SemiAnalysis: Claude Plans Deliver 5x ChatGPT's API Value on Opus 5.5 vs. GPT-6.1 Sol

TL;DR

A SemiAnalysis report published October 5 metered every consumer plan from Anthropic and OpenAI, then priced each plan's full monthly limit at API list rates. On an agentic coding workload, a $200 Claude Max 20x plan carries $11,726 a month of Claude Opus 5.5. A $200 ChatGPT Pro plan carries $2,084 of GPT-6.1 Sol. The ratio stays above 5x at the $100 and $20 tiers. On flagship models the two labs are close, and OpenAI is slightly ahead. The gap opened because OpenAI halved its $200 plan on September 29. The same report estimates that a fully used Opus plan runs at a -369% gross margin, so treat these numbers as a snapshot of a subsidy both labs are trying to shrink.

How you price a meter with no numbers on it

Subscription plans never tell you how many tokens you bought. Tools like Claude Code and Codex show a 0 to 100% usage meter over 5-hour and 7-day windows, and the plan tiers are described only relative to each other.

So SemiAnalysis authors Andrew Megalaa, Max Kan and Dylan Patel measured the plans from the outside. They ran repeated calls built to be almost entirely one token type (fresh input, cache writes, cache reads or output), counted the tokens the provider billed, and recorded how far the meter moved. The input prompts were chunks of War and Peace, because some models refuse to answer large blocks of gibberish. Tolstoy has had worse readers.

Think of a fuel gauge with no numbers on the dial. You burn a measured amount of fuel, wait for the needle to drop one notch, and repeat until you know what a notch holds. SemiAnalysis kept adding notches until each rate was pinned within ±5%.

The rest is arithmetic. The weekly limit caps the month, each token type takes a measured share of that limit, and the plan's capacity gets priced at API rates for a fixed workload mix. The "agentic" mix is SemiAnalysis's own September usage: 0.4% input, 96.6% cached input, 2.6% cache writes and 0.3% output.

Mid-tier models: more than 5x at every price

Opus 5.5 and GPT-6.1 Sol are the models each lab pitches as the daily driver. This is where SemiAnalysis found the widest gap.

The report's own summary: "Anthropic is an overwhelmingly better deal, offering ~5x the API-equivalent value across the board."

Dollar value flatters Anthropic somewhat, because Opus 5.5 costs about twice what Sol does per token at API prices ($0.41 against $0.20 per million tokens on this mix, by SemiAnalysis's blended figures). Count raw tokens and the lead shrinks to 28.6 billion a month on Max 20x against 10.2 billion on Pro 200. That is still 2.8x.

Flagship models: close to a tie

Move up to the top models and the gap closes. A $200 ChatGPT Pro plan carries $2,897 of GPT-6 Astra. Claude Max 20x carries $2,485 of Claude Fable 5.1. At $100 it is $1,322 against $1,273. Fable is not on the $20 Claude Pro plan at all, while ChatGPT Plus includes $162 of Astra.

One difference matters if you mix models. Anthropic caps Fable at half of a plan's usage limit, so a Max 20x subscriber who burns the full $2,485 of Fable still has the other half of the plan for Opus or Sonnet 5.5. The ChatGPT plan is empty after $2,897 of Astra.

What OpenAI changed on September 29

On September 10, OpenAI stopped selling the $200 plan because Astra demand was eating its compute. On September 29 the plan came back at half strength, next to a new $500 tier.

Thibault Sottiaux, who leads Codex at OpenAI, said as much when he announced it, as Implicator reported: "In effect, if you do the math, it will net out at half the dollar in API spend compared to the old Pro $200 plan." SemiAnalysis's meter readings confirm his math.

• Pro 200 on Astra. $5,734 of API value became $2,897, and 3.8 billion tokens became 1.9 billion. Both fell 49%.
• Pro 200 on Sol. Tokens fell 48%, from 19.6 billion to 10.2 billion. API value fell 65%, from $5,904 to $2,084, because GPT-6.1 Sol also arrived with cheaper cached input.
• Pro 500. The fee is 150% higher than the old $200 plan. It buys 21% more Astra and 9% less Sol-tier value than the old plan did.

In fairness to the $500 plan, its headline feature is Ultrafast, the 300-tokens-per-second speed tier, and SemiAnalysis says it is still testing those limits. Subscribers who bought Pro 200 before the cut keep the old limits until October 29.

The cut also flattened OpenAI's price ladder. The old $200 plan gave roughly twice the per-dollar value of the $100 plan. Now Pro 100, Pro 200 and Pro 500 all buy the same tokens per dollar, which is how Anthropic's tiers already worked.

It is a fast reversal. In June, SemiAnalysis put a fully used $200 ChatGPT plan at $14,000 of API value and Claude Max 20x at roughly $8,000, as TechSpot reported at the time. Models and prices have changed since, so the dollar figures do not line up one to one, but the order has flipped. The new report says OpenAI "used to be lauded by indie developers for their generous subscription limits compared to Anthropic, but this is simply no longer true today."

API price cuts mostly do not reach your plan

The finding you can reuse beyond this month is what happens to a subscription when the API price drops. If the lab holds your token limit flat, you get the same tokens for the same fee and the lab's cost of serving you falls. On paper, your "API-equivalent value" shrinks.

