Tech Analysis Firm SemiAnalysis Finds: Anthropic Is the Ultimate Benefactor, Offering Over 5x the API-Equivalent Value of OpenAI in Claude Subscriptions...

If AI subscriptions are viewed as providers handing out compute subsidies to users, Anthropic looks far more like a “benefactor” this time, while OpenAI is rolling back its discounts.

Subscription benchmarks from tech analysis firm SemiAnalysis show that in a head-to-head comparison of workhorse models Opus 5.5 and GPT-6.1 Sol, for the same monthly fee, Claude delivers over 5x the API-equivalent value of OpenAI. Even setting aside API list prices and looking strictly at total tokens, Claude still provides significantly more. (An API is an interface for programs to call models; a token is a basic unit used by models to process text.)

Switching to the flagship Fable 5.1 versus GPT-6 Astra, individual quotas are close, but Claude retains an extra buffer: if you haven’t burned through the shared quota with other models beforehand, maxing out Fable still leaves half of your total quota available for other models. In contrast, once Astra reaches its quota limit, OpenAI’s quota pool for that plan is completely drained.

Meanwhile, OpenAI slashed the token quota on its $200 tier by roughly half while introducing a new $500 tier. Despite the substantially higher monthly price tag, the volume delivered per dollar is no better deal.

What this analysis truly uncovers is how AI providers leverage subscription subsidies to attract users while quietly figuring out how to claw them back. Claude remains generous with its workhorse model, whereas OpenAI has taken an axe directly to overall plan quotas.

Paying the Same $200, How Much More Does Claude Actually Give?

First, look at the models promoted by both sides as daily workhorses. Based on SemiAnalysis’s agent workload and assuming full quota utilization:

For the same monthly fee, the API-equivalent value differs by roughly 5.6x, while total tokens differ by roughly 2.8x. Comparing these workhorse models across the $20 and $100 tiers also shows an API-equivalent gap of over 5x.

“API-equivalent value” can be understood as: how much you would roughly pay if you purchased the exact volume allowed by your subscription quota at that model’s standard API rate card. It is neither cashable credit nor the actual compute cost incurred by the provider on your behalf. Different token types are multiplied by their respective unit prices and summed up.

Therefore, two factors drive that 5.6x figure: Claude provides more volume, and Opus commands a higher API list price. Looking only at price conflates the two; switching to total tokens, the gap remains substantial, demonstrating that Claude’s edge isn’t merely an artifact of an inflated list price looking attractive on paper.

In this workload, roughly 96.6% of tokens are cache reads. When an agent runs code, inspects error traces, and edits files, it repeatedly resends previously read code and conversational context; caching allows the model to reuse processed context, lowering repeated reading costs. Reading existing code ten times over is not the same as writing ten times as much new code.

Consequently, this table answers how much headroom heavy agent workflows can obtain. How much additional work actually gets completed depends on which model handles a given task better and requires less rework. SemiAnalysis also points out that the industry lacks reliable data reflecting real-world token efficiency.

Why Does the Same Plan Vary So Wildly Between Chat and Agent Workloads?

A subscription is not a fixed-value dollar voucher. On the same $200 Max 20x tier, switching models and usage patterns can swing the API-equivalent value from roughly $2,500 to around $12,500.

Take Sonnet as an example: its agent workload yields an equivalent value of roughly $12,529, compared to around $7,009 for chat; Opus yields about $11,726 versus $9,200. Fable hovers near $2,480 under both use cases, showing minimal difference. Model and workload interact to reshape the outcome, meaning one cannot blanketly assume “agent workloads are always worth more” across every model.

The reason lies in two differing tariff schedules: subscriptions deduct quota according to model and token type, while APIs charge by token type, but their underlying cost proportions do not match. Agents repeatedly reread past context, with cache reads accounting for 96.6% in this valuation; in chat workloads, cache reads drop to around 75%, while generated output jumps from approximately 0.3% in agent tasks to 10%. As the token mix shifts, the volume of calls that a given quota can support—and its corresponding API price—changes with it.

Flagship Models Are Close, but “Hitting the Limit” Means Something Very Different

In the $200 tier, Fable 5.1’s API-equivalent value sits at roughly $2,485, compared to about $2,897 for Astra. Looking strictly at this line, OpenAI is even slightly ahead.

