Blog · 2026-08-05 · Vynaris Team

Claude Max vs API: 119 Sessions Is the Break-Even at 90% Cached Input

At 90% cached input, Claude Max 5x equals 119 Fable 5 sessions, 237 Opus 5 sessions, or 592 Sonnet 5 sessions at API rates.

Claude Max 5x reaches $100 of API-equivalent value after 119 one-hour coding sessions on Claude Fable 5, 237 on Claude Opus 5, or 592 on Claude Sonnet 5 at its introductory price. Those are economic break-evens, not promised plan capacities. Anthropic publishes relative limits, not a fixed token allowance. Prices verified 2026-08-05.

TL;DR

Max 5x costs $100 per month. Max 20x costs $200. Anthropic describes them as 5x and 20x Pro usage per five-hour session, with separate weekly limits.
On Anthropic's own one-hour coding shape of 50,000 input and 15,000 output tokens, a 90%-cached session costs $0.845 on Fable 5, $0.4225 on Opus 5, $0.169 on Sonnet 5 through 2026-08-31, and $0.0845 on Claude Haiku 4.5.
Max 5x equals 119 Fable sessions or 1,184 Haiku sessions at API rates. Max 20x needs twice those counts because it costs twice as much.
The subscription wins only if you can consume the allowance. Session and weekly caps mean this table cannot prove how many sessions Max will permit.

Verdict table

Model                            API $/1M in / cached / out  Cost per 90%-cached session  Sessions to $100  Sessions to $200
-------------------------------  --------------------------  ---------------------------  ----------------  ----------------
Claude Fable 5                   $10 / $1 / $50              $0.8450                      119               237
Claude Opus 5                    $5 / $0.50 / $25            $0.4225                      237               474
Claude Sonnet 5, through Aug 31  $2 / $0.20 / $10            $0.1690                      592               1,184
Claude Sonnet 5, from Sep 1      $3 / $0.30 / $15            $0.2535                      395               789
Claude Haiku 4.5                 $1 / $0.10 / $5             $0.0845                      1,184             2,367

The decision is blunt. Buy Max for heavy, interactive use where interruptions have a cost. Use the API when volume is low, work is automatable, or you need exact control over the model routing, prompt caching, and spend cap. Do not call 119 sessions a guaranteed Max allowance. It is the point where 119 Fable sessions would have cost $100 on the API.

What we calculated

Anthropic's pricing documentation gives a one-hour coding example with 50,000 input tokens and 15,000 output tokens. We use that public shape rather than inventing a tiny chat prompt and calling it coding.

The brief asks what happens at a 90% cache hit rate. We therefore split the 50,000 input tokens into 5,000 fresh and 45,000 cache reads. The 15,000 output tokens stay unchanged. The formula is:

session cost =
  5,000 x fresh-input rate / 1,000,000
  + 45,000 x cache-read rate / 1,000,000
  + 15,000 x output rate / 1,000,000

break-even sessions = ceiling(plan price / session cost)

Every count is editable in claude-max-vs-api-pay-per-token-break-even-coding-agent-cost-math.py. The prices come from Anthropic's live pages. We did not use Claude plan telemetry, private account data, or a community invoice as a price receipt.

Assumption             Value                Source or status
---------------------  -------------------  ------------------------------------------------------
Coding session input   50,000 tokens        Anthropic pricing worked example
Coding session output  15,000 tokens        Anthropic pricing worked example
Cache hit rate         90%                  Explicit scenario assumption
Max 5x price           $100/month           Anthropic Max support page
Max 20x price          $200/month           Anthropic Max support page
Plan allowance         Not fixed in tokens  Anthropic publishes relative session and weekly limits

The API-equivalent session bill

Fable 5 costs $0.845 for the assumed session. The arithmetic is $0.05 of fresh input, $0.045 of cache reads, and $0.75 of output. Output is 88.8% of the bill. Better input caching cannot remove it.

Opus 5 is exactly half that price on every token category, so the same session costs $0.4225. Sonnet 5 costs $0.169 during its introductory window. Its rates rise 50% on 2026-09-01, taking the session to $0.2535. Haiku 4.5 lands at $0.0845.

All five rows save 32.4% versus an uncached session. That uniform saving is not a coincidence. Each model's cache-read rate is one tenth of its fresh-input rate, and only the input side receives the discount. The output-heavy session prevents a 90% cache hit from becoming a 90% bill reduction. We showed the same trap from a different angle in the reasoning-token tax analysis: expensive output can dominate even when the visible answer looks modest.

Use the calculator at this point to replace the 50,000/15,000 shape with one session from your own usage export. The session shape matters more than the plan label.

Horizontal bar chart of API-equivalent sessions needed to reach the $100 Claude Max 5x price at 90% cached input. Fable 5 needs 119 sessions, Opus 5 needs 237, Sonnet 5 needs 592 during introductory pricing and 395 after September 1, and Haiku 4.5 needs 1,184. — Claude Max 5x economic break-even at 90% cached input. Log scale. Source: Anthropic pricing and plan pages, verified 2026-08-05. Session shape: 50k input and 15k output.

