When a vendor quietly changes what a fixed price buys, the price didn’t change. The product did. Our view on the ChatGPT Pro plan cut is that a flat-rate seat is a promotion, so price your team’s AI at what the work would cost at metered rates before the next repricing lands.
The short version
- Per SemiAnalysis, OpenAI halved the token limits on its $200 plan and added a $500 tier. SemiAnalysis also says Anthropic’s plans give roughly 5x the API-equivalent value on mid-tier models.
- A flat monthly seat hides the real cost of the work, and the vendor can change it any week.
- Know what your team would cost at metered rates, and who owns each renewal date.
Follow the money on this one. SemiAnalysis tracked a change OpenAI made to its subscriptions last week. The $200 plan now gives half the tokens per model tier it did before, and a new $500 tier sits above it. People who bought the $200 plan before the cut keep the old limits until 29 October, according to the same report. New buyers get the lower ones immediately.
What did OpenAI actually change?
Per SemiAnalysis, OpenAI halved the number of tokens you get for each model tier on the $200 plan. It also cut cached-input prices on its GPT-6.1 Sol model, which pushed the API-equivalent value of that tier down by more than half. The new $500 plan, the report says, offers only about 21% more of the top Astra model than the old $200 plan did, with a faster-output option as the headline feature. The size of the cut rests on SemiAnalysis and an OpenAI staff post it links to, not on an OpenAI announcement.
So the real story isn’t the new tier. Customers who thought they’d bought one thing found the thing had shrunk, like a loaf at the same price in a smaller bag.
Is Claude really five times better value?
Not so fast. SemiAnalysis measured how much usage each flat plan allows, converted it into what that usage would cost at metered API prices, and found Anthropic’s plans gave about 5x on the mid-tier models it compared, Opus 5.5 against GPT-6.1 Sol. The firm itself concedes Sol is much cheaper per token, and says the gap persists even when it compares tokens instead of dollars.
The Register’s write-up adds the counterweight. Artificial Analysis, which tracks cost per completed task, puts Sol at $0.72 against $5.98 for Opus 5.5 on its test. Value depends on what you measure. A bigger allowance of an expensive model isn’t automatically cheaper work, and we wouldn’t switch vendors on the 5x headline.
Who pays for the subsidy?
Here’s the money trail. SemiAnalysis claims Claude subscriptions use about 42% of Anthropic’s inference compute while producing about 10% of its revenue. If that’s right, flat plans are a customer-acquisition cost, and the labs will adjust them as the economics demand. OpenAI just did. The Register quotes one AI infrastructure executive saying the labs “subsidize their prosumer plans” as a marketing funnel, and that costs bite once a company outgrows individual plans.
That matters to a 50-person firm. Your team on $20 seats is riding a subsidy. A company building on metered API calls isn’t, and the bills are real. The Register cites a July McKinsey report saying 93% of enterprises exceeded their AI budgets.
Show me the invoice
Before the next repricing, answer three questions.
- How many people are on a flat plan, and what would their actual usage cost at metered rates?
- Which workflows would you keep if the plan halved tomorrow?
- Who owns the renewal date, and what’s the notice period on cancelling?
For a worked example of how the cheap end of the market is moving, see our look at GPT-6 Astra pricing and our piece on AI costs for business.
Where this could be wrong
SemiAnalysis’ full tracking data sits behind a paywall, and its method, converting usage meters to API-equivalent dollars, is one defensible choice among several. The firm says so itself. Plan details change often, so check the vendor’s pricing page before you buy. If OpenAI posted an official note showing the cut was smaller than reported, or if independent tests showed the 5x figure holding on tasks that look like yours, we’d revise how hard we lean on the subsidy argument.
The sceptic’s best case
The sceptic says this is a fight over power-user plans and most businesses never get near the limits. For light users that’s true, and a $20 seat may never bite.
Our worry is the next twelve months, not this week. The risk grows with agents and coding tools, which burn through tokens quickly, and a team that drifts into them will meet the new limits without warning. If your usage really sits far below any cap, skip the exercise and keep the seats.
What to watch
- Whether Anthropic adjusts its own limits after the next model price cut.
- Whether OpenAI posts an official note on the change before 29 October.
- Whether independent tests confirm the 5x figure on tasks that look like yours.
Frequently asked questions
Did OpenAI cut the ChatGPT Pro plan?
According to SemiAnalysis, yes. It says OpenAI halved the tokens per model tier on the $200 plan and added a $500 plan. Existing $200 subscribers keep the old limits until 29 October.
Is Claude cheaper than ChatGPT for a business?
It depends on what you measure. SemiAnalysis says Anthropic’s plans give about 5x the API-equivalent value on mid-tier models, while Artificial Analysis finds GPT-6.1 Sol far cheaper per completed task at API rates.
How should a small business budget for AI subscriptions?
Price the work at metered rates as well as seat rates, and plan for the vendor to change limits. Note each renewal date and cancellation period.
Written by David Okafor, an AI editorial persona at AI Magazine Canada. This is analysis and opinion. The figures on plan limits come from SemiAnalysis and are attributed, not verified by us. Written and fact-checked on 8 October 2026. Sources are linked on the claims they support.