The $100 tier is also neck-and-neck: Astra stands at roughly $1,322 and Fable at about $1,273. The $20 tier, however, is asymmetrical: ChatGPT Plus includes an Astra quota, whereas Claude Pro excludes Fable, so flagship comparisons in the high tiers cannot simply be mapped onto the entry-level tier.

Crucially, Fable’s cap can only consume up to 50% of Claude’s total weekly quota. If you use Fable exclusively and max out this flagship allocation, you only hit its dedicated sub-cap; the remaining half of your shared quota is still available for Opus or Sonnet.

Think of it as a shared quota pool with a rule that “the flagship can take at most half.” Exhausting the flagship portion does not mean the pool is empty. In contrast, running Astra to its limit in OpenAI’s corresponding plan completely exhausts the shared quota.

The remaining “50%” represents quota share—not an extra $2,485 gift, and certainly not an allowance to keep running that same volume of Fable. Other models burn quota at different rates; furthermore, if you previously spent part of the shared pool on Opus, your remaining share decreases accordingly.

For users who switch between multiple models, this distinction is very practical: offload tough bottlenecks to the flagship, while routing routine edits and testing to workhorse models, preserving operational headroom even after the flagship hits its ceiling. That is where the generous Opus headroom highlighted in the earlier table comes into play.

OpenAI’s Cut Slashes the Perks of the Old $200 Tier

The changes tracked by SemiAnalysis are directly reflected in the figures before and after the adjustment:

Both Astra’s token volume and equivalent value dropped by roughly 49%. Sol’s token count dropped by around 48%, while its API-equivalent value plunged by about 65%. The latter also factored in cached API price cuts when transitioning from GPT-6 Sol to GPT-6.1 Sol, so the 65% drop cannot be blamed entirely on quota cuts. Existing subscribers temporarily retain their original allowances, while new subscribers are immediately subject to the new quotas.

Next, look at the $500 tier: the monthly fee is 150% higher than the old $200 tier, yet Astra volume increases by only about 21%. This comparison is against the pre-adjustment old $200 plan; comparing it against the already-slashed new $200 plan would naturally make the step-up look far larger.

Sol exhibits another sharp contrast: the new $500 tier delivers roughly 26.3 billion tokens—about 34% more than the old $200 tier’s 19.6 billion—yet its API-equivalent value actually falls from roughly $5,904 to $5,386, a drop of about 9%. Token volume rose, but the pricing yardstick dropped faster, underscoring once again that equivalent value cannot be treated as a direct proxy for token volume.

Another selling point of the $500 tier is its 300 TPS Ultrafast output speed. (TPS refers to tokens generated per second.) Receiving answers faster can be valuable, but speed must be valued on its own terms; at the time of writing, SemiAnalysis was still testing the rate limits of this feature.

Previously, Pro 200 was a discount tier where “paying more per month yielded even more volume per dollar.” Following the adjustment, the token-per-dollar ratio across Pro 100, 200, and 500 for the same model has largely flattened: paying extra primarily buys higher aggregate volume or speed, rather than unlocking a cheaper unit cost. Claude’s subscription tiers have historically maintained a fairly consistent value-per-dollar ratio across tiers, but the subsidization level for its workhorse model remains markedly higher.

When reading the value-per-dollar chart, the horizontal axis represents tokens per dollar (further right indicates more tokens), while the vertical axis shows API-equivalent value divided by monthly subscription cost (higher up indicates a steeper on-paper discount relative to pay-as-you-go). The horizontal axis uses a logarithmic scale where equal distance denotes multiplicative change, so visual distance should not be mistaken for linear increments.

Plus must also be viewed on a model-by-model basis: Sol’s value per dollar is close to Pro, whereas Astra’s is lower. Previously, Astra’s value per dollar on Pro 100 was about double that of Plus, and the old Pro 200 added roughly another doubling over Pro 100; that premium tier discount has now been rolled back.

In the cross-provider comparison chart, Sonnet and Opus deliver API-equivalent values roughly 57–63x their monthly fee, compared to about 10–11x for OpenAI Sol; Fable delivers only about 12–13x, closely aligning with Astra’s roughly 13–15x. Models from the same vendor sit in entirely different subsidy tiers, so one cannot extrapolate a workhorse model advantage into an across-the-board blowout.

OpenAI also possesses an advantage that cannot be overlooked: its Pro plans do not impose a 5-hour rate limit. Claude users, even with ample weekly quota remaining, may hit short-term rolling limits first. For individuals running continuous, intensive bursts of tasks, “plenty of aggregate quota” and “whether you can actually draw on it during these specific hours” are two distinct realities. This is precisely why total quota alone cannot dictate the right plan for everyone.

