The short version
The monthly price of an AI coding tool is becoming a terrible proxy for how much compute you are actually consuming.
A $20 subscription can include hundreds of dollars of API-equivalent usage. Heavy $100 to $200 plans can represent thousands. In one unusually intensive real-world Claude Max 20x workload, $200 of subscription access corresponded to more than $26,000 at API list prices.
That does not mean every user gets $26,000 of value. It means AI software pricing has split into two very different markets: flat subscriptions with hidden usage allowances, and metered APIs where every token has a visible cost.
For founders, IT teams and finance leaders, that distinction matters.
What the subscriptions can be worth
Claude Code
OpenAI Codex
Grok Bot (via Cursor)
For Grok Bot, we use the Cursor plans because that's where we have the cleanest actual Bot-meter observations.
Gemini / Google Antigravity
GitHub Copilot
Copilot is the outlier here. Its multiples sit close to 1:1, a reminder that not every subscription is built to hide runaway compute the way Claude Code, Codex and Grok Bot plans are.
Important: these are API-equivalent estimates, not credits that vendors give you. Actual value changes with model choice, caching, context length, output length, workload and how much of the allowance you actually use.
Why these numbers are so much higher than the sticker price
Traditional SaaS is easy to understand. If a seat costs $20, the vendor is broadly selling you a $20 seat.
AI software is different.
A coding agent may read a repository repeatedly, hold large contexts, call tools, run sub-agents and generate thousands of tokens while you experience it as one task. Subscription providers can bundle this compute aggressively because not every user hits the limits, caching can reduce their serving cost, and usage is controlled with five-hour and weekly allowances.
That creates a strange outcome: the same workload can be dramatically cheaper inside a subscription than through an API.
Claude Code
Anthropic currently prices Claude Pro at $20 per month, Max 5x at $100 and Max 20x at $200. Claude Code is included. Anthropic says Max 5x users can typically send roughly 50 to 200 Claude Code prompts every five hours, while Max 20x users can send roughly 200 to 800, with other limits also applying.
Independent usage measurements show how large the gap can become. A separate audit of one extremely heavy Claude Max 20x seat measured 33.3 billion input tokens in a month and priced the same workload at $26,187 using Opus 5 API list rates.
The lesson is not that Claude Max is worth $26,187 to everyone. It is that the ceiling is enormous compared with the subscription price.
Codex
OpenAI currently includes Codex in ChatGPT Plus and Pro. Plus is $20 per month, Pro has $100 5x and $200 20x usage tiers. OpenAI's current Codex pricing page shows that, depending on model and workload, Plus can allow approximately 10 to 100 GPT-5.6 Sol local messages per five-hour period, Pro 5x approximately 50 to 500 and Pro 20x approximately 200 to 2,000. Weekly limits can also apply.
One current wrinkle: OpenAI has temporarily paused new sign-ups and upgrades to the $200 Pro 20x plan, although existing subscriptions continue.
Grok Bot
Grok Bot is particularly interesting because it is bundled into Cursor Pro, which starts at $20 per month, and can also be accessed by linking qualifying SuperGrok subscriptions.
Cursor says Grok Bot gets its own weekly included usage pool. It resets weekly, and once that allowance is exhausted, additional usage can spill into on-demand billing if enabled.
Cursor does not publish the numeric size of that weekly Grok Bot pool, which is why the Cursor Pro, Pro+ and Ultra figures above are the most reliable read on its real value.
Gemini and Google Antigravity
Google's AI Pro, AI Ultra 5x and AI Ultra 20x plans show a more moderate curve than Claude Code or Codex at the entry tier, but the 20x plan still lands near 53x, in the same range as the highest multiples on this list. That consistency across tiers suggests Google is metering its allowances more predictably than some competitors.
GitHub Copilot
Copilot's value multiples are the lowest of any tool here, from 1.5x on Pro to 2x on Max. That is not necessarily a knock on Copilot. It suggests GitHub is pricing closer to actual compute cost rather than bundling a large hidden allowance, which matters if you are comparing tools on value density versus predictability.
The bigger story: AI spend is becoming invisible
This is the part companies should pay attention to.
Imagine a 30-person company where developers independently buy Claude, ChatGPT, Cursor, Gemini and Copilot. The finance system sees five neat recurring charges. It does not see which tool is being used, whether two tools are doing the same job, who owns each subscription, what usage sits behind the flat fee, or whether the company is about to start paying API overages on top.
The sticker price is increasingly the least interesting number.
AI spend now comes in several forms at once: seats, usage-based APIs, credits, token bundles, agent allowances, on-demand overages and subscriptions with opaque weekly caps. Two employees can pay exactly the same subscription price while creating radically different amounts of compute consumption.
That is why "how much are we spending on AI?" is becoming surprisingly difficult to answer.
What teams should track instead
Instead of tracking only the invoice amount, companies should know:
- Which AI tools are active across the company.
- Who owns and uses each subscription.
- Whether tools overlap in capability.
- The fixed subscription cost versus metered or on-demand spend.
- Renewal dates and overage exposure.
- Whether expensive seats are actually being used.
- Where one team is paying for something another team already has.
This is exactly the problem we are building ELI to solve.
ELI gives companies one place to see their software and AI stack, understand spend, spot overlapping tools and keep track of renewals, instead of reconstructing it manually from cards, invoices and inboxes.
If your company has started accumulating Claude, ChatGPT, Cursor, Gemini, Copilot and a dozen other AI subscriptions, you can see what is actually in your stack at eli.work.
Methodology and caveats
There is no perfectly clean apples-to-apples conversion between a subscription allowance and an API bill. Vendors meter subscriptions differently, model mixes change, caching changes effective API cost, and user workloads vary dramatically.
For that reason, the figures above draw on three types of evidence:
- Official: published subscription prices, plan ratios and usage guidance from Anthropic, OpenAI, Cursor, Google and GitHub.
- Observed: API-equivalent costs reconstructed from user logs or usage dashboards.
- Derived: simple proportional estimates based on official plan multipliers or observed weekly allowance consumption.
The goal is not to claim that every subscriber receives a fixed dollar amount of API credit. The goal is to show the order of magnitude, and why companies need to look beyond the subscription price when managing AI spend.
