SemiAnalysis: ChatGPT Pro Costs OpenAI $14,000 at Full Use — The Agentic Subsidy Exposed

A SemiAnalysis researcher who maxed out a $200 ChatGPT Pro subscription with long-horizon coding and agent tasks consumed compute that would have cost $14,000 at standard API pricing. That is a 70x gap between what a power user pays and what OpenAI spends to serve them.

The research firm published its findings in June 2026 after buying one of every subscription tier from both OpenAI and Anthropic, running agentic workflows until weekly limits were exhausted, and mapping actual token consumption to public API rates. The numbers are not close. The subscription pricing model that drove AI adoption was, by OpenAI's own head of ChatGPT's admission, "accidental" — designed for casual chat users, now being stress-tested by engineers running autonomous agents all day.

The SemiAnalysis Methodology

SemiAnalysis purchased every available tier from OpenAI and Anthropic: ChatGPT Free, ChatGPT Plus ($20/month), ChatGPT Pro 5x ($100/month), ChatGPT Pro 20x ($200/month), Claude Pro ($20/month), Claude Max 5x ($100/month), and Claude Max 20x ($200/month).

For each tier, researchers ran long-horizon coding tasks and agent-style workflows — the kind of work that software engineers, data scientists, and DevOps teams actually do with these tools — until the weekly usage limits were fully exhausted. They then calculated the token volume consumed and converted it to cost using publicly posted API pricing for the equivalent models (GPT-4o, o3, claude-sonnet-4-6, claude-opus-4-8).

This is the correct methodology for understanding subscription economics. Surveying average users gives you the median case. Maxing the limits gives you the upper bound — and the upper bound is what the flat-rate pricing model has to sustain.

The Numbers: $200 In, $14,000 Out

The 70x multiplier on ChatGPT Pro is the headline. A user who genuinely maxes their subscription every week for four weeks gives OpenAI a compute bill 70 times larger than the subscription revenue. Anthropic's equivalent is 40x — still catastrophic at scale, but significantly better insulated than OpenAI.

The difference between the two companies on this metric matters. Anthropic's inference costs per token are lower relative to its subscription caps, and the Claude Max architecture allows Anthropic to sustain heavier use before reaching the same loss threshold. Both companies lose money on heavy users. OpenAI loses money faster.

The Agentic Token Explosion

The reason these numbers are so extreme is not that the models got more expensive. It is that how people use them changed completely.

A standard ChatGPT conversation in 2023 consumed roughly 500-2,000 tokens per exchange. A user asking questions, getting answers, following up — the token volume was bounded by how fast a human could type and read.

An agentic task in 2026 is different. The typical agent job consumes 96,000 tokens before producing a single answer — more text than the entire novel The Great Gatsby. An autonomous coding agent reviewing a codebase, writing tests, running them, reading error output, revising, and committing can consume hundreds of thousands of tokens in a single session. SemiAnalysis estimates agentic workflows use up to 1,000 times more tokens than a standard prompt-and-response exchange.

OpenAI priced ChatGPT Pro for the prompt-and-response world. The product is now used primarily for agent workflows by the engineers who pay $200/month. The pricing model has not caught up.

The Utilization Breakpoints: When the Losses Start

SemiAnalysis calculated the utilization rate at which each tier crosses from profit to loss:

OpenAI:

• ChatGPT Plus and Pro 5x: loses money above 11.4% utilization
• Top-tier Pro 20x: reaches zero gross margin at 5.7% utilization

Anthropic:

• Claude Pro and Max 5x: breaks even at approximately 20% utilization
• Top-tier Max 20x: reaches zero gross margin at approximately 10% utilization

These numbers mean OpenAI is structurally more exposed than Anthropic. An OpenAI Pro 20x subscriber who uses the product at even moderate intensity — not maxing it, just using it daily for real work — crosses the loss threshold at 5.7% of theoretical maximum capacity. That is not a power user edge case. That is a working developer using an AI coding assistant for a few hours a week.

The business model works because most subscribers do not use their plan anywhere close to its limits. The casual user who asks ChatGPT a few questions a day subsidises the engineer running agents all night. This cross-subsidy has funded AI adoption at scale. The SemiAnalysis data shows how thin the margin for that cross-subsidy actually is.

OpenAI Already Knows: "Accidental" Pricing and the $100 Pro Lite Move

OpenAI's head of ChatGPT described the current pricing model as "accidental" in early 2026 — designed when the product was a chatbot, not an agent platform. The statement is an unusually candid acknowledgment that the economics are wrong.

The company has already started restructuring. On April 9, 2026, OpenAI launched ChatGPT Pro Lite at $100/month — a new tier slotting between Plus ($20) and Pro ($200). The tier was discovered in ChatGPT's web app code before launch, confirming that OpenAI was actively segmenting the market to separate moderate users (who can be profitable at $100) from heavy users (who remain a loss at $200 but potentially profitable at a higher future price).

