A Month of Agentic Coding for Under $20
Right now, if you want Claude or Codex to write code for you, here are your options:
These companies are spending billions on GPU infrastructure. They're subsidizing consumer access to build habit and dependency. The $20/100/200 plans lose money. Everyone knows and accepts this. It's a land grab: get users locked in then raise their prices. Make them pay more for less.
What if instead of paying one company whatever they decide to charge, you could tap a network of competing inference providers--each running the same open models, each bidding to serve your requests--and the price was set by supply and demand, not by a pricing committee in San Francisco?
That's not hypothetical. That's what a free market for AI inference looks like.
The Problem: Monopoly Pricing
Today's AI inference market is a collection of monopolies. OpenAI sets GPT-5.4's price. Anthropic sets Claude's price. Google sets Gemini's price. You pay what they ask or you don't get to use it.
There's no competition on price for the same model. If you want Claude Opus, you buy it from Anthropic or a small selection of third-party routers. You pay Anthropic's price. Period.
But here's what most people haven't internalized: The frontier models are temporary monopolies. The open-source models are commodities. And commodities get cheaper when markets are allowed to work.
The numbers tell the story/. According to Epoch AI, ope-weight models now trail frontier proprietary models by three months on average. DeepSeek v3.2 delivers 90% of GPT-5.4's quality at 1/50th the cost. GLM-5 is within 3 points of Opus 4.6 on SWE-bench. Qwen 3.5 397B scores within striking distance of Claude Sonnet and GPT-5.4 on HumanEval+ and MBPP+. Llama 4 has a 10-million-token context window.
The 1,000x Cost Collapse
The cost of inference has dropped by approximated 1000x in three years. In early 2023, GPT-3.5-level performance cost around $400 per million tokens. Today, equivalent performance costs 40 cents or less.
Four factors compound: hardware improvements, software optimizations, model architecture efficiency, and quantization. Each improvement multiplies with the other leading to a 10x reduction in comparable performance inference cost annually.
But who captures those savings? Right now: the providers. Together AI charges $0.88/million tokens for Llama 70B Turbo. Groq charges 59 and 79 cents for input and output, respectively. The actual cost to serve these tokens--the electricity, amortized GPU, bandwidth--is a fraction of that. The difference is margin, overhead, and the absence of competition.
The Solution: A Two-Sided Marketplace
Imagine an inference marketplace that works like a commodity exchange:
Supply side: Anyone with GPU capacity can register as a provider. You run Llama 4 on your hardware, you connect to the network, you set your price per million tokens. If someone with a Blackwell cluster can serve it for 8 cents/million profitably, they offer it for $0.08.
Demand side: Developers and agents submit inference requests. They specify the model, their latency and token speed requirements, and the upper threshold they are willing to pay. They don't care who serves it, they care that a) the model is correct and unmodified, b) the latency is acceptable, and c) the price is the best available.
The routing layer: Sits between supply and demand. It verifies providers are running unmodified models via hardware attestation in TEEs. It tracks provider quality, latency, throughput, and reliability. It routes each request to the best provider based on the users' constraints. It handles payments per-token, per-request, per-stream, all at machine speed.
This is how commodity markets work in every other industry. Electricity. Bandwidth. Compute. The product is fungible--a token generated by GLM 5.1 on GPU cluster A is identical to one generated on GPU cluster B. When the product is fungible, the market drives the prices to the marginal cost.
What This Means for Developers
Ina competitive marketplace with 50 providers bidding to serve the same open model:
Recommended by LinkedIn
At that blended rate, a month of intensive AI-assisted coding--20 million tokens of total throughput--costs $12-20. Less than a single month of the anemic Claude Pro subscription. With no usage caps, not rate limits, no throttling, and no ads. The best model for each task, selected automatically.
And here's the part that should worry Anthropic and OpenAI: This price doesn't depend on subsidy. It's the value put on inference by the free market, set by competition, converging on the actual cost of production. It goes down over time, not up.
The Subsidy Trap
The current pricing from the major providers is not sustainable. The current pricing from the major providers is customer acquisition.
OpenAI launched ChatGPT plus at $20/month in early 2023. Three years later they are setting $200 tiers. The $20 plan still exists, but it gets progressively less capable relative to the higher tiers. Rate limits get tighter. The best models get reserved for premium plans. The experience degrades until you upgrade.
This is the classic SaaS playbook: hook at $20, extract at $200. It works when there's no alternative.
An open marketplace is the alternative. When the model weights are public and anyone can serve them, the subscription lock-in breaks. You're not paying for access to a proprietary model, you're paying for compute and compute is a commodity. The providers compete on price, latency, throughput, and reliability. The price converges on cost, not "what the market will bear."
The monopoly pricing of frontier models is temporary. Every proprietary capability gets replicated in open source within months. The subsidy pricing of consumer plans is a strategy not a service. An open marketplace makes both irrelevant.
The Trust Layer
"But how do I know some random GPU provider is actually running the correct model?"
This is the real engineering problem, and it's solvable with hardware. Trusted Execution Environments create hardware-sealed compute enclaves. The model weights are loaded into the enclave and the hardware generates a cryptographic attestation proving exactly which model is running. No one--not the operator, nor the ISP, or even the NSA can modify what is inside the enclave or read the data being processed.
This means:
The Bigger Picture
The centralized API model where a handful of companies set prices and terms for the entire industry is a temporary artifact that only works when the models are proprietary and the providers are few. That era is ending. Open-weight models trail frontier by three months, a gap which closes with each release. DeepSeek, Qwen, GLM, Llama--every quarter brings another open model that matches what was SOTA six months ago.
As open models reach parity, the value shifts from "who has the best model" to "who serves it cheapest and most reliably." That's a commodity market, and commodity markets have an inevitable outcome.
The price drops to the cost of production.
The companies that win in this world aren't the ones running the GPUs. They're the ones building the routing, reputation, verification, and payment infrastructure that makes the market work. The ones who make it so a developer or an AI agent can access any model, from any provider, at the best available price, with cryptographic guarantees that the inference is correct and private.
The future of AI isn't $200/month or $2,000/month subscriptions to one company's walled garden. It's an open marketplace where hundreds of providers compete to serve you.
That's what I'm building at https://usepod.ai
