A $200 Claude plan and a $200 ChatGPT plan are not the same size. SemiAnalysis, which limit-tested subscriptions from Anthropic, OpenAI, Meta, SpaceXAI, MiniMax, Moonshot, Z.ai, Cursor and Cognition, reports that on each lab's mid-tier model, Claude plans offer about 5x the API-equivalent value of ChatGPT plans. The firm's headline example is stark: a $200 Claude Max 20x plan is worth roughly $11,726 of Opus 5.5 usage per month on an agentic workload, while ChatGPT Pro 200 is worth roughly $2,084 of GPT-6.1 Sol.
At the flagship tier the picture flips to near parity. This post summarizes the published figures, explains how the measurement works, covers what the numbers do not prove, and shows what to do with them. Everything attributed here comes from SemiAnalysis's October 2026 article by Andrew Megalaa, Max Kan and Dylan Patel; the full dashboard is paywalled, so we report only what the public portion states.
TL;DR: the questions people are asking
| Question | Short answer |
|---|---|
| Which plan gives more API value? | Claude, by about 5x, on mid-tier models (Opus 5.5 vs GPT-6.1 Sol) |
| What about flagship models? | Close: ChatGPT Pro 200 on Astra about $2,897 vs Claude Max 20x on Fable 5.1 about $2,485 |
| Why is Fable different? | Fable 5.1 can use only up to half of the plan limit |
| What is the best-value Claude combo? | Sonnet 5.5 on Max 20x: about $12,529 agentic, 62.6x the fee |
| Did OpenAI cut the $200 plan? | Yes, API-equivalent value roughly halved; old plans keep old limits until Oct 29 |
| Is the new $500 plan a big upgrade? | Only about 21% more Astra than the old $200 plan, plus Ultrafast speed |
| Are Chinese plans worth it? | Still subsidized; per-dollar value averages a little under the roughly 12x OpenAI gives |
| Are third-party wrappers a good deal? | SemiAnalysis says they are worse than first-party plans |
What "API-equivalent value" measures
Subscriptions do not publish token allowances. They show a usage meter from 0 to 100 percent, usually over a 5-hour and a 7-day window, and sometimes a separate meter for a premium model. Your monthly fee buys credits, and each combination of model and token type spends a different number of credits.
That is why SemiAnalysis argues a plan has no single worth. "It doesn't make sense to say a plan is worth $X in isolation," it writes; you need the full tuple of plan, model and workload. The firm defines API-equivalent value as "the plan's full monthly usage limit priced at first-party list API rates." The multiplier shown next to each bar is that value divided by the fee, so $11,726 on a $200 plan is 58.6x.
Two workload shapes are used. The agentic mix is 0.4 percent fresh input, 96.6 percent cached input, 2.6 percent cache writes and 0.3 percent output, which reflects SemiAnalysis's own September usage ratios. The chat mix is 2 percent input, 75 percent cached input, 13 percent cache writes and 10 percent output.
The headline numbers
The lab below redraws the published figures so you can switch between views. The tables that follow hold the same data, so the page is complete without it.
Mid-tier models: where the 5x comes from
| Plan fee | ChatGPT, GPT-6.1 Sol | Claude, Opus 5.5 | Claude advantage |
|---|---|---|---|
| $200 per month | $2,084 (10.4x) on Pro 200 | $11,726 (58.6x) on Max 20x | about 5.6x |
| $100 per month | $1,055 (10.6x) on Pro 100 | $5,725 (57.3x) on Max 5x | about 5.4x |
| $20 per month | $211 (10.6x) on Plus | $1,178 (58.9x) on Pro | about 5.6x |
SemiAnalysis calls this tier "marketed by both companies as the intended daily driver for most users" and says Anthropic is "an overwhelmingly better deal." It anticipates the fairness objection that Sol is much cheaper per token than Opus, and says the gap remains large when measured in raw tokens per dollar, not dollars of API value.
Flagship models: roughly level
| Plan fee | ChatGPT, GPT-6 Astra | Claude, Fable 5.1 |
|---|---|---|
| $200 per month | $2,897 (14.5x) | $2,485 (12.4x) |
| $100 per month | $1,322 (13.2x) | $1,273 (12.7x) |
| $20 per month | $162 (8.1x) | Not on Claude Pro |
The footnote matters: Fable 5.1 can be used for only up to half of the plan's usage limit. SemiAnalysis frames it this way: after spending $2,485 of Fable, your $200 Anthropic plan "would still have 50% left," while the equivalent OpenAI plan "would be fully exhausted after $2,897 of Astra." That reserve is what pushes Claude ahead once you move down to Opus.
