Gumloop's own docs say bring-your-own-key gets you 50% off credits. Read that again.

Disclosure before anything else: we make BizOps Pilot, which competes with Gumloop. You should read this the way you would read any piece written by someone with a stake in your conclusion. What we can offer in exchange is that the central fact here comes from Gumloop's own documentation, that Gumloop's genuine strengths get stated plainly and without a "but" bolted on, and that we will tell you where Gumloop beats us. It does, in at least one measurable way.

Here is the fact, from Gumloop's own AI models documentation:

"Provide your own provider API key to cut model costs. BYOK gives you 50% off AI model credits in both agents and workflow AI nodes, and it also applies to voice transcription."

It requires the Pro plan or higher, and works with an OpenAI, Anthropic, Google AI, Perplexity or SpaceXAI account. Image generation is a flat 30 credits per image, 15 with BYOK. Audio transcription runs roughly 1 to 2 credits per minute, also halved with BYOK.

Gumloop's documentation carries no publication date, so we cannot tell you when this was written or last changed. We read it on 22 July 2026. That is the honest boundary of what we know about its timing.

What that arrangement actually is

Walk the money.

You go to Anthropic. You create an account, put a card on it, and generate an API key. You paste that key into Gumloop. From then on, when a workflow calls a model, the tokens are billed to you by Anthropic at Anthropic's prices. Gumloop does not buy those tokens. Gumloop does not resell those tokens. Gumloop is not in that transaction at all.

And Gumloop still charges you 50% of the credit meter for the call.

So the token cost lands on your Anthropic invoice, and half the credit cost lands on your Gumloop allowance. Two bills, one model call. That is what "you pay twice" means, and we did not have to characterise anything to get there. It falls out of their own sentence.

Now the part most competitor blogs would leave out

A 50% credit charge on a call you paid for yourself is defensible on its face, and we are not going to pretend it is theft.

Gumloop is still doing real work on that call, and real work costs them real money. They run the orchestration. They hold the workflow state. They handle the retries when the provider 529s. They queue it, schedule it, log it, surface the trace when it goes wrong, and keep the thing running when you are asleep. None of that is free, and none of it goes away because the tokens moved onto your invoice. A vendor that charged you exactly nothing for a BYOK model call would be giving away the orchestration that is arguably their entire product.

So the honest version of the criticism is not "Gumloop charges for something worthless." It is narrower than that, and it is this: the credit meter is denominated in tokens, and after BYOK the tokens are not theirs.

Per Gumloop's own tokens-and-costs material, credits are calculated on two things: tokens analyzed (every input and output token the AI processes, with output costing "2-4x more expensive" than input), and tool calls to external data sources. Tools that stay inside Gumloop, such as sending a Slack