Part 1: New plans, cheaper models and meaningful percentage comparisons
For people who use AI regularly, answer quality is no longer the only consideration. The bill matters too: How much does a subscription cost, how far does its allowance stretch, and what does a task cost when an agent spends longer working on it?
A number of things have changed since September. However, this does not justify saying that AI as a whole has become a particular percentage more expensive. New premium subscriptions sit alongside lower API prices. A change in allowances can also make a service effectively more expensive even when the monthly bill stays the same.
We therefore separate subscription prices, technical usage prices and the cost of a useful result. Without that distinction, percentages can be misleading.
What are we comparing?
All API amounts below are in US dollars per million tokens. Tokens are billing units, not a fixed number of questions. This is a selection of publicly documented changes, not a comprehensive market index. Taxes, exchange rates and individual contracts are excluded.
The calculation is: (new value ÷ previous value − 1) × 100. Negative percentages indicate a reduction. For allowances, we also consider the inverse relationship: if the bill remains unchanged, a smaller included quantity costs more per unit.
These percentages must not be averaged into a general “AI inflation rate”. A plan upgrade, a smaller weekly allowance and a cheaper token price describe different things.
| Comparison | Previous or reference value | New value | Change | What it means |
|---|---|---|---|---|
| ChatGPT: Pro 200 compared with Pro 500 | USD 200/month | USD 500/month | +150% | Moving to a different plan, not a demonstrated increase in the existing subscription price [1] |
| Claude Code: weekly allowance after the promotion | 150% of the original allowance | 125% from 14 September | −16.7% allowance | At an unchanged subscription price, the calculated cost per allowance unit rises by 20%; the bill itself does not increase [2] |
| Claude Opus 5 → Opus 5.5: input | USD 5 | USD 4 | −20% | Comparing model generations, not retrospectively changing the older model's price [3] |
| Claude Opus 5 → Opus 5.5: output | USD 25 | USD 20 | −20% | The same billing category, but different models [3] |
| Claude Sonnet 5.5: cache reads | USD 0.20 | USD 0.10 from 7 October | −50% | An actual reduction in one price component for the same model [4] |
Why 16.7% less allowance means 20% more per unit
Anthropic temporarily increased Claude Code's weekly allowance by 50%. From 14 September, a permanent increase of 25% over the original allowance remains. In an illustrative index, 150 units therefore become 125. The five-hour limits were not covered by this promotion. [2]
At the same subscription price, the ratio of calculated costs per allowance unit is 150 ÷ 125 = 1.2: 20% higher than during the promotion. Compared with the original allowance, more is still included.
This helps explain why an unchanged bill can feel like a price increase. It does not prove that every user can complete 16.7% less work. Tasks, model choices and consumption vary; allowance is not a guaranteed quantity of completed results.
A new premium plan is not automatically a price increase
OpenAI lists Pro tiers at USD 100, 200 and 500 per month. Moving from the USD 200 tier to the USD 500 tier increases a subscriber's monthly budget by 150%. [1]
That figure describes the upgrade cost. It does not establish an equivalent increase in the existing plan. To assess that, we would need to compare the actual benefits and conditions before and after a change. The introduction of a higher tier alone is insufficient evidence.
Cheaper tokens do not guarantee cheaper projects
A lower unit price helps, but it is only part of the bill. An agent may read files repeatedly, invoke additional tools or revise its solution several times. Software development also requires checks, tests and sometimes rework.
Conversely, a more expensive model may be more economical on a particular task if it needs substantially fewer attempts. That is something to measure in a real workflow, not a general promise.
A deliberately simplified example illustrates the difference: reducing the price per million output tokens from USD 25 to USD 20 lowers that unit price by 20%. If the required output quantity rises by 50%, the calculated cost becomes USD 30 instead of USD 25—a 20% increase. This example deliberately excludes input, caching and tools.
Our view at TNSystem
We want to know what a checked result costs and whether the work can be planned reliably—not simply which provider displays the smallest number beside a plan.
A useful comparison needs the same task, quality requirement and total effort. Waiting, checking and corrections belong in that calculation. Temporary promotional benefits should be recorded separately from the ongoing offer.
Our conclusion: AI can become more expensive for an individual user. The figures examined here do not establish a blanket price increase across the market; genuine reductions exist too. The key question is what we receive for our money and how predictable that relationship remains.
Part 2 examines user reactions and why usage limits cause particular frustration.
Sources
[1] OpenAI: current ChatGPT pricing, accessed 11 October 2026
Read source ↗
[2] Anthropic: Claude Code promotion and permanent weekly allowance
Read source ↗
[3] Anthropic: Opus 5.5 pricing compared with Opus 5, 22 September 2026
Read source ↗
[4] Anthropic: Haiku 5.5 and the Sonnet cache price reduction, 7 October 2026
Read source ↗
Continue with Part 2
Usage limits and user reactions
Read source ↗