Inside one Claude plan
| Model on Max 20x ($200) | Agentic | Chat |
|---|---|---|
| Sonnet 5.5 | $12,529 (62.6x) | $7,009 (35.0x) |
| Opus 5.5 | $11,726 (58.6x) | $9,200 (46.0x) |
| Fable 5.1 | $2,485 (12.4x) | $2,480 (12.4x) |
The ladder is the story. Within the same fee, value falls sharply as you move to more powerful and more expensive models. That is by design, as the section on strategy below explains.
How SemiAnalysis measured the limits
The method is simple enough to understand, and careful enough to trust more than a screenshot of a meter.
- Isolate one token type per experiment. Input, cache write and cache read tests use a passage of War and Peace, because some models refuse gibberish. Input tests add a random tag to every call so nothing is cached. Cache write tests mark the prompt for caching with a new tag each time. Cache read tests use a fixed tag, so the first call writes and repeats read. Output tests use a technical essay to force long generations.
- Count in steps, not requests. A single request often does not move the meter. The firm records the running token total each time the meter rises, defines the tokens between two moves as one step, and drops the partial first and last steps to avoid bias.
- Keep adding steps until the range is within plus or minus 5 percent. Because the meter is read once per request, a single step can be off by up to one request, so the result is tracked as a range that narrows as steps accumulate.
- Subtract the fixed overhead. No request has only one token type, and some plans require standing instructions on every call, so the cost of those extras is removed using the rates already measured.
- Treat very cheap reads as free. If the meter does not move after 500 million cache read tokens, the firm assumes cache reads are free on that plan.
- Convert to tokens per window and per month. If a million tokens use 5 percent of a limit, the limit holds 20 million. The 5-hour window resets many times a week, so a month is capped by the weekly limits.
- Price the result. Consuming 1 percent of a $200 plan's monthly limit equals $2. Multiplying the achievable token mix by blended API list prices gives the API-equivalent value.
One side finding deserves attention. One of three identical accounts at a provider had roughly 20 percent lower limits. The provider later confirmed it was part of an "extremely tiny" A/B test on limits, and said it did not "decrease limits wholesale" but was testing "how to better balance when people hit limits." SemiAnalysis draws two lessons: providers can silently change limits at any time, and its method is sensitive enough to catch small shifts.
Why OpenAI and Anthropic behave differently
SemiAnalysis argues subscriptions matter far more to lab economics than their revenue share suggests. For Anthropic it estimates they are about 10 percent of revenue but over 40 percent of inference compute, lowering blended revenue per megawatt by roughly $36 million. These are the firm's modeled figures from its Tokenomics Model, not company disclosures.
The two labs have chosen different ways to reduce that subsidy:
- Anthropic lowers value as models get stronger. Sonnet to Opus is a small drop; Opus to Fable is a big one. On a maxed-out plan with 100 percent utilization and 92 percent API gross margins, SemiAnalysis models Opus 5.5 and Fable 5.1 at about negative 369 percent and 1 percent gross margin respectively. At an assumed 20 percent average utilization those become roughly 6 percent and 80 percent. In its view, Anthropic would have "software like margins" if everyone used only Fable.
- OpenAI cut limits across the board. SemiAnalysis describes it as the "nuclear option" of moving straight to Fable-level limits. Its read is that the DevDay news cycle and a one-month grandfathering window blunted the backlash.
For context on that cut, see our coverage of the Pro 200 halving on October 30 and the earlier Pro 200 reopening at half the API usage.
What changed at OpenAI
SemiAnalysis confirmed the 50 percent value reduction on the $200 plan. OpenAI halved the tokens per model tier, and also cut cached input pricing for GPT-6.1 Sol, which together pushed Sol-class API-equivalent value down by more than half. Plans bought before the cut keep the old limits until October 29, and new purchases get the lower limits immediately.
The new $500 plan offers only about 21 percent more Astra than the old $200 plan, according to the firm. Its headline feature is Ultrafast speed, quoted in the article at 300 tokens per second, which we covered in the Ultrafast and Pro 500 breakdown and explain in our tokens per second guide. SemiAnalysis says it is still testing Ultrafast limits and usage through Sign in With ChatGPT in third-party tools such as Devin, and will publish those results to subscribers.
