What Your $20 AI Subscription Actually Costs the Company (2026)

A $200 ChatGPT Pro subscription can cost OpenAI $14,000 in compute if a single user maximizes it. That finding from SemiAnalysis is striking on its own. But it becomes more interesting when you put it next to what every other major AI company is doing at the same price points — and realize that the $20 AI subscription, across every platform, is one of the most heavily subsidized products in the history of consumer software.

Here is a complete breakdown of what you pay, what you get, and what it actually costs the company to serve you.

Why AI Subscriptions Are Structurally Unusual

Most SaaS subscriptions are profitable or near-profitable because the marginal cost of serving an additional user is near zero — another Spotify user streaming costs fractions of a cent. AI is different. Every query to a frontier model involves significant GPU compute. The more capable the model and the longer the context, the more expensive each inference.

This means AI companies have made a bet: subsidize users heavily now to build habits and market share, generate revenue from enterprises and API customers, and eventually either raise consumer prices or let usage naturally throttle as users hit limits.

The SemiAnalysis $14,000 figure is the extreme case — a user who hammers o1 Pro reasoning queries all day, generates hours of Sora video, runs research agents constantly. But it illustrates the math underneath every subscription tier.

The $20 Tier: Who Gives What

All four major AI platforms land at approximately $20/month for their entry paid tier. They are not selling the same thing.

*Fable 5 currently suspended by US export control order as of June 12, 2026.

ChatGPT Plus — $20/month

What you get: GPT-4o access without rate limits, DALL-E 3 image generation, web browsing, code interpreter, limited access to o1 (not o1 Pro), GPT-5.5 access limited.

What you don't get: The models that matter most — o1 Pro, GPT-5.5 at full capability, Sora video generation — are all behind the $200 Pro tier. At $20, ChatGPT Plus is essentially "GPT-4o with extras." That is still genuinely useful, but it is not the product OpenAI showcases in its demos.

Estimated cost to OpenAI per average user: $10–40/month depending on usage. The $20 tier is where OpenAI is roughly breaking even or slightly profitable on typical users.

Claude Pro — $20/month

What you get: Claude Opus 4.8 as the primary model, limited Fable 5 credits (when restored), Claude Code included, 5x more usage than the free tier, priority access during peak hours.

The Fable 5 situation: Before June 12, Pro subscribers had free access to Fable 5 through June 22 as a launch promotion. The model was suspended by US government export control order on June 12. When it returns, Fable 5 usage will require credits at $10/M input tokens and $50/M output tokens — some credits bundled into Pro, more into Max tiers. See our full Fable 5 return tracker for current status.

What makes it stand out: Claude Code — the agentic coding tool — is included at the Pro tier. No other $20 plan includes an equivalent. For developers, this is the differentiating feature.

Estimated cost to Anthropic per average user: $25–60/month. Pro is a loss-leader. The Max tiers at $100–200/month are where Anthropic starts recovering costs on heavy users.

Gemini Advanced — $19.99/month

What you get: Gemini 2.0 Pro access (Google's best currently publicly available model), deep research mode, integration across Google Workspace (Docs, Sheets, Gmail), and 2TB of Google One storage that replaces or extends your existing storage plan.

The bundle angle: This is the most differentiated product in the $20 tier because it bundles storage. If you already pay Google for 200GB or 2TB of Drive storage ($3–10/month), Gemini Advanced partially offsets that cost. The effective AI-only price becomes $10–17/month depending on your existing storage spend.

What you don't get: Google has not released Gemini Ultra publicly at the $20 tier — the most capable models remain in limited access or enterprise channels.

Estimated cost to Google per average user: Hard to isolate. Google's infrastructure scale likely makes per-query costs lower than competitors, but Gemini Pro inference is still expensive at the frontier tier.

Copilot Pro — $20/month

What you get: Priority access to GPT-4o, DALL-E 3, real-time image creation in Bing, integration into Word, Excel, PowerPoint, and Outlook if you have a Microsoft 365 subscription.

The catch: The Office integration only works if you also have Microsoft 365 Personal or Family ($7–10/month). Without that, Copilot Pro is a web/mobile interface to GPT-4o — similar to ChatGPT Plus — without the ecosystem advantage that makes it worth considering over ChatGPT directly.

Best for: Users already in the Microsoft 365 ecosystem who want AI natively in their Office tools.

Perplexity Pro — $20/month

What you get: Unlimited Pro searches powered by the best available model (rotates between Claude Opus, GPT-4o, Gemini, and Perplexity's own models), $5/month in API credits, image generation, file analysis.

The unusual value proposition: You are not subscribing to one AI — you are subscribing to a search product that routes queries to whichever frontier model performs best for that task type. During the Fable 5 suspension, Perplexity Pro users can still access Claude Opus 4.8 and GPT-5.5 through the routing layer.

The $200 Tier: Where the Real Losses Are

ChatGPT Pro — $200/month and the $14,000 Problem

OpenAI's Pro tier unlocks the features it actually wants to sell: o1 Pro mode (which uses substantially more compute than standard o1 for each query), GPT-5.5 at full capability, Sora video generation, unlimited deep research, and advanced voice.

SemiAnalysis, a semiconductor and AI infrastructure research firm, modeled what happens when a single user maximizes all of these features simultaneously:

• o1 Pro reasoning queries at maximum rate: each query uses extended chain-of-thought reasoning that can consume 10–50x the tokens of a standard GPT-4o query
• Sora video generation at high resolution and long duration: video generation is among the most compute-intensive tasks in the consumer AI stack
• Deep research running long multi-step research agents: each research session can consume hours of agent execution

The result: $14,000 in estimated compute cost for what OpenAI charges $200. The average Pro user does not come close to this — but the tail of heavy users creates enormous per-user losses, and the average usage patterns mean OpenAI is losing money on most Pro subscribers at current compute costs.