Read alongside the SemiAnalysis findings: the $100 Pro Lite tier is OpenAI's first structural response to the subsidy problem. The next move is almost certainly a price increase on Pro 20x or a hard token cap that converts the unlimited plan into a capped one. OpenAI has 900 million weekly active users and 50 million paying subscribers. Even a small fraction of those subscribers running agentic workflows at scale creates a compute bill that the current pricing cannot cover.

This connects directly to why OpenAI is still losing $14 billion per year despite $2 billion in monthly revenue. See our OpenAI $122B funding and $852B valuation post for the full financial picture.

What This Means for Developers: Subscription vs. API Math

For developers and engineering teams, the SemiAnalysis data has an immediate practical implication: if you are a heavy user, you are currently getting a deal that cannot last.

The arbitrage works like this. ChatGPT Pro at $200/month gives you access to GPT-4o and o3 for tasks that would cost $14,000 at API rates. Claude Max at $200/month gives you access to claude-opus-4-8 and claude-sonnet-4-6 for tasks that would cost $8,000 at API rates. As a developer running agentic coding workflows, the subscription is a 40-70x cost advantage over using the API directly.

That advantage exists because OpenAI and Anthropic are subsidising it. When the subsidy ends — through price increases, usage caps, or tiered token limits — the economics flip. The switch GitHub Copilot made in June 2026 from flat-rate to usage-based AI Credits billing is exactly what happens when a subscription business discovers its heavy users are destroying unit economics. See our GitHub Copilot billing change post for the detailed breakdown of what that transition looks like.

Current developer guidance:

• If you are running agentic workflows heavily, extract maximum value from the current flat-rate plans now
• Build your cost models assuming subscription prices double or token caps appear within 12-18 months
• For production workloads, use the API rather than subscriptions — it scales predictably and will not have its economics changed overnight

For current model API pricing across OpenAI, Anthropic, and alternatives, see the LLM API Pricing Tracker.

Our Analysis: The Flat-Rate Bet That Worked Until Agents Arrived

The $20 ChatGPT Plus plan launched in February 2023 was a brilliant customer acquisition move. Flat-rate pricing removed purchase anxiety, drove mass adoption, and generated the subscriber base that justified the company's valuation. OpenAI and Anthropic both bet that casual users would subsidize the heavy ones, and the bet paid off through 2024.

The problem is that the product category changed underneath the pricing model. In 2023, an AI subscription competed with Google Search — a tool for quick answers. In 2026, it competes with hiring a software engineer — a tool for extended autonomous work. The token consumption profile of those two use cases is not in the same order of magnitude.

SemiAnalysis's 96,000-token-per-agent-job figure is the crux of the issue. That number is not going down. As models get better at autonomous coding, research, and operations tasks, the jobs users give them get longer and more complex. The average token consumption per paying subscriber is a ratchet that only turns one way.

Both OpenAI and Anthropic know this. The $100 Pro Lite tier, the GitHub Copilot credit switch, the Anthropic Max usage caps — these are all early signals of an industry-wide repricing that will happen gradually and then suddenly. The SemiAnalysis report just made the math visible.

Key Takeaways

• ChatGPT Pro ($200/month) costs OpenAI $14,000 in compute at full use: SemiAnalysis bought every tier and maxed weekly usage with agentic coding tasks — the 70x gap between subscription revenue and compute cost is the clearest measure of the subsidy problem
• Claude Max ($200/month) costs Anthropic $8,000 at full use: Anthropic is better insulated at 40x (vs OpenAI at 70x) because its inference costs per token are lower relative to subscription caps
• OpenAI loses money at 5.7% utilization on top-tier plans: the 11.4% breakeven for Plus/Pro 5x means a working developer using AI daily crosses into loss territory well before hitting any usage limit
• Agentic tasks use 1,000x more tokens than standard prompts: a typical agent job consumes 96,000 tokens — more than The Great Gatsby — before producing a single answer; this is what the flat-rate model was never priced for
• OpenAI already responded: Pro Lite at $100/month launched April 9, 2026; pricing described internally as "accidental"; price increases or token caps on Pro 20x are the logical next move
• Developer implication: heavy users are extracting 40-70x value from subscriptions that cannot remain subsidised indefinitely — use it while it lasts, model your production costs assuming usage-based billing within 12-18 months

Sources

• TechSpot — A $200 ChatGPT subscription could cost OpenAI $14,000 if used to its full potential
• SemiAnalysis — Tokenomics model
• CyberNews — ChatGPT and Claude subscriptions may cost AI firms thousands per user
• Let's Data Science — ChatGPT $200 plan implies up to $14,000 API costs
• Superhuman AI — SemiAnalysis exposes AI subscription economics
• WinBuzzer — OpenAI rethinks ChatGPT pricing, calls current model accidental
• Tom's Hardware — AI costs spike as subscriptions hit pricing wall