Before the cut, the $200 ChatGPT plan was significantly more generous than OpenAI's others: Pro 100 gave about 2x Plus per dollar on Astra, and Pro 200 about 2x again. Now Pro 100, 200 and 500 offer the same tokens per dollar on every model, and Plus is comparable for Sol but worse for Astra. Anthropic's tiers already gave the same per-dollar value across the board.
SemiAnalysis also points to one real counter-argument for OpenAI: none of its Pro plans has a 5-hour limit, which makes it easier to use a high share of the monthly allowance. It does not think that offsets a roughly 4x value gap versus Opus 5.5.
Do labs adjust limits when API prices fall?
Not fully. The article lists recent price cuts: Fable 5.1 cut cache reads by 75 percent versus Fable 5; Opus 5.5 cut input and output pricing by 20 percent and cache reads by 60 percent versus Opus 5; GPT-6.1 Sol cut cache reads by 50 percent versus GPT-6 Sol.
OpenAI did not change Sol token limits on the $200 plan when 6.1 Sol arrived, so API-equivalent value fell about 30 percent. Anthropic did not raise Fable limits for 5.1, but did raise Opus limits with 5.5: roughly 20 percent more tokens on Max plans and about 50 percent more on Pro. That was not enough to offset the price cut, so Opus API-equivalent value still fell. The practical lesson: when list prices drop, the headline "value" of a plan can fall even if you get more tokens.
Other plans in the study
SemiAnalysis says Chinese labs still offer subsidized plans, with large differences between providers, and value per dollar rises on higher tiers. On average their API-equivalent value per dollar is a little under the roughly 12x OpenAI offers. It also reports that third-party plans, such as Cursor and Cognition's Devin, are worse than first-party plans for the same underlying models. The detail on those sits behind the paywall.
What this does not tell you
A good analysis still has limits, and this one says so itself.
- API list price is not what you would pay. Value is computed at list rates. If you would never buy that volume through the API, the number is a comparison index, not savings.
- The workload is SemiAnalysis's own. A mix that is 96.6 percent cached input rewards plans and models with cheap cache reads. Your usage may look different.
- Token efficiency is unmeasured. The firm says the industry lacks reliable data here and does not trust the benchmark tasks in the Artificial Analysis index as representative of real work. If one model needs fewer tokens to finish a task, a smaller plan can go further. Our GPT-6.1 Sol cost efficiency post covers that angle.
- Quality is not in the chart. Opus 5.5 and Sol are different models, so more value in dollars does not mean better results per task.
- It is a snapshot. Labs change credit costs without notice. SemiAnalysis suggests rechecking every combination daily, which is what its dashboard does.
- We saw only the public portion. Methodology and headline charts are public; the dashboard and several sections are subscriber-only.
What this means for what you build or pay
If you code with an agent daily, the practical read is straightforward.
- Match the model to the task. On Claude, Sonnet 5.5 and Opus 5.5 give by far the most value; keep Fable for what needs it. On ChatGPT, Sol stretches your allowance further than Astra per task if quality is enough.
- Do not buy up for value alone. On ChatGPT, Pro 100, 200 and 500 now give the same tokens per dollar, so you pay for headroom and speed, not a better rate.
- Test on your own workload. Use a cheaper plan for a week and watch the meter on your real tasks before committing to a $200 tier.
- Keep a second harness ready. Limits change quickly. Our Claude Code vs Codex comparison and usage limits timeline show how often.
- Plan around the date. If you hold a $200 ChatGPT plan bought before the cut, the higher limits end on October 29.
Related reading on explainx.ai
- Dots shipped, Pro 200 halved, Sol got a global reset
- ChatGPT Pro 200 reopens at half the API usage
- Ultrafast and Pro 500 at DevDay 2026
- GPT-6.1 Sol cost efficiency versus Astra
- Claude usage limits timeline explained
- Claude Code vs Codex after the rate limit boost
- The true cost of AI subscriptions
- What is TPS? Tokens per second explained
All figures are from SemiAnalysis's October 2026 subscription limit-testing article and are accurate as of its publication; labs can change limits at any time. Charts here are redrawn by explainx.ai from the published numbers, not copies of SemiAnalysis's images. The firm's dashboard and full methodology are subscriber-only.