All news

OpenAI cuts the price of frontier reasoning from both ends

A graphite bench vise closing from both sides on a small luminous glass cube

GPT-5 arrived at half the input price of GPT-4o, two days after OpenAI's first open-weight models in six years. Value shifts from model access to products.

GPT-5 costs $1.25 per million input tokens in OpenAI's API, half the $2.50 charged for GPT-4o, and OpenAI presents it as its strongest model. It arrived on 7 August, two days after the company released gpt-oss-120b and gpt-oss-20b, its first open-weight models since GPT-2.

In the space of three days OpenAI reset the price of capable models from both directions, and the margin that gives way is the one at the model layer. A frontier lab now sells its best model for less than the one before it and hands out a strong reasoning model for anyone to host, the move DeepSeek made in January. That compression lands while the labs commit tens of billions of dollars to compute. They can carry it through scale and distribution, but their valuations now rest on volume and on the products they build on top, not on a premium per token.

The open end of the squeeze matters as much as the discount. Both open models ship under the Apache 2.0 license, the larger one has 117 billion parameters and runs on a single 80 GB GPU, and Amazon put both on Bedrock, the first OpenAI models offered there. GPT-5 itself reached free ChatGPT users from day one and bills $10 per million output tokens. Companies whose value is reselling model access with a markup lose ground. Money moves to whoever owns the workflow, the proprietary data and the customer relationship.

Cost per task fell again, and a capable model can now run inside a customer's own environment. That opens regulated buyers in health, finance and the public sector that will not send data to an outside API. The sound design routes between a cheap closed model, a self-hosted open one and a premium tier, with gross margin built on the assumption that today's price is a ceiling. Reasoning tokens bill as output, so the real cost depends on how hard the model thinks at each step, and products that set reasoning effort per task will hold margin better than products that leave it on high.

One lab's discount does not yet amount to a price war. Anthropic and Google decide that: cuts on their top models this quarter would confirm one at the frontier. How fast the large clouds and enterprise software vendors offer gpt-oss as a managed service behind company firewalls will show how quickly the open models reach regulated buyers. OpenAI's own revenue is the final test. If usage grows faster than price falls, lab economics hold, and if it does not, the labs will push harder into applications and compete with their own customers.

Intelligence is getting cheaper faster than most business plans assume, which makes the pricing power of any single model a poor thing to bet on. Fanaura would rather own companies that turn cheaper models into durable revenue, the core of how it invests in AI-native companies.

For founders

Building products that run open-weight models inside regulated customers? Send us the project.

Fanaura invests its own capital and the capital of its limited partners (LPs), usually through a separate SPV for each deal. If we see a fit, we get in touch.

Apply in five minutes
Subscribe

Get the news by email

One short note each weekday. No marketing, and you can unsubscribe at any time.