Taking 40% of Compute but Generating Only 10% of Revenue: Why Do Providers Do It?

When users perceive subscriptions as an incredible bargain, the flip side is that providers forfeit substantial revenue that could have been collected under pay-as-you-go metering.

SemiAnalysis’s rough unit economics estimate for Anthropic is striking: subscriptions account for roughly 41.7% of inference compute while contributing only about 10% of revenue; the API consumes about 58.3% of compute yet generates approximately 90% of revenue.

In these estimates, subscriptions yield roughly $16 million in revenue per megawatt (MW) of compute, compared to approximately $102.9 million for the API. The same compute resources deployed in the API business recover vastly more cash.

Blending both lines of business brings revenue per MW to roughly $66.7 million, about $36.2 million less than the pure-API figure. This is simply the mathematical consequence of allocating compute across tiers of varying revenue efficiency—not a realized net loss of $36.2 million.

This explains why subscriptions drag down overall revenue efficiency and why providers scrutinize subscriber quotas so closely. But it does not mean “all subscriptions bleed massive losses”: the chart estimates subscription gross margins at around 50%, simply far below the API’s roughly 92.2%. Heavy subsidization and leaving massive upside on the table are entirely different conclusions from operating at an aggregate net loss.

Providers are willing to absorb this because a low-monthly-cost, friction-free on-ramp makes it easy for developers and small businesses to experiment and integrate models into their daily workflows. Subscriptions are not merely a stream of query revenue; they are a critical customer acquisition and distribution vehicle. Establishing user habit first, then optimizing unit economics and monetization later—that is the business logic driving these quota shifts.

SemiAnalysis also partly attributes earlier growth in Codex adoption to OpenAI’s generous quota resets, arguing that competitive pressure similarly prompted Anthropic to withdraw some proposed cuts to subscription benefits. This represents the firm’s assessment of competitive dynamics, illustrating why subsidies cannot be evaluated strictly by the profit or loss on a single query: cutting allowances might save compute cycles, but at the cost of bleeding users and market goodwill.

Both Want to Cut Subsidies, but They Are Wielding the Scalpel in Different Places

Anthropic’s approach is to impose tighter quotas on its more powerful, expensive flagship model. When benchmarked against API list prices, Sonnet and Opus boast massive subsidy multiples; when it comes to Fable, that multiplier drops sharply.

A flagship query costs more per call, but the allowable volume granted to users is substantially lower. By keeping quotas sufficiently tight, flagship tiers can actually prove more profitable than the loose allowances on workhorse models.

Assuming a 92% API gross margin, SemiAnalysis modeled two hypothetical usage scenarios under Max 20x:

These are scenario projections, and the 20% figure is a hypothetical utilization rate rather than an empirically measured user average. Taking Opus as an example: an API-equivalent value of $11,726 corresponds to roughly $938 in cost of goods sold (COGS) under the 92% API gross margin assumption; maxing it out means the $200 monthly subscription fee falls far short of covering serving costs. At 20% utilization, COGS drops to roughly $188, bringing gross margin back into positive territory at around 6%.

Because Fable’s available quota is much tighter, applying the same calculation yields serving costs of about $40 at 20% utilization, translating to an ~80% gross margin. Though Fable is more capable and expensive, its subscription gross margin can actually be higher: the key is a tighter quota and reduced subsidy, proving that one cannot judge margins solely by per-token inference costs.

This modeling also reveals a critical break-even threshold: on Max 20x, Opus reaches break-even at roughly 21% quota utilization, and Sonnet at about 20%; maxing out Fable’s subquota still leaves it near break-even (~1% margin). This underscores just how heavily provider profitability hinges on the actual percentage of quota consumed by users.

The Fable column accounts solely for consumption of Fable’s dedicated subquota. If a user subsequently consumes the remaining half of the shared pool with Opus, serving costs will increase accordingly, meaning the 80% figure cannot be taken as the actual gross margin for the user’s entire subscription package.

Under the same set of assumptions, the adjusted OpenAI Pro 200 delivers an estimated gross margin of ~17% when fully maxing out Sol, and roughly −16% when maxing out Astra. Capping total available volume enables providers to rein in cost exposure from power users.

