The 2023-2026 model price war — and why your cost model is already obsolete — step 1 of 8
Two orders of magnitude in three years
GPT-4 launched in March 2023 at $30 per million input tokens and $60 per million output tokens. That price was the assumption every serious LLM company built their unit economics on for the next twelve months. Funding rounds got raised against it. Pricing pages got written against it. Headcount got hired against it.
By May 2026, the frontier looks like this:
- Claude Haiku 4.5 — $1 input / $5 output per 1M tokens
- Claude Sonnet 4.6 — ~$3 input / $15 output per 1M tokens
- Claude Opus 4.7 — $5 input / $25 output per 1M tokens
- GPT and Gemini frontier tiers track within a small multiple of these
Read those numbers carefully. Haiku 4.5 input is 30× cheaper than GPT-4 input three years earlier. Sonnet 4.6, which is a genuinely better reasoner than the original GPT-4, is 10× cheaper on input and 4× cheaper on output. The "expensive" tier in 2026 (Opus) is the same price as the cheap tier in 2024.
This is not a one-time drop. It's a power-law collapse.
What the curve actually looks like
A16z's LLM pricing tracker (the canonical chart in the industry, updated quarterly) shows the per-token cost of "GPT-4-class intelligence" falling roughly 10× every twelve months since the original release. Dylan Patel's SemiAnalysis breakdowns of the inference economics give you the why: smaller-better models, better serving infrastructure, FP8/FP4 quantization, speculative decoding, and a brutal multi-lab competition for developer mindshare.
The shape of it:
GPT-4 launch (Mar 2023): $30 / $60 per 1M ← baseline
GPT-4 Turbo (Nov 2023): $10 / $30 ← 3x cheaper input
Claude 3 Sonnet (Mar 2024): $3 / $15 ← 10x cheaper input
GPT-4o (May 2024): $5 / $15
Claude 3.5 Sonnet (Jun 2024):$3 / $15
Haiku 3.5 (Oct 2024): $1 / $5 ← cheap-tier emerges
Sonnet 4.5 (mid 2025): $3 / $15
Haiku 4.5 (Oct 2025): $1 / $5 ← 30x cheaper input
Sonnet 4.6 (early 2026): $3 / $15
Opus 4.7 (April 2026): $5 / $25
The frontier label sits at roughly the same price point each year ($3/$15 for "best general workhorse"), because the labs keep relabeling. But the capability that used to live at $3/$15 keeps sliding down to $1/$5 within twelve months. Yesterday's flagship becomes today's cheap tier.
Why this matters more than the model itself
If you built a product in 2023 with GPT-4 cost assumptions baked in, you over-paid by 99% for three years. If you build a product in 2026 with Sonnet 4.6 cost assumptions baked in, you will probably overpay by 90% by 2027.
The strategically interesting question is not "what does my feature cost today?" It is "what does my feature cost when the model I'm using is one-tenth the price in twelve months?" Features that are unaffordable at Sonnet prices but viable at Haiku-4.5 prices are features your competitors will ship in 2027 even if you can't ship them today.
Three implications worth burning into your head before the next step:
- The unit economics of any LLM feature are dated within twelve months. You are not pricing against today's model. You are pricing against the model that will exist when the contract renews.
- The biggest lever is "which model" — not prompt engineering. Switching from Opus to Sonnet is a ~1.7× cost reduction. Switching from Sonnet to Haiku is another 3× reduction. No amount of prompt trimming touches that.
- The companies that won the last three years were not the ones with the best 2023 product. They were the ones who restructured their cost model fastest as prices fell. The next two case studies show this in detail.
The 2023-2026 model price war — and why your cost model is already obsolete — step 1 of 8
Two orders of magnitude in three years
GPT-4 launched in March 2023 at $30 per million input tokens and $60 per million output tokens. That price was the assumption every serious LLM company built their unit economics on for the next twelve months. Funding rounds got raised against it. Pricing pages got written against it. Headcount got hired against it.
By May 2026, the frontier looks like this:
- Claude Haiku 4.5 — $1 input / $5 output per 1M tokens
- Claude Sonnet 4.6 — ~$3 input / $15 output per 1M tokens
- Claude Opus 4.7 — $5 input / $25 output per 1M tokens
- GPT and Gemini frontier tiers track within a small multiple of these
Read those numbers carefully. Haiku 4.5 input is 30× cheaper than GPT-4 input three years earlier. Sonnet 4.6, which is a genuinely better reasoner than the original GPT-4, is 10× cheaper on input and 4× cheaper on output. The "expensive" tier in 2026 (Opus) is the same price as the cheap tier in 2024.
This is not a one-time drop. It's a power-law collapse.
What the curve actually looks like
A16z's LLM pricing tracker (the canonical chart in the industry, updated quarterly) shows the per-token cost of "GPT-4-class intelligence" falling roughly 10× every twelve months since the original release. Dylan Patel's SemiAnalysis breakdowns of the inference economics give you the why: smaller-better models, better serving infrastructure, FP8/FP4 quantization, speculative decoding, and a brutal multi-lab competition for developer mindshare.
The shape of it:
GPT-4 launch (Mar 2023): $30 / $60 per 1M ← baseline
GPT-4 Turbo (Nov 2023): $10 / $30 ← 3x cheaper input
Claude 3 Sonnet (Mar 2024): $3 / $15 ← 10x cheaper input
GPT-4o (May 2024): $5 / $15
Claude 3.5 Sonnet (Jun 2024):$3 / $15
Haiku 3.5 (Oct 2024): $1 / $5 ← cheap-tier emerges
Sonnet 4.5 (mid 2025): $3 / $15
Haiku 4.5 (Oct 2025): $1 / $5 ← 30x cheaper input
Sonnet 4.6 (early 2026): $3 / $15
Opus 4.7 (April 2026): $5 / $25
The frontier label sits at roughly the same price point each year ($3/$15 for "best general workhorse"), because the labs keep relabeling. But the capability that used to live at $3/$15 keeps sliding down to $1/$5 within twelve months. Yesterday's flagship becomes today's cheap tier.
Why this matters more than the model itself
If you built a product in 2023 with GPT-4 cost assumptions baked in, you over-paid by 99% for three years. If you build a product in 2026 with Sonnet 4.6 cost assumptions baked in, you will probably overpay by 90% by 2027.
The strategically interesting question is not "what does my feature cost today?" It is "what does my feature cost when the model I'm using is one-tenth the price in twelve months?" Features that are unaffordable at Sonnet prices but viable at Haiku-4.5 prices are features your competitors will ship in 2027 even if you can't ship them today.
Three implications worth burning into your head before the next step:
- The unit economics of any LLM feature are dated within twelve months. You are not pricing against today's model. You are pricing against the model that will exist when the contract renews.
- The biggest lever is "which model" — not prompt engineering. Switching from Opus to Sonnet is a ~1.7× cost reduction. Switching from Sonnet to Haiku is another 3× reduction. No amount of prompt trimming touches that.
- The companies that won the last three years were not the ones with the best 2023 product. They were the ones who restructured their cost model fastest as prices fell. The next two case studies show this in detail.