Why the model changes the answer by 10x

The subscription price does not change when you select a model. The API bill does. That creates the 119-to-1,184-session range in the table.

Fable 5 consumes plan capacity under a special rule. Anthropic's plan comparison says Fable access on Max is limited to 50% of the weekly allowance. That makes the 119-session economic break-even easier to reach in dollars but potentially harder to reach inside the subscription. The priciest API model is not automatically the easiest subscription win.

Haiku creates the opposite case. The API is so cheap that a developer must run 1,184 of these assumed sessions before paying $100. At one session per working day, the API stays far below Max. At continuous agent volume, a consumer subscription is the wrong unit anyway. Use an API with a budget and an audit trail.

Sonnet 5 shows why the date stamp matters. Through 2026-08-31, $100 buys 592 API-equivalent sessions. From 2026-09-01, it buys 395. Nothing about the session changed. The input, cache-read, and output prices all rise 50%. We published the full transition math in the Sonnet 5 introductory-pricing analysis.

The limit math Anthropic does not publish

Anthropic says Max 5x provides five times Pro usage per session and Max 20x provides twenty times Pro usage. A session window lasts five hours. The plan also has two weekly limits: one across all models and one for Sonnet models. Anthropic may apply other model, feature, weekly, or monthly caps.

That is enough to compare tiers. Max 20x offers four times the relative session usage of Max 5x for twice the price. If you repeatedly exhaust Max 5x and can use the extra capacity, Max 20x has the better allowance-per-dollar ratio.

It is not enough to convert a plan into a fixed token bucket. Message length, conversation length, model choice, files, and features affect usage. A table claiming “Max includes N million tokens” would be invented precision. We will not print it.

The correct test is empirical and account-specific:

Export one representative week of API usage or record the token shape of your coding sessions.
Compute the API-equivalent dollars with the current price sheet.
Compare that value with $100 and $200.
Check whether Max actually lets you finish the same work before its session or weekly limits stop you.
Count usage credits separately. Anthropic bills those at standard API rates after plan limits, so they are not free subscription capacity.

Where the API still wins above break-even

The subscription can win on API-equivalent dollars and still be the wrong operating model.

The API gives you programmatic calls, deterministic per-request cost, per-user attribution, and hard budget controls. It also lets you send easy steps to Haiku and keep Fable for work that earns the premium. Our coding-agent cost report shows why a blended workflow rarely matches the cost of running every step on one model.

Max wins on convenience. Claude Code is included in paid plans and shares the account's usage pool. You pay a fixed monthly price and avoid metering each terminal turn until a limit interrupts you. That predictability has value for one human working interactively.

Do not share a consumer plan across production workers or use it as infrastructure. The API exists for workloads, service accounts, logging, and billing. A subscription exists for a person.

The honest tradeoff

This model prices sessions as if the API and Max produce identical useful work. They may not. Plan-side tools, feature access, rate limits, model availability, and hidden product optimizations can change the result. A $0.845 API-equivalent Fable session that finishes a task is cheaper than a “free” subscription session that hits a cap halfway through. The reverse is also true when Max absorbs long interactive work without extra credits.

Caching is another source of error. We assumed a 90% hit rate because stable coding-agent prefixes often repeat. If your agent rewrites its tool schemas or system prefix mid-session, the real rate falls. Our cache-write churn study shows how prefix instability turns cheap reads back into premium writes.

Treat the table as a break-even calculator, not a purchase verdict. Run your token shape, then test the plan's actual limits for a week.

FAQ

How many sessions make Claude Max 5x cheaper than the API? On the stated 50,000-input, 15,000-output session with 90% cached input, 119 Fable 5 sessions, 237 Opus 5 sessions, 592 Sonnet 5 sessions at introductory pricing, or 1,184 Haiku 4.5 sessions equal the $100 plan price.

Is Max 20x a better deal than Max 5x? It offers four times the relative per-session usage for twice the monthly price. That is better allowance per dollar if you can use it. Separate weekly and model limits still apply.

Does Claude Max include unlimited Claude Code? No. Claude Code draws from the same plan pool. Max has five-hour session limits, two weekly limits, and possible additional caps.

Can I convert the Max allowance into tokens? Not honestly from public information. Anthropic publishes relative usage, not a fixed token allowance. Session shape, model, files, and features affect consumption.

What happens after I hit a Max limit? Paid plans can enable usage credits. Anthropic says those credits are billed at standard API rates, so add them to the subscription price when comparing the invoice.

Sources

Anthropic Claude plans and pricing, captured 2026-08-05: Max from $100, 5x/20x relative usage, Fable weekly-limit rule, credits at API rates.
Anthropic Max plan support page, captured 2026-08-05: Max 5x $100, Max 20x $200, session and weekly limits.
Anthropic API pricing, captured 2026-08-05: model, cache, and output rates; 50k/15k one-hour coding example.
Hacker News Max-vs-API discussion, captured 2026-08-05: 116 points and 196 comments; demand signal only, not a price receipt.

Prices change. Re-run the attached script before buying a plan on this math.