OpenRouter alternative for Kimi K3: standing 10% off

If you are choosing a Kimi K3 endpoint for a production agent, the number that matters is not today's cheapest row. It is the rate you can put in next quarter's budget. In the hour we spent checking this piece, the cheapest rows on OpenRouter's K3 list moved. Our rate did not: $2.70 / $0.27 / $13.50 per million tokens, 10% off the official list, published as a standing rate rather than a promo window.

That is the pitch, and it is narrower than "cheapest". On the official-list tier, we are about 15% below OpenRouter's official K3 endpoint after its 5.5% top-up fee. On the fp4 tier, OpenRouter is cheaper today; those rows are also the ones that moved while we were writing. If you want the absolute floor and can accept the quantization and the price risk, that is a legitimate choice. If you want a rate you can budget, read the arithmetic below.

Disclosure: Wallaby Token sells API access to Kimi K3. OpenRouter figures are two snapshots from its public models API and pricing page on 28 October 2026, about an hour apart. Our rate is from our published pricing file, last checked 9 October 2026. Endpoint prices move; re-check before budgeting.

The floor moved while we checked

Same public endpoint list, two checks about an hour apart. Some rows moved down, some moved up; the point is not the direction. The point is that the cheapest row is not a contract. A budget built on it is a budget built on a number that can change between your invoice and your forecast.

We are not claiming the official list never moves. It does, and our price is 10% off whatever the list is. What is standing is the discount and the no-fee top-up, not a promise that the list stays frozen.

The like-for-like bill

The fair comparison is the official-list tier: OpenRouter's official Moonshot endpoint versus us, after OpenRouter's fee. Prices per million tokens, three workloads:

OpenRouter's fee is 5.5% on non-crypto credit purchases, with a $0.80 minimum; crypto is 5.0% with no minimum. The fee is added to the purchase: a $100 credit purchase costs $105.50, a $20 purchase costs $21.10, and a $100 budget buys $94.79 of credits. Wallaby absorbs the card processing fee: a $20 top-up credits $20, and the metered rate is the rate.

At $1,000 a month, the fee difference is $55; at $10,000, it is $550. That is the part of the bill that does not depend on which K3 row is cheapest today.

The fp4 tier

The fp4 tier is cheaper today. That is why the pitch is not "cheapest": on the same three workloads, the cheapest total row on OpenRouter's list is well below us:

Those totals include OpenRouter's 5.5% fee. They are also the rows that moved between our two checks, and the cheapest total row was a different provider on different workloads. If quantization is acceptable to you and you can re-shop the floor every month, OpenRouter is the cheaper buy today.

We are not going to pretend that gap away. What we sell is the other thing: one serving path, per-request thinking lines, a hard per-key dollar cap, and a standing discount instead of a moving floor.

What a standing rate buys

• A budget line that does not need re-shopping. The discount is 10% off the official list, ongoing; it has been the published rate since 15 September 2026. If the list moves, your rate moves with it by the same 10%.
• No top-up fee. The fee is 5.5% at OpenRouter, $0.80 minimum; ours is zero. At $20 that is $1.10; at $1,000 it is $55.
• Itemized thinking tokens. Every request is metered as input, cached input, output, and thinking, on separate lines. On a reasoning model, the thinking line is often the one you want to audit.
• A hard per-key dollar cap. You can hand a key to a teammate with a ceiling enforced in dollars, not tokens.

What you give up is real: no multi-model catalog, no cross-provider routing, and no ability to shop across 30 endpoints. We run two provider keys per model with request-level failover, but that is resilience inside one model, not a marketplace.

What OpenRouter does better

This is not a close call on breadth. OpenRouter lists 464 models in its public API today, one key, one bill, provider routing, and failover across hosts. If your stack runs Kimi K3 next to Claude, GPT, and two image models, a single gateway is the sane architecture, and the 5.5% buys consolidation.

It is also the better choice if you want to shop quantizations: 30 endpoints means you can pick fp4 for cost, fp8 for fidelity, or the official endpoint for the reference path, and switch per request. A single-endpoint provider cannot offer that.

Who should pick which

Pick OpenRouter if you run more than one model family, need one invoice across providers, or your workload can use fp4 and you are willing to re-check the floor.

Pick a direct K3 endpoint if K3 is the product, your bill is output-heavy, or you want a rate you can put in a budget without re-shopping it every month.

The third option is to do both: keep OpenRouter for breadth and fp4 experiments, and route the K3 production loop to a direct endpoint. That is the setup that usually makes sense for teams that care about both.

Try the standing rate

A new Wallaby account starts with $0.50 of trial credit, enough to run a real Codex or OpenCode loop before you top up. Run your own workload for a week, then compare the bill against your current official-endpoint bill. If it does not beat it, do not switch. The minimum top-up is $20, and the K3 rate is $2.70 / $0.27 / $13.50 per million tokens, 10% off the official list, with no top-up fee. Create an account.

For the full 18-provider price table, see Kimi K3 API pricing. For where K3 is available and what each route costs, see every access route compared. For the top-up fee math, see card payment and the $20 minimum.

Sources: OpenRouter public models API and pricing page, two snapshots on 28 October 2026; OpenRouter fee announcement (updated 11 June 2026); Moonshot AI pricing page; Wallaby pricing file, last checked 9 October 2026. Prices and endpoint counts change; re-check before committing budget.