ChatGPT API costs dropped 93% between 2024 and 2026. What cost $0.03 per thousand tokens now costs $0.002. The models that commanded $60 per million output tokens two years ago approach $8 today.
This is the fastest pricing collapse in enterprise software history, and it’s not over.
How We Got Here
The pricing war has three origins: DeepSeek, hardware advances, and good old-fashioned competition.
DeepSeek broke the market. When the Chinese startup released R1 in late 2025 at $0.55 per million input tokens—96% cheaper than OpenAI’s o1—Western providers had to respond. DeepSeek proved you could build competitive reasoning models without burning $100 million on training runs. The efficiency gap wasn’t sustainable.
Hardware caught up. GPU generations from H100 to H200 to B200, combined with custom chips from Google, Amazon, and Meta, reduced computational costs roughly 10x in 18 months. Meanwhile, engineering improvements—quantization, speculative decoding, KV-cache optimization—boosted efficiency another 3-5x. The same capability costs a fraction of what it did.
Five-way competition keeps pressure on. OpenAI, Anthropic, Google, Meta (open-source), and DeepSeek now all offer credible frontier models. Each price cut forces responses from the others. When GPT-4.1 launched 83% cheaper than GPT-4o mini, competitors matched within weeks.
Where Prices Stand Now
Current API pricing per million tokens:
OpenAI:
- GPT-4.1: ~$8 (output), down from $60 in 2024
- GPT-4.1 nano: $0.20
- GPT-5.2: Premium tier, approximately $15
Anthropic:
- Claude Opus 4.6: $15 (output), down from $75
- Claude Sonnet 4.6: Identical to Sonnet 4.5 pricing with better performance
- Claude Haiku: Sub-$1 for most workloads
Google:
- Gemini 3 Pro: $2 (input), $12 (output)
- Gemini 3 Flash-Lite: $0.25 input
- Gemini costs roughly one-third of 2025 levels
DeepSeek:
- R1: $0.55 (input), $2.19 (output)
- V3.2: $0.27 per million tokens
- Free tier available for low-volume users
March 2026 saw unprecedented volatility: 114 out of 483 tracked models changed prices in a single month—nearly 24% of all monitored models.
The Subsidy Problem
This pricing can’t last.
OpenAI burns approximately $14 billion annually. Both OpenAI and Anthropic are preparing IPOs. Current prices are, as OpenAI’s head of ChatGPT called them, “accidental”—a side effect of competitive pressure rather than sustainable economics.
Consumer subscription prices already diverge from API costs. ChatGPT Plus remains $20/month, unchanged since launch. Premium features command $30-50/month. The companies are discovering that enterprise contracts and subscriptions subsidize their race-to-the-bottom API pricing.
The business model tension is real: give away API access cheap to capture developers, then charge enough elsewhere to cover costs. It works while investor money flows. When the IPOs arrive, expect recalibration.
Who Wins
Developers and startups. Building AI products just got dramatically cheaper. A startup that would have spent $100,000/month on API calls in 2024 now spends $7,000 for equivalent usage. This lowers barriers to entry and accelerates experimentation.
Budget-conscious users. Pay-as-you-go pricing beats subscriptions for anyone using AI less than 20 days per month. A $0.99 day pass outperforms a $20 monthly subscription at lower usage volumes.
Emerging markets. Lower prices expand the addressable market. Users in India and Southeast Asia who couldn’t justify $20/month subscriptions can now access frontier models through API credits or free tiers.
Open-source infrastructure. Every price cut narrows the gap between proprietary and self-hosted models. When commercial APIs cost $0.20 per million tokens, running your own Llama instance makes less financial sense—which paradoxically validates the open-source approach.
Who Loses
Subscription models. The math no longer works for irregular users. Why pay $20/month for ChatGPT Plus when equivalent API access costs $3-5 for typical monthly usage?
Premium tutoring and professional services. If a $8 API call answers the same question as a $200/hour consultant, the consultant needs a new value proposition.
AI wrappers. Thin-margin businesses built on reselling API access face compression from both directions: providers cutting prices and users going direct.
Companies banking on pricing moats. If your competitive advantage was “we have API access and you don’t,” that advantage is evaporating.
What Happens Next
The consensus forecast: API prices fall another 40-60% per year for the next 18 months. Models costing $8 today approach $1-2 by late 2027.
But the war isn’t purely downward. March 2026 saw 114 pricing changes—some increases, particularly for newer capabilities. Companies are learning to segment: commodity features get cheaper, advanced reasoning and long-context become premium tiers.
The pricing war also masks growing revenue. Despite DeepSeek pressure, Anthropic’s revenue run rate jumped from $1 billion to $9 billion in one year. OpenAI expects $20 billion in 2025 revenue. Microsoft and Google saw cloud revenues grow 26% and 48% respectively since DeepSeek’s debut. Lower prices generate more usage, and more usage generates more revenue.
The question isn’t whether AI gets cheaper. It’s whether the companies selling it can make money at these prices. The current answer—heavily subsidized by venture capital expecting future returns—is unstable.
When IPOs arrive, so will pricing reality.