OpenAI's 2023 price cut turned chat completion into a commodity

A price attached to a new endpoint

On 1 March 2023 OpenAI published the announcement that opened its chat model to outside developers through an API. The post states the new model, gpt-3.5-turbo, was "priced at $0.002 per 1k tokens, which is 10x cheaper than our existing GPT-3.5 models", and it credits "a series of system-wide optimizations" for a stated 90 percent cost reduction in serving ChatGPT since the previous December. A vendor's own account of why a price fell is a claim about engineering, not an audited cost breakdown, but the price itself is a dated fact: a specific rate, posted on a specific day, for a specific model snapshot.

What the number was measured against

The tenfold comparison in the announcement is to "our existing GPT-3.5 models", principally text-davinci-003. OpenAI's own current deprecations page, a living document reflecting its state on 16 September 2026, records that text-davinci-003 was priced at $20.00 per 1 million tokens, equivalent to $0.02 per 1,000 tokens - ten times the $0.002 rate quoted for gpt-3.5-turbo. That page also shows text-davinci-003 was later retired, with a shutdown date of 4 January 2024, and recommends migrating to gpt-3.5-turbo-instruct. The announcement further set a support window for the specific snapshot released that day, gpt-3.5-turbo-0301, promising it would remain available "through at least June 1st".

A record, not a rate card

None of this establishes what a token costs today, or what any given task should cost to automate. The announcement is a single vendor's dated price for one model family, quoted against a comparison of the same vendor's choosing. It says nothing about a task's total cost, which depends on prompt length, output length, retries, and the surrounding system. It also predates a naming change: OpenAI later added a note that "the ChatGPT API name has been discontinued" and that references to it should be read as the GPT-3.5 Turbo API, a reminder that even a vendor's own historical post gets edited after the fact. Treat a headline price as evidence of what happened on 1 March 2023, not as a benchmark for what a workflow will cost when read years later.

• What is the current, dated price for the specific model version under evaluation, and does the source state whether that price is stable or provisional?
• Does a per-token comparison account for typical prompt and completion lengths for the task at hand?
• Has the cited model snapshot been deprecated, and what does the vendor's own deprecation record say about a replacement?

A price cut can change which tasks are worth automating, but only for as long as the price holds and the model behind it stays available. Reading a 2023 announcement well after the fact means treating it as a dated record of a decision, not as a quote a reader could still act on.