This is not sustainable indefinitely. It represents a land-grab strategy: establish the Pro tier as the standard for what serious AI use looks like, acquire heavy users before competitors do, and either reduce compute costs through efficiency gains or raise prices later.

Claude Max — $100/month and $200/month

Anthropic offers two Max tiers:

Max at $100/month: 5x more usage than Pro. Includes Claude Code, priority access, and bundled Fable 5 credits when the model is restored.

Max at $200/month: 20x more usage than Pro. Same inclusions, but at a scale designed for developers and power users who are running Claude for hours per day.

The Fable 5 credit inclusion is the key differentiator for the Max tiers. When Fable 5 returns, at $10/M input and $50/M output tokens, a power user running long-context sessions could spend $50–200+ in credits per day. Max tiers bundle a meaningful credit allowance to offset this. Exact credit amounts per tier have not been fully published.

Estimated cost to Anthropic: Fable 5 inference is among the most expensive Anthropic runs. A Max user running Fable 5 at heavy usage could plausibly cost Anthropic $500–800+/month to serve. The $200 price point is a loss at heavy usage, though less extreme than OpenAI's $14,000 scenario.

For the full Claude pricing breakdown including Claude Code specifics, see our Claude Code pricing guide.

What You Actually Get Per Dollar: The Real Comparison

Rather than comparing sticker prices, here is what matters by use case:

For general AI chat and research

Best value: Gemini Advanced ($19.99) — competitive model, 2TB storage bundle, deep research. Effective AI-only cost is $10–17/month if it replaces existing storage.

Runner-up: Perplexity Pro ($20) — multi-model routing means you get the best available model for each query type without vendor lock-in.

For coding and software development

Best value: Claude Pro ($20) — Claude Code is included and Claude Opus 4.8 outperforms GPT-4o on most agentic coding benchmarks. When Fable 5 returns, the included credits make it even more compelling.

Runner-up: GitHub Copilot ($10–19/month) — purpose-built for in-IDE use, integrates with VS Code and JetBrains, now routes to multiple models.

For creative work (writing, images, video)

Best value at $20: ChatGPT Plus — DALL-E 3 integration, voice, and GPT-4o's strong creative output.

If you need video: ChatGPT Pro ($200) — Sora is the only frontier video generation model in a consumer subscription. No other plan at any price point currently includes it.

For heavy professional use

Best value at $200: This is genuinely context-dependent. ChatGPT Pro gives you the most raw capability (o1 Pro, GPT-5.5, Sora) but the models are expensive enough that OpenAI rate-limits even Pro users in practice during peak hours. Claude Max ($200) gives you 20x usage with more predictable access patterns.

The Hidden Math: Why These Prices Are All Temporary

Every AI company is betting that one or more of the following will happen before the subsidy model collapses:

1. Inference costs fall dramatically. GPU efficiency, model quantization, and architectural improvements (mixture-of-experts, speculative decoding) are reducing the cost per token at roughly 2–3x per year. A query that costs $0.10 today may cost $0.03 in 18 months. If this holds, the $14,000 theoretical max for ChatGPT Pro might drop to $5,000 — still a loss, but a more survivable one.

2. Enterprise revenue subsidizes consumers. Claude Enterprise, ChatGPT Enterprise, and Gemini for Workspace are all priced at $30–60+/user/month with usage limits that make them profitable. The consumer products are marketing channels for the enterprise products. As long as enterprise adoption grows, consumer losses are fundable.

3. Usage habits establish lock-in before price increases. The expectation is that by the time prices need to rise, users will be dependent enough on their preferred AI tool that they accept higher prices. This is the same playbook as streaming services in 2015–2020.

4. API and developer revenue grows. The OpenAI API, Anthropic API, and Gemini API are all priced to be profitable at scale. As enterprise API spend has surged, this revenue stream increasingly cross-subsidizes consumer subscriptions.

The implication: current subscription prices are promotional pricing, not long-term sustainable pricing. The $20 tier will either rise, get capped more aggressively, or have its best features migrated to higher tiers.

The Fable 5 Wildcard

The US export control suspension of Fable 5 on June 12 adds a new dimension to the Claude value calculation. Anthropic's most capable model is temporarily offline, which affects:

• Claude Pro: The free Fable 5 trial period (to June 22) is running out while the model is unavailable
• Claude Max: The Fable 5 credit bundle is effectively paused
• The competitive position: GLM-5.2 from Zhipu AI is now beating Fable 5 on BridgeBench Reasoning at 1/10th the cost, and Kimi K2.7-Code is available open-source — both affecting how Claude subscribers evaluate their options

The rational move for existing Claude subscribers right now: use the rate limit reset Anthropic issued on June 13 to maximize Opus 4.8 usage, and wait for clarity on the June 22 pricing transition before deciding whether to upgrade or downgrade tiers. See our full Fable 5 status tracker.

Quick Decision Guide

The honest summary: at $20, you are buying access to GPT-4o or equivalent — a genuinely useful but not bleeding-edge model. At $200, you are buying access to the actual frontier — and the company is losing money to give it to you. Neither is priced to last forever.

For more on how AI costs affect enterprise budgets, see when AI token spend stops looking like a SaaS line item. For developer-specific Claude pricing including API rates, see the Claude Code pricing guide.