The cheapest Claude API is usually not a different vendor — it is the same one with prompt caching turned on and a smaller model doing the routine work. Those two levers are several-fold each, they apply without changing anything about who you buy from, and they are missing from every result on this search, which instead argues about subscriptions versus API access. This page puts the levers in the order of how much they actually save, which means our own product is fourth.
Bias declared: Kunavo is one of the gateways in lever four. Putting it last is the honest ordering rather than a pose — if you act on levers one and two and stop reading, you will have captured most of the available saving.
The levers, in order of size
| # | Lever | Applies when | Vendor change needed |
|---|---|---|---|
| 1 | Prompt caching | Requests share a long identical prefix | No |
| 2 | Smaller model for routine calls | Not every call needs a frontier model | No |
| 3 | Subscription instead of API | Interactive use you can saturate | No — different product, same vendor |
| 4 | A gateway listing under published rates | You pay public list today | Yes |
1. Prompt caching — the biggest and most ignored
Prompt caching bills repeated prefixes at a fraction of fresh input. On Kunavo, Claude Sonnet 5 reads cached input at $0.20 per 1M against $2.00 for fresh input — and the pattern it rewards is precisely what agent and coding workloads do: the same long system prompt, the same document, the same codebase context, on every turn.
If your requests have that shape and caching is off, this is the cheapest change available to you and it does not involve moving anything. If your requests are all one-off and share no prefix, this lever does nothing — check before assuming either way.
2. Model choice — a bigger gap than any vendor difference
Claude Haiku 4.5 at $0.40 / $2.00 per 1M, Claude Sonnet 5 at $2.00 / $10.00, Claude Opus 5 at $2.00 / $10.00. The distance between tiers is larger than the distance between suppliers of the same tier, so routing classification, extraction and summarisation to a smaller model saves more than any switch of provider will. The discipline is having evals good enough to know which calls can drop a tier — without them this becomes a quality cut rather than a cost cut. Cost optimization in practice works through the routing patterns.
3. Subscription instead of API, when it fits
The loudest claim on this search is that a Claude plan is many times cheaper than API access, sometimes quoted at 36×. It is arithmetically fine under an assumption its sources rarely state: a subscription used to saturation, against the token spend of the same workload. For interactive daily coding that regularly reaches the usage window, it genuinely holds, and plan pricing is the input.
It stops holding for bursty months, unattended and CI runs, work that must continue after a window is spent, and per-project spend records. Which source to point Claude Code at works that decision through in full; for what the published rates themselves are, see Claude API pricing.
4. A gateway that lists under published rates
A reseller that buys capacity can list below the published price. Kunavo lists Claude Sonnet 5 at $2.00 / $10.00 per 1M against Anthropic's $2.00 / $10.00 — at list on this model today, pay-as-you-go from $10, with caching supported so lever one still applies. The per-model rates are published rather than expressed as a headline percentage, because a catalog discounts unevenly and a single figure would not be checkable.
When this lever does not apply: if you hold a negotiated enterprise agreement with Anthropic, that rate almost certainly beats any public reseller price, and nothing here restores a discount you already have.
A warning this results page does not carry
Marketplace listings selling Claude API keys for a flat price appear on this very search. Buying one means using someone else's account credentials. That violates Anthropic's terms, the key can be revoked at any moment with no recourse and no refund, you cannot know how the underlying account was funded, and every prompt you send transits an account you do not control — including whatever proprietary code or customer data is in your context window.
A key that costs suspiciously little is cheap for a reason, and the reason is generally that somebody else is paying for it. Every legitimate lever above is available without trusting an anonymous seller, and the first two are free.
FAQ
What is the cheapest way to use the Claude API?
In order of how much they usually save: turn on prompt caching if your requests share a long prefix, because cached input is billed far below fresh input; use the smallest model that passes your evals, since the gap between tiers is several-fold; move to a Claude subscription instead of the API if your usage is interactive and steady enough to saturate a plan; and buy through a gateway that lists Claude under Anthropic's published rates. The first two need no vendor change and are where most of the savings are.
Is the Claude subscription cheaper than the API?
For saturated interactive use, yes — substantially, and that is the basis for the "up to 36x cheaper" claims. The comparison assumes you use the plan to its usage limits. It stops holding when your usage is bursty, when you need unattended or CI runs, when work must continue after a plan's usage window is spent, or when you need per-project spend records. Those cases are what per-token billing is for, and the two can be combined on the same machine.
How much does prompt caching save on Claude?
Cached input is billed far below fresh input — on Kunavo, Claude Sonnet 5 reads cached input at $0.20 per 1M against $2.00 for fresh input. It applies when requests share a long identical prefix: a big system prompt, a document, a codebase context. That is the normal shape of an agent or coding workload, which is why prompt caching is usually the single largest lever on the bill, ahead of any choice of vendor.
Are cheap Claude API keys sold on marketplaces safe?
No, and it is worth being direct about it: buying an API key from a marketplace listing means using someone else's account credentials. That violates Anthropic's terms, the key can be revoked at any time with no recourse and no refund, you have no idea how the account was funded, and anything you send through it transits an account you do not control. A key that costs suspiciously little is cheap for a reason. Legitimate ways to pay less exist — caching, model choice, a subscription, or a gateway that publishes its per-model rates — and none of them requires trusting an anonymous seller.
Which Claude model is cheapest?
Claude Haiku 4.5 is the cheapest of the current family at $0.40 / $2.00 per 1M tokens on Kunavo, against Claude Sonnet 5 at $2.00 / $10.00 and Claude Opus 5 at $2.00 / $10.00. The gap between tiers is larger than the gap between vendors, so routing routine calls to a smaller model moves a bill more than switching provider does.
Can a gateway be cheaper than Anthropic directly?
It can, because a reseller that buys capacity can list below the published rate — Kunavo lists Claude Sonnet 5 at $2.00 / $10.00 per 1M against Anthropic's $2.00 / $10.00, which is at list for this model today. It is not cheaper in every case: if you hold a negotiated enterprise agreement with Anthropic, that rate almost always beats a public reseller price, and no gateway can restore a discount you already had.