The 2023-2026 model price war — and why your cost model is already obsolete — step 2 of 8
Three bets and how they aged
Theory only sticks if you can see it in actual P&Ls. Three companies, three different cost-model bets in 2023, three very different outcomes by 2026.
Bet 1: Jasper — "GPT-4 unit economics are the unit economics"
Jasper was the marketing-copy company. By 2022 they were doing ~$80M ARR generating ad copy and blog posts on top of OpenAI. In October 2022 they raised $125M at a $1.5B valuation, the textbook "AI-native" win.
Their cost model assumed GPT-3.5 and GPT-4 prices were permanent floors. They priced subscriptions around them. They hired around them. They built features around them.
Then the prices fell, and the floor became a ceiling. Three things happened in parallel:
- Their margins on individual API calls became absurd as costs fell — for about six months — until competitors realized and undercut on price.
- New entrants (Copy.ai, then a wave of vertical AI writers) shipped the same feature on cheaper models and charged $20/month where Jasper charged $99.
- ChatGPT released a free consumer tier, then Anthropic and Google followed, then everyone shipped "Write copy with AI" as a feature inside the tools the customer already paid for (Notion, HubSpot, Mailchimp).
Jasper laid off ~30% of staff in 2023, restructured, and the path to IPO that the $1.5B valuation implied has not materialized. The cost model they priced against didn't exist twelve months later. They survived as a smaller company, but they did not capture the upside their cap table assumed.
Lesson: pricing your product against today's model costs is pricing against a melting ice cube. Anyone who shipped the same feature at Haiku prices six months later took the customer.
Bet 2: Perplexity — "route to the cheapest model that meets the bar"
Perplexity launched as an "answer engine" in late 2022. They had a fundamentally harder cost problem than Jasper: every user query required (a) a search step, (b) a synthesis call against a frontier model, (c) a re-ranking pass. Three LLM calls per query, often more for complex questions, at GPT-4 prices.
If Perplexity had taken the Jasper bet — assume GPT-4 economics forever — they would have died inside a year.
Instead they built dynamic model routing in early 2024. Simple queries got routed to a smaller, cheaper model (initially Claude Haiku 3, later Sonnet-3.5 once Anthropic released it). Complex queries got routed to the frontier. They ran their own evals continuously to re-score which queries needed which tier, and as cheaper models got better, they kept moving the line.
Their cost-per-query fell ~6× between 2024 and 2026, faster than the raw model price curve, because they kept downgrading the routing tier as cheaper models crossed the quality bar. They survived the price war by assuming it would happen and pricing the product ($20/month Pro) against where the cost curve would be in twelve months, not where it was today.
Lesson: if your product is alive in three years, the model underneath it has been swapped two or three times. Build the swap in from the start.
Bet 3: Cursor — "Haiku-class models are free, use them everywhere"
Cursor (the AI code editor) made the most aggressive bet of the three. Starting in late 2023, they architected the product around a tiered model strategy:
- Frontier model (Claude Sonnet, then Opus) for the high-stakes "compose / refactor / explain this codebase" actions where quality matters and the user is waiting on the answer.
- Cheap model (Haiku, plus their own fine-tuned tab-complete model) for the invisible actions — autocomplete predictions, syntax fixups, format suggestions, draft summaries — that run hundreds of times per session per user.
The frontier calls cost real money. The cheap calls were essentially free per call but ran at ~100× the volume of the frontier calls. Total spend was dominated by the cheap tier, but the cheap tier scaled roughly with the Haiku price curve — which collapsed in 2024-2025.
Cursor printed money. The "Tab" autocomplete feature became their signature, and they could afford to run it constantly because each call cost a fraction of a cent. By the time the rest of the IDE market realized, Cursor had a year of compounding usage data, their own fine-tuned models, and the developer mindshare. They hit ~$300M ARR by late 2025 with a small team.
Lesson: features that were "too expensive to run constantly" in 2023 became free in 2026. Whoever shipped them first owned the category. The cost curve isn't just a margin question — it unlocks new product categories every twelve months.
What ties the three together
The three companies had access to the same models, the same APIs, the same prompt-engineering literature. The only difference was their assumption about the price curve:
- Jasper assumed it was flat. Wrong.
- Perplexity assumed it would fall and built routing. Right.
- Cursor assumed it would fall and unlock new product shapes. Most right.
The next step is a multiple-choice on this exact pattern, applied to four hypothetical 2023 startups. Practice spotting which cost-model assumption survives.
