Admix already offer access to multiple models at lower prices ($8-24/month). If aggregators capture enough market share, the major providers may need to respond with lower prices or better bundles.

Open-source quality parity. If open-source models reach true parity with commercial ones, paying $20/month for a marginal quality advantage becomes indefensible. We're not quite there yet, but the gap closes with every release.

A price war. One major provider could decide to cut prices aggressively to grab market share. Google is the most likely candidate because they can subsidize AI with ad revenue. If Google dropped Gemini Advanced to $10/month, OpenAI and Anthropic would face pressure to match.

Usage-based pricing. The flat subscription model may give way to pay-per-use pricing. This would let light users pay less and heavy users pay more. Some API providers already use this model. If it extends to consumer products, average prices could drop because most users consume far less than $20/month worth of compute.

My Predictions

Short term (2026-2027): Subscription prices stay mostly flat. Providers add more value at the same price. Aggregators gain market share at lower price points.

Medium term (2027-2028): One major provider breaks the price equilibrium, either through a direct price cut or through bundling that makes the effective AI price much lower. Others follow within 6-12 months.

Long term (2028+): AI access becomes cheap or free for basic capabilities, similar to how email went from paid to free. Revenue shifts to enterprise, API access, and premium features. Consumer AI becomes a commodity priced at $5-10/month for full access, or free with ads.

What To Do Now

Don't overpay while waiting for prices to drop. If you're spending $40-60/month across multiple AI subscriptions, you can get equivalent access today through an aggregator for a fraction of the cost.

Admix offers 350+ AI models starting at $10/month (or $8/month billed annually), which is where I think mainstream pricing will eventually settle anyway. You might as well pay that price now instead of waiting two years for the big providers to catch up.

The AI pricing future favors buyers, not sellers. But there's no reason to overpay in the meantime.