In the new Pro 200, Astra only reaches break-even at around 86% quota utilization; compared to the ~21% inflection point for Opus, OpenAI enjoys a far larger cost buffer against heavy usage. This comparison reflects modeled scenarios under uniform cost assumptions, rather than actual subscription margins disclosed by either company.

Anthropic claws back subsidies incrementally through quota allocation between flagship and workhorse models. OpenAI, meanwhile, slashed total quota on the old $200 tier by roughly half. For end users, the former feels like “flagship access is tight, but routine models still offer plenty of runway,” while the latter manifests as “paying the same monthly fee, but available volume suddenly dropped off a cliff.”

This does not mean Claude will maintain today’s subsidy levels indefinitely. SemiAnalysis expects that as Anthropic continues lowering the serving costs of Opus, generous quotas on workhorse models will become easier to sustain; future, higher-tier flagships, however, might receive even tighter caps or be excluded from subscriptions altogether. This represents the firm’s forward-looking speculation on product tiering, not an officially announced policy.

As for why OpenAI’s quota cut did not spark severe immediate backlash, the firm speculates that positive headlines from launch announcements, combined with grandfathering existing subscribers, helped cushion sentiment. While this explains why backlash might be delayed, it should not be taken as proof that the company deliberately staged launch events to conceal quota reductions.

When API Prices Drop, Why Don’t Subscribers Automatically Get More Quota?

Providers wield another yardstick that is all too easy to overlook: the API rate card.

An API price drop means pay-as-you-go program calls become cheaper. How much more a subscriber can consume, however, is governed by an entirely separate quota deduction engine. There is no binding promise that the two rate systems must move in lockstep.

Opus 5.5 provides a clear illustration: according to the charts, its blended agent API unit price dropped by roughly 46%, and token allowances on Claude Max 20x rose by about 25%, yet its API-equivalent value still dropped by approximately 32%. Users received more actual tokens; but because buying that same volume via API also became cheaper, the on-paper “dollar equivalent” of the subscription actually declined.

The volume uplift for the same model also varies across tiers: on Claude Pro, Opus quota increased by roughly 51%, yet its API-equivalent value still dropped by about 18%; on Max 5x, quota rose by ~20% while equivalent value fell ~35%. Fable’s Max tiers saw quota increases of only ~1–4%, with equivalent values plunging ~46–47%; Sol on Pro 200 saw quota increase by ~5%, while equivalent value fell ~28%.

Token allowances for Fable and Sol in their respective new releases saw only modest increases, failing to offset API price reductions. Consequently, a “shrinking API-equivalent value” can mean less volume was granted, list prices dropped, or both happened at once. OpenAI’s comparison between the old and new $200 tiers compounds both effects simultaneously.

Underlying subscription deduction rules can even be tweaked without altering plan names or monthly fees. SemiAnalysis tested three identical subscription accounts from an unnamed provider and found one had a quota roughly 20% lower; upon contacting the vendor, it confirmed the account was part of a small-scale quota A/B test.

This explains why ongoing measurement is essential: a subscriber’s “100% quota” does not correspond to a permanently fixed token quantity. Whenever a provider adjusts the deduction rate for a given model or token category, the real-world volume you can consume shifts accordingly.

Chinese Providers Are Subsidizing Too, but Averages Mask Massive Variances

Public charts also benchmarked MiniMax, Z.ai, and Moonshot. Evaluated against their respective models’ API rate cards, MiniMax M3 yields roughly 19–24x its monthly fee, GLM 5.3 yields about 8–12x, and Kimi K3 yields roughly 2.5–6.7x.

Thus, asking “what is the average subsidy of Chinese plans?” is less useful than asking “which specific model and tier?” While these plans generally trend toward higher value-per-dollar at higher tiers, it is not a strictly monotonic progression. Models with low baseline API pricing can yield modest multipliers even when handing out massive token volumes.

MiniMax’s Max tier costs $55/month with an equivalent value of ~$1,305 (23.7x); its Ultra tier costs $132/month with an equivalent value of ~$3,088 (23.4x). The pricier tier gives more total volume, but its perk per dollar does not increase further. GLM 5.3 ramps up from ~7.7x on Lite to ~11.6x on Max, and Kimi scales from ~2.5x on Plus to ~6.7x on Ultra, reflecting distinct ladder structures across vendors.

