What a coding plan actually delivers

A dated snapshot of three subscription plans and the five ways you can buy them. It knows nothing about your usage until you type it, and nothing you type leaves this page.

Only models all three plans list. Their rates differ by vendor, and each option uses its own vendor's numbers.

Output tokens cost several times what input tokens cost, so this bar moves the answer the most.

Share of input tokens served from cache. Cache reads are the cheapest rate on every plan.

Total tokens in and out, in millions. This is the demand line on the chart.

Your pacing

These three decide how much of the allowance is reachable at all. They are tagged assumed until you move them, and the window verdict stays withheld until all three have been touched.

A plan caps how much of the allowance one five-hour window can hold, so how many separate sittings you take in a day changes how much you can reach at all.

The answer

At 65.0M tokens a month, Command Code Go is the cheapest option that covers it: $1.00 a month for 69.8M tokens of reachable allowance. Pacing is still assumed (20 active days, 2h sessions, 1 sitting a day), so the window verdict is withheld: move all three pacing controls to unlock it.

Pacing is assumed until you move all three pacing controls. Until then the window verdict is withheld, because "the binding limit is the weekly window" is a claim about your rhythm and not about the plan.

Cost against what you can actually spend

Each line is one option: a fixed monthly cost, drawn only as far as its reachable allowance goes. Where a line stops early, the rest of what you paid for cannot be reached at this pacing, and the label past the stop is that dead allowance.

All five options

Option Cost a month Tokens covered Binding limit Cost per 1M tokens Beaten above

Cheapest first. "Cost per 1M tokens" is the price of using the whole reachable allowance at this pacing, so it is the value of the deal rather than the price of a token, and it moves when the pacing does.

What would have to be true

The same numbers read as conditions rather than as a ranking. No slider moves on its own.

Assumptions and sources

This tool reports the capacity reachable at the pacing you enter, assuming you fill every five-hour window you touch. It does not project the rate you are actually burning now, and it prints no realised-waste percentage: that is a different question, and the one it answers is whether the plan's window shape can deliver the allowance you would be paying for.

Where these numbers come from

Transcribed from those pages into rates.json, which is the source of record for every number on this page, and which carries the date as_of: 2026-11-05.

none, and permitted by the 2026-11-05 ruling: no browser existed on the host at capture time A capture record is the evidence of record: it says what was read and when, and it cannot prove a page still reads that way today. Your verification path is the source link plus the as_of date.

This is a dated snapshot, taken 2026-11-05. Vendors change rates and window shapes without notice, and this page is not monitored: nothing here refreshes itself, alerts anyone, or tries to stay true. If a number looks wrong, the source links above are where to check it.

Questions this file closed, and how
ItemWhat conflictedClosed as
GOAT allowance "up to $100 / 10x" (docs) vs "$70 credits / 7x" (launch announcement) no band; per-model published values used. $100 is the plan maximum, reached only by the boosted models (GLM-5.3 Flash, DeepSeek V4.1 Flash, Ling 3.1 Flash). $70 is the allowance of GLM-5.2 / Tencent Hy3 / Qwen 3.8 27B, which is where the older "70 credits" figure came from.
CC Go allowance flat $10 (pre-2026-09-28 subscribers) vs per-model $6..$10 per-model published values used; a legacy flat-$10 subscriber is not modelled (the page states the capture date and the plan shape it assumes)
MiniMax M3 50% off on Command Code vs <=/>512K tiers on OpenCode no band. Command Code reaches it by a 50% deal collapsible across both context tiers; OpenCode reaches the same numbers by pricing the <=512K tier there and doubling above 512K (context_tier). The equivalence is a coincidence of the deal and is declared as one.
DeepSeek V4.1 Flash / V4 Pro peak/off-peak peak vs off-peak (01-04 and 06-10 UTC, Mon-Fri) on both vendors off-peak is the base rate set; the peak set is carried in `peak` with the windows attached
context-tier breakpoints (GPT-6 Luna, Claude Haiku 5.5, Qwen 3.8 Max) OpenCode doubles the rate above 272K / 100K / 256K; Command Code publishes one row base tier is the rate set; the above-threshold set is carried in `context_tier` for the page. Qwen 3.8 Max is not tiered on either vendor at capture time; Qwen3.7 Plus is, and closed.
OpenCode Go Plus / Teams multipliers unstated Go Plus ($40) and Teams ($20/member) are not modelled and no output references them
Claude models on CC Go "available on every plan" is asserted for some models only Claude Haiku 5.5 and Claude Sonnet 5.5 are GOAT-and-above only; Claude is excluded from the intersection. "Available on every plan, including Go" is printed on the Command Code page for GPT-5.6 Luna, Grok 4.5 and the Muse Spark Contributor models only, and those three are in the intersection.