Research note 001 / Subscription economics
Claude vs ChatGPT: the subscription value gap
An interactive reading of the public numbers: equal monthly fees, unequal allowances, and why the model and workload change the answer.
An AI subscription sells access at a fixed price. The useful question is how much of your work that price can cover.
For the agentic scenario in SemiAnalysis’s public snapshot, the Opus 5.5 plans carry roughly five and a half times the API-equivalent value of the same-price GPT-6.1 Sol plans. The token gap is smaller. [1]
What does the same monthly fee buy?
Start with the model pair and subscription price. Then compare the reported monthly allowance, its API valuation, or value per dollar of subscription fee. The utilization control adds our own linear scenario to the source’s full-allowance estimates.
Monthly API-equivalent value
The vertical mark is the subscription fee; values above it exceed that fee at API list prices.
Reported estimates at full utilization
| Fee / month | GPT-6.1 Sol | Opus 5.5 |
|---|---|---|
| $20 | $211 1.0B tokens | $1,178 2.9B tokens |
| $100 | $1,055 5.1B tokens | $5,725 14.0B tokens |
| $200 | $2,084 10.2B tokens | $11,726 28.6B tokens |
Value and token labels were published separately and are rounded. Their ratios should be read as approximate.
At the full allowance for a $200/month plan, the reported Opus 5.5 estimate is 5.6× GPT-6.1 Sol’s API value and 2.8× its monthly token count.
More tokens explain only part of the difference
At $200, the public labels report $11,726 and 28.6 billion tokens for Opus 5.5, versus $2,084 and 10.2 billion for GPT-6.1 Sol. Dividing value by token count separates the quantity difference from the API-price yardstick.
The $200 comparison, decomposed
- 1
Monthly token ratio
2.8×28.6B ÷ 10.2B. The quantity of tokens in the reported allowance.
- 2
API value per token ratio
2.0×($11,726 ÷ 28.6B) ÷ ($2,084 ÷ 10.2B). The blended price yardstick.
- 3
Monthly API-value ratio
5.6×$11,726 ÷ $2,084. Quantity ratio × blended-price ratio.
This makes API-equivalent value useful for comparison, but it also makes it sensitive to list pricing. A model with expensive API tokens can look valuable even when its allowance contains fewer tokens. Neither measure tells us how many tasks a user completes successfully.
One $200 plan has several different values
Keep the subscription fixed and change the model or token mix. The source provides two scenarios for Claude Max 20x. Switch between them to see how the same fee produces a different API valuation.
Monthly API value by model and scenario
- Fresh input
- 0.4%
- Cache reads
- 96.6%
- Cache writes
- 2.6%
- Output
- 0.3%
The model with the largest API-value bar can change when the workload changes. A single headline about an entire subscription hides that sensitivity. Match a scenario to your usage before treating its full-allowance value as something you could consume.
The $200 Pro plan before and after September 29
Allowance and API-value changes
| Model | API value / month | Tokens / month |
|---|---|---|
| GPT-6 Astra | $5,734 → $2,897-49.5% | 3.8B → 1.9B-50.0% |
| GPT-6 Sol → GPT-6.1 Sol | $5,904 → $2,084-64.7% | 19.6B → 10.2B-48.0% |
The Sol row also changes model version. Its API-value decline combines changes in token quantity and the pricing applied to those tokens; it should not be interpreted as a pure measure of the quota change. A dated comparison also cannot promise what a future subscription will hold.
The $500 tier against the former $200 tier
| Model family | Former $200 plan | $500 plan |
|---|---|---|
| GPT-6 Astra | $5,734 API value 3.8B tokens | $6,955 API value 4.6B tokens |
| GPT-6.1 SolPreviously GPT-6 Sol | $5,904 API value 19.6B tokens | $5,386 API value 26.3B tokens |
Chinese coding plans add another price range
Different monthly fees make raw dollar bars difficult to compare. The source’s public coding-plan labels are shown below; the per-dollar view divides each API value by its subscription fee.
API value per subscription dollar
| Plan / model | Fee / month | Value / $1 fee |
|---|---|---|
| Z.ai · GLM Coding LiteGLM-5.3 | $18 | 7.7× |
| Z.ai · GLM Coding LiteGLM-5.3-Flash | $18 | 1.3× |
| Z.ai · GLM Coding ProGLM-5.3 | $80 | 10.4× |
| Z.ai · GLM Coding ProGLM-5.3-Flash | $80 | 1.8× |
| Z.ai · GLM Coding MaxGLM-5.3 | $168 | 11.6× |
| Z.ai · GLM Coding MaxGLM-5.3-Flash | $168 | 2.0× |
| MiniMax · Token PlusMiniMax-M3 | $22 | 19.4× |
| MiniMax · Token MaxMiniMax-M3 | $55 | 23.7× |
| MiniMax · Token UltraMiniMax-M3 | $132 | 23.4× |
| Moonshot · Kimi Code PlusKimi K3 | $19 | 2.5× |
| Moonshot · Kimi Code ProKimi K3 | $39 | 5.4× |
| Moonshot · Kimi Code MaxKimi K3 | $99 | 6.5× |
| Moonshot · Kimi Code UltraKimi K3 | $199 | 6.7× |
What these numbers measure
SemiAnalysis describes experiments that isolate token types, observe complete usage-meter steps, estimate rates for each usage window, and blend them into workload scenarios before applying API list prices. Source methodology [1].
The public charts provide final estimates, not the raw experiment logs. We transcribed their labels and calculated ratios, percentage changes, and proportional-use scenarios. We have not independently repeated these experiments.
Full capacity and useful value are different questions
API-equivalent value prices an estimated allowance. It is not spendable API credit. Multiplying it by a utilization percentage assumes the workload mix stays constant and usage scales proportionally. Session limits, model caps, and changes in behavior can prevent that scenario from being reached.
The arithmetic gives a comparison under stated assumptions. It does not measure output quality, task success, response speed, or whether a particular user benefits from upgrading. Our subscription comparison and local benchmark methodology use separate data and methods.
References
Original research by Andrew Megalaa, Max Kan, Dylan Patel.
- SemiAnalysis: Anthropic Subscriptions Offer 5x+ More Value Than OpenAI — original public article and methodology.
- OpenAI pricing — provider plan prices and usage guidance.
- What is the Max plan? — Max tiers and usage limits.
- Claude Code with Pro or Max — subscription use in Claude Code.
- tokenusage.dev methodology — our separate local benchmark approach.
Source published October 5, 2026, at 20:01 UTC; accessed October 12, 2026. Its displayed date can vary by time zone. Dollar values and token counts retain the source labels’ rounding. No paid dashboard data is used.