The price of AI inference is in freefall. Not gradually. Not modestly.
In 2026, the cost of calling a frontier AI model via API has dropped in ways that would have seemed implausible two years ago. OpenAI’s current pricing page reflects meaningfully lower rates than the company charged at GPT-4’s original commercial launch. AI industry tracking reports a reduction of approximately 60% for GPT-4 API access, though that figure comes from a single aggregator source and hasn’t been independently audited. Anthropic’s published Claude pricing shows its most efficient model at $0.80 per million tokens, a price point that reflects the same competitive pressure. A specific reduction percentage for Claude 3.5 Sonnet wasn’t confirmed at publication.
The 60% figure may or may not be exactly right. The direction is beyond dispute.
Why Costs Are Falling This Fast
Three forces are working simultaneously, and they’re all pointing the same way.
The first is compute efficiency. The cost of running a model isn’t fixed, it tracks with chip performance improvements and inference optimization research. Every hardware generation, every flash attention update, every quantization breakthrough brings the per-token cost down. Labs that aren’t passing those savings to customers lose developers to those that do.
The second is open-source competition. Models like Meta’s Llama series have given enterprise engineering teams a credible alternative to paying API fees at all. When capable open-weight models are freely available and deployable on commodity hardware, the frontier labs face a ceiling on what they can charge for API access. The competitive floor isn’t just other commercial providers, it’s free.
The third is Chinese AI providers. DeepSeek’s emergence in early 2025 demonstrated that state-subsidized or highly efficient Chinese models could deliver competitive performance at dramatically lower price points. That introduced a new category of competitive pressure that Western providers couldn’t simply ignore. Pricing had to respond.
These three forces aren’t temporary. They compound over time. The trajectory of falling inference costs isn’t a 2026 aberration, it’s the permanent operating condition of this market.
What OpenAI and Anthropic’s Moves Signal
The providers making pricing adjustments aren’t doing so reluctantly. They’re doing so strategically.
For OpenAI, lower API prices expand the addressable market. An enterprise that couldn’t economically justify running GPT-4 inference at 2024 pricing can recalculate at 2026 pricing. Volume offsets margin compression if the market grows fast enough. This is a deliberate land-grab for enterprise infrastructure dependency.
Anthropic’s position is slightly different. Reported pricing adjustments to Claude’s API come alongside Anthropic’s push into enterprise contracts. Lower API costs reduce the switching cost for companies evaluating Claude as an alternative to OpenAI. That’s a customer acquisition mechanism dressed up as a pricing decision.
The aggregate claim, that AI API costs have dropped 40 to 70 percent across major providers in 2026, comes from a single aggregator source and should be treated accordingly. But the competitive logic supporting that general direction is independently sound.
What Remains Unknown
Two things matter that aren’t confirmed.
First, the specific percentage reductions for individual models haven’t been verified against historical pricing records. The 60% GPT-4 figure, and whatever the Anthropic Sonnet reduction is, need a confirmed baseline price to be meaningful. Without that baseline, percentage claims are suggestive, not definitive.
Second, Google’s pricing position is absent from this analysis because it wasn’t in the verified package. In any serious treatment of the AI inference pricing landscape, Google’s Gemini API rates are relevant data. Readers tracking enterprise AI procurement should pull Google’s current pricing independently.
Enterprise Budget Implications
Three practical implications follow from the verified directional facts.
Contracts negotiated before mid-2025 likely don’t reflect current pricing. Enterprise AI agreements with volume commitments and fixed per-token rates are worth reviewing. The question isn’t whether prices have fallen, it’s whether your agreement captures the current market.
Build-vs-buy calculations have shifted. If your 2024 analysis concluded that an open-source deployment was economically superior to a commercial API, that math has changed. Lower commercial API costs narrow the total cost of ownership gap.
Provider lock-in risk is lower than it was. When all three major providers are cutting prices in response to the same competitive pressures, switching costs decrease. Enterprise buyers have more negotiating leverage now than at any point in the GPT-4 era.
The structural condition of this market is persistent deflation. Budget for it.