• Opus 5 to Opus 5.5. The blended API price fell 46%. Anthropic raised token limits 51% on Pro, 20% on Max 5x and 25% on Max 20x, which went from 22.9 billion tokens a month to 28.6 billion.
• Fable 5 to Fable 5.1. The blended price fell 48%. Token limits moved 1% to 4%.
• GPT-6 Sol to GPT-6.1 Sol. The blended price fell 32%. Token limits rose 19% on Plus and 5% to 8% on the Pro plans.

In all three cases API-equivalent value fell, because limits rose by less than prices dropped. Opus 5.5 is the only one where subscribers got a real share of the saving.

The margin math says this will not last

SemiAnalysis treats subscriptions as a marketing expense that shows up in the compute bill. Its rough model of Anthropic has subscriptions bringing in 10% of revenue while taking 41.7% of inference compute. It estimates $16 million of revenue per megawatt at a 50% gross margin for subscriptions, against $102.9 million per megawatt at 92.2% for the API.

On a single plan the numbers turn ugly fast. Max 20x on Opus 5.5 carries 58.6 times its fee in API value, so the plan breaks even at 21% utilization.

At what SemiAnalysis calls a "more realistic average utilization of 20%," the Opus margin is 6% and the Fable margin is 80%. Its conclusion: "Anthropic would likely already have software like margins on their subscription plans if everyone only used Fable."

A plan that hands over $58 of API-priced usage for every dollar of fee is a customer acquisition strategy right up until the customers use it.

The two labs are closing the subsidy in different ways. Anthropic gives each plan less API value on its priciest model: 12.4 times the fee on Fable against 58.6 times on Opus. OpenAI, in the report's words, "picked the nuclear option of just immediately cutting to Fable-level limits across the board." SemiAnalysis expects any Anthropic tier above Fable to come with lower limits still, if plans include it at all.

Chinese plans, and the part behind the paywall

The report also meters coding plans from Z.ai, MiniMax and Moonshot. All three subsidize their plans, by very different amounts:

• MiniMax Token Ultra, $132 a month: $3,088 of MiniMax-M3, or 23.4 times the fee.
• Z.ai GLM Coding Max, $168 a month: $1,942 of GLM-5.3, or 11.6 times.
• Moonshot Kimi Code Ultra, $199 a month: $1,343 of Kimi K3, or 6.7 times.

The final section compares first-party plans with Cursor and Cognition. Its heading says third-party plans are worse. The evidence sits behind SemiAnalysis's paywall, so we could not check it.

Caveats

• Value is not finished work. These figures price tokens. They do not count tasks completed. Sottiaux's defense of the cut, per Implicator, is that cheaper and more efficient models let developers "still get more work done than if you were on the Pro $200 subscription one month ago." SemiAnalysis concedes that the industry "lacks reliable data" on token efficiency.
• The mix is one team's usage. A 96.6% cached-input share describes long agent sessions. On a chat-style mix, the same Max 20x plan is worth $9,200 of Opus 5.5 instead of $11,726.
• Full utilization is a ceiling. Claude plans have a 5-hour limit and OpenAI's Pro plans do not, which makes the OpenAI allowance easier to use up. SemiAnalysis does not think that offsets the gap.
• Limits move without notice. One of three identical accounts the team tested had limits about 20% lower than the other two. The provider, which the report does not name, confirmed an "extremely tiny" A/B test.
• The margins are modeled. They assume a 92% API gross margin and come from SemiAnalysis's Tokenomics Model, a paid product the article also promotes. Neither lab publishes these figures.

What to do with this

If you run coding agents on a flat plan, the subsidy today sits in the mid-tier Claude models. Opus 5.5 and Sonnet 5.5 on any Claude plan return 57 to 63 times the fee at full use. Nothing OpenAI sells returns more than 15 times.

If you bought ChatGPT Pro 200 before September 29, your limits halve on October 29. Check your current usage against the new ceiling before then.

If your product or your budget depends on someone else's flat-rate limits, treat those limits as a promotion. One lab halved them this quarter with a month of grace for existing subscribers, and an A/B test can move them with no notice at all.

Key Takeaways

• SemiAnalysis metered each plan's limits and priced them at API list rates, on an agentic mix that is 96.6% cached input.
• $200 buys $11,726 of Opus 5.5 on Claude Max 20x and $2,084 of GPT-6.1 Sol on ChatGPT Pro 200. The ratio is 5.4x to 5.6x at every tier, and 2.8x in raw tokens at $200.
• Flagship models are close: $2,897 of GPT-6 Astra against $2,485 of Fable 5.1, with Fable capped at half the plan.
• OpenAI halved Pro 200 on September 29. The new $500 plan costs 150% more than the old $200 plan and carries 21% more Astra. Grandfathered limits end October 29.
• When API prices fell, subscription token limits mostly stayed put. Opus 5.5 was the partial exception.
• A fully used Opus plan models out at a -369% gross margin, so expect both labs to keep trimming.

Sources: SemiAnalysis: Anthropic Subscriptions Offer 5x+ More Value Than OpenAI, Implicator, Techmeme, TechSpot (June estimate), SemiAnalysis Subscriptions Dashboard, Claude pricing