The 2023-2026 model price war — and why your cost model is already obsolete — step 2 of 8
Three bets and how they aged
Theory only sticks if you can see it in actual P&Ls. Three companies, three different cost-model bets in 2023, three very different outcomes by 2026.
Bet 1: Jasper — "GPT-4 unit economics are the unit economics"
Jasper was the marketing-copy company. By 2022 they were doing ~$80M ARR generating ad copy and blog posts on top of OpenAI. In October 2022 they raised $125M at a $1.5B valuation, the textbook "AI-native" win.
Their cost model assumed GPT-3.5 and GPT-4 prices were permanent floors. They priced subscriptions around them. They hired around them. They built features around them.
Then the prices fell, and the floor became a ceiling. Three things happened in parallel:
- Their margins on individual API calls became absurd as costs fell — for about six months — until competitors realized and undercut on price.
- New entrants (Copy.ai, then a wave of vertical AI writers) shipped the same feature on cheaper models and charged $20/month where Jasper charged $99.
- ChatGPT released a free consumer tier, then Anthropic and Google followed, then everyone shipped "Write copy with AI" as a feature inside the tools the customer already paid for (Notion, HubSpot, Mailchimp).
Jasper laid off ~30% of staff in 2023, restructured, and the path to IPO that the $1.5B valuation implied has not materialized. The cost model they priced against didn't exist twelve months later. They survived as a smaller company, but they did not capture the upside their cap table assumed.
Lesson: pricing your product against today's model costs is pricing against a melting ice cube. Anyone who shipped the same feature at Haiku prices six months later took the customer.
Bet 2: Perplexity — "route to the cheapest model that meets the bar"
Perplexity launched as an "answer engine" in late 2022. They had a fundamentally harder cost problem than Jasper: every user query required (a) a search step, (b) a synthesis call against a frontier model, (c) a re-ranking pass. Three LLM calls per query, often more for complex questions, at GPT-4 prices.
If Perplexity had taken the Jasper bet — assume GPT-4 economics forever — they would have died inside a year.
Instead they built dynamic model routing in early 2024. Simple queries got routed to a smaller, cheaper model (initially Claude Haiku 3, later Sonnet-3.5 once Anthropic released it). Complex queries got routed to the frontier. They ran their own evals continuously to re-score which queries needed which tier, and as cheaper models got better, they kept moving the line.
Their cost-per-query fell ~6× between 2024 and 2026, faster than the raw model price curve, because they kept downgrading the routing tier as cheaper models crossed the quality bar. They survived the price war by assuming it would happen and pricing the product ($20/month Pro) against where the cost curve would be in twelve months, not where it was today.
Lesson: if your product is alive in three years, the model underneath it has been swapped two or three times. Build the swap in from the start.
Bet 3: Cursor — "Haiku-class models are free, use them everywhere"
Cursor (the AI code editor) made the most aggressive bet of the three. Starting in late 2023, they architected the product around a tiered model strategy:
- Frontier model (Claude Sonnet, then Opus) for the high-stakes "compose / refactor / explain this codebase" actions where quality matters and the user is waiting on the answer.
- Cheap model (Haiku, plus their own fine-tuned tab-complete model) for the invisible actions — autocomplete predictions, syntax fixups, format suggestions, draft summaries — that run hundreds of times per session per user.
The frontier calls cost real money. The cheap calls were essentially free per call but ran at ~100× the volume of the frontier calls. Total spend was dominated by the cheap tier, but the cheap tier scaled roughly with the Haiku price curve — which collapsed in 2024-2025.
Cursor printed money. The "Tab" autocomplete feature became their signature, and they could afford to run it constantly because each call cost a fraction of a cent. By the time the rest of the IDE market realized, Cursor had a year of compounding usage data, their own fine-tuned models, and the developer mindshare. They hit ~$300M ARR by late 2025 with a small team.
Lesson: features that were "too expensive to run constantly" in 2023 became free in 2026. Whoever shipped them first owned the category. The cost curve isn't just a margin question — it unlocks new product categories every twelve months.
What ties the three together
The three companies had access to the same models, the same APIs, the same prompt-engineering literature. The only difference was their assumption about the price curve:
- Jasper assumed it was flat. Wrong.
- Perplexity assumed it would fall and built routing. Right.
- Cursor assumed it would fall and unlock new product shapes. Most right.
The next step is a multiple-choice on this exact pattern, applied to four hypothetical 2023 startups. Practice spotting which cost-model assumption survives.