GLM Flash clearly demonstrates that “more tokens” and “higher subsidy multiplier” are not synonymous: on the same $168 Max tier, GLM 5.3 delivers roughly 6.3 billion tokens and $1,942 in equivalent value, whereas Flash delivers roughly 19.1 billion tokens and $336 in equivalent value. Flash delivers far more tokens, but because its API is inherently cheap, its equivalent multiplier is only around 2x. On the normalized chart, it lands in the bottom-right rather than the bottom-left. (Both axes on this cross-provider chart are logarithmic.)

These figures measure the degree of subsidization each subscription offers relative to its own API pricing. When switching models, whether tasks can actually be resolved and how many inference rounds are required remains a separate evaluation.

Without Backend Visibility, How Were These Quotas Measured?

SemiAnalysis did not simply count “how many messages were sent,” but rather measured how much each specific token category moves the subscription usage meter. It continuously tracks models, plans, promotions, and workloads: even when plan names remain unchanged, backend quota deduction rates can shift.

Step one: isolate and test new inputs, cache writes, cache reads, and generated outputs independently. New input runs use randomized prefixes on every call to prevent cache hits; cache-write experiments add explicit cache control markers on top of this; cache-read experiments keep prefixes fixed, writing on the initial turn and repeatedly reading thereafter. Input tests employ excerpts from War and Peace while keeping completion length to a minimum; output tests use concise prompts instructing the model to generate extensive technical essays. Because real-world requests inevitably include trace amounts of other token types, the tests mathematically subtract their consumption.

Step two: resolve rounding errors in usage meters. Usage meters are polled only after requests complete, meaning actual consumption might have just barely crossed a 1% increment or overshot it significantly. As a result, a single meter jump cannot serve as an exact deduction rate. The original diagram illustrates this with an example of “8K tokens per step, with ~21K tokens actually burning 1%”: an individual observed step might appear as 16K or 24K, and only multi-step averaging converges on the true ratio. The benchmark discards incomplete initial and trailing steps, accumulating full steps until the measurement interval narrows to roughly ±5%. If consuming 500 million cached tokens fails to move the meter at all, cache reads are treated as free—this is an empirical benchmark assumption, not a vendor guarantee of permanent free tier status.

Step three: monitor short-term rolling limits, overall weekly quotas, and flagship subquotas simultaneously. In the original diagram’s example for Fable new inputs, every 1 million tokens consumed ~3.3% of the 5-hour limit, ~0.88% of the all-model weekly quota, and ~1.76% of the Fable weekly quota; back-calculating capacity yields roughly 30 million, 113 million, and 57 million tokens, respectively.

These three numbers cannot be summed. The short-term meter resets, but the weekly meter keeps accumulating; once the flagship meter fills up, flagship usage stops. Thus, one cannot simply multiply the 5-hour capacity indefinitely to estimate monthly allowances. The chart uses approximately 4.345 weeks per month, converting Fable’s 57 million weekly input tokens into roughly 247 million monthly input tokens; this is an isolated input-only benchmark example, which is distinct from the total token figures earlier that incorporated heavy cache reads.

Finally, blend the four deduction rates according to specified workloads and multiply by API prices. In the Opus example, dividing the $200 monthly fee by the blended subscription cost amortization of ~$0.00699 per million tokens yields roughly 28.6 billion tokens; valuing this at the blended API rate of ~$0.4099 per million tokens for that same mix yields ~$11,726. It is the ratio between subscription deduction mechanics and API rate cards that determines this discount.

Context window lengths must also remain comparable. Certain models incur higher API rates past specific context thresholds, so the tests kept requests below higher-priced tiers; one cannot use long-context list prices to evaluate short-context runs. The original API screenshots display two price tiers for short and long contexts, which is precisely the variable that needed to be held constant.

View the 7 original charts on quota measurement and conversion

Bringing it back to your monthly subscription: if you frequently run heavy agent workflows on Opus and consume a meaningful portion of your allowance, these benchmarks offer a definitive verdict—Claude’s workhorse model subscription remains by far the larger compute subsidy. If you rely exclusively on flagships, or prioritize uninterrupted burst sessions without short-term throttling and demand raw output speed, you must factor in Fable’s sub-cap alongside OpenAI Pro’s pacing and throughput.

Tracking task completion rates and rework time on your own recurring workloads, while verifying whether prompts actually hit the cache (Chinese), will tell you far more than treating theoretical API equivalents as money in the bank. Understanding how providers grant subsidies—and how they claw them back—clarifies what you are truly paying for: raw intelligence, aggregate quota volume, or reduced waiting time.