Will ChatGPT get more expensive?

ChatGPT Plus has cost $20 a month since it launched in February 2023. In September 2024 the New York Times reported OpenAI projections of $44 a month by 2029, but two years later Plus is still $20. OpenAI has added tiers around it instead: Go at $8 and Pro from $100.

Why is AI getting cheaper?

Labs find ways to reach the same ability with smaller, more efficient models, and once several labs can reach a level they compete on its price. Epoch AI found costs fall fastest right after a level first appears, about 66% a quarter, slowing to about 32% a quarter two years later.

Are Chinese AI models cheaper than American ones?

At the low end in October 2026, the gap is small. On Artificial Analysis's figures, OpenAI's GPT-6 Luna and Anthropic's Claude Haiku 5.5 reach scores similar to DeepSeek V4.1 Flash at a lower cost per task. Open Chinese models can still come out cheaper through other hosts, because dozens of companies compete to serve them.

What is the cheapest way to use AI?

ChatGPT's free plan costs nothing and runs GPT-6 Luna, but it may show ads and OpenAI trains on your chats unless you opt out. The cheapest paid plans start at $4.99 a month for Google AI Plus and $5 a month for Plainly.

Will AI get cheaper? What the prices show

The same AI ability gets about 13 times cheaper each year (Epoch AI). Top models and $20 plans hold their price. Dated prices, and my forecast.

Will AI get cheaper?

For the same level of ability, yes, and fast. Epoch AI estimates that the cost of reaching a given level of AI performance has fallen about 47% a quarter since 2023, roughly 13 times a year. The newest top models and the $20 monthly plans have held their price, and my forecast comes from that split.

How fast AI prices have fallen

Epoch AI's report "The plunging price of thought" (Luke Emberson and David Roodman, 22 September 2026) tracks the cheapest model that reaches a set score on five benchmarks, counting the extra tokens reasoning models use. Costs fall fastest right after a level of performance first becomes state of the art, about 66% a quarter. Two years later the fall has slowed to about 32% a quarter. Epoch's explanation is that a lab can charge a premium for new top performance until competitors catch up.

An earlier Epoch analysis (March 2025) gives a like-for-like pair. GPT-4 scored 86 on the MMLU knowledge test in March 2023. Google's Gemini 2.0 Flash scored 88 in February 2025 at about a two-hundredth of GPT-4's price per token, in under two years.

List prices show the same pattern, and they show where it stops. These are per million tokens of input and output, from each company's pricing page on 11 October 2026.

Model	Tier	First sold	Price (input / output)
GPT-4	OpenAI's best	Mar 2023	$30 / $60
GPT-4 Turbo	OpenAI's best	Nov 2023	$10 / $30
GPT-4o mini	OpenAI's small model	Jul 2024	$0.15 / $0.60
GPT-6 Luna	OpenAI's small model	Sep 2026	$0.10 / $0.50
GPT-6 Astra	OpenAI's top GPT-6 model	2026	$10 / $50
Claude Haiku 5.5	Anthropic's cheapest	Oct 2026	$0.10 / $0.50 (prompts up to 100K tokens)
Claude Opus 5.5	Anthropic's highest-scoring	2026	$4 / $20
Gemini 3.8 Flash	Google's fast model	2026	$0.75 / $3.75 to 31 Dec 2026, then $1.50 / $7.50

GPT-6 Luna, OpenAI's newest small model, costs a three-hundredth of GPT-4's 2023 input price. GPT-6 Astra, its top GPT-6 model, still charges about five-sixths of GPT-4's 2023 output price.

Prices also go up. Google's pricing page (updated 9 October 2026) doubles Gemini 3.8 Flash's rates on 1 January 2027. DeepSeek changed its API prices in August 2026, a sharp rise according to The Next Web (10 September 2026), and cut them on 10 September (DeepSeek's change log).

What $20 a month buys

ChatGPT Plus has cost $20 a month since February 2023, and Claude Pro $20 since September 2023. Neither has come down. What changed is what the money gets you: ChatGPT's free plan now runs GPT-6 Luna, which launched on 22 September 2026, and Plus runs GPT-6.1 Sol.

The other change is new tiers. Below $20 there is ChatGPT Go at $8, Google AI Plus at $4.99 and Plainly at $5. Above it there is ChatGPT Pro from $100 and Claude Max at $100 and $200. Why the big plans cost what they do is covered in why is ChatGPT so expensive.

My own bills show both ends. I pay $100 a month for Claude Max 5× and still hit its usage limit, which is a common complaint about Claude's plans. Over the month from 16 August to 15 September 2026, the two everyday chat models I run through my own OpenRouter account (DeepSeek V4 Flash and GLM 5.3 Flash) cost me $18.90 between them.

My forecast, made in August 2026

China keeps shipping hyper-efficient models and absorbs every low-to-medium use case in a race to the bottom on cost. The US stays at the frontier and keeps fat margins on the most intense tasks. American models are too clunky and bloated to compete on efficiency, so the cheap end belongs to China.

The evidence against it

Two months later, the US labs are competing at the low end. Artificial Analysis runs every model through the same test set and reports what each run cost per task. These are its figures on 11 October 2026 (Intelligence Index version 4.3.2).

Model	Lab	Intelligence Index	Cost per task
Claude Opus 5.5	Anthropic (US)	58	$5.98
GPT-6.1 Sol	OpenAI (US)	52	$0.72
Claude Haiku 5.5	Anthropic (US)	43	$0.21
GLM-5.3-Flash	Z.ai (China)	42	$0.25
Gemini 3.8 Flash	Google (US)	41	$1.24
DeepSeek V4.1 Flash	DeepSeek (China)	39	$0.27
GPT-6 Luna	OpenAI (US)	38	$0.07

Each model is shown at its highest listed reasoning setting. GPT-6 Luna scores within one point of DeepSeek V4.1 Flash for about a quarter of the cost per task. Claude Haiku 5.5 scores higher than both and costs less per task than DeepSeek's model. OpenAI also gives Luna away in ChatGPT's free plan. Even my own OpenRouter month included a US model: Google's Gemini 3.7 Flash, at $1.93.

Epoch's report also found that near the frontier, price competition has mostly been between closed models: an open-weight model was the runner-up in only one of the eight price races it traced.

The evidence for it

The top end has held its price, as forecast. Claude Opus 5.5 costs $5.98 per task on the same test, more than 20 times DeepSeek V4.1 Flash. Claude Max at $100 a month still runs out for a heavy user like me.

Google's fast model is getting more expensive. Gemini 3.8 Flash already costs several times more per task than the cheap Chinese models at a similar score, and its rates double in January.

The Chinese models are also open, so competing hosts set their price. Twenty-nine companies were selling DeepSeek V4.1 Flash on OpenRouter on 11 October 2026. GPT-6 Luna had three hosts, all within 10% of OpenAI's own price. How that competition works is in why China open-sources its AI models, and why Chinese labs had to get efficient in the first place is in why Chinese AI models cost less.

On these numbers, the first half of the forecast holds: the US labs keep the top end and its margins. The second half, that US models can't compete on efficiency, doesn't match October 2026 prices: OpenAI and Anthropic now sell small models as cheap as the Chinese ones.

What would change the picture

A supply shock that cuts China's access to chips would slow the Chinese labs. A turn away from open weights by Chinese labs would end the host competition that keeps their prices low. If US labs follow Google and raise prices on their cheap models, the low end tips toward China.

What cheaper models mean for a $5 plan

When the cost of a level of ability falls 13 times a year, a fixed monthly price buys a better model each year. Plainly runs one capable open-source model and switches when a better one comes out. You never pick, and the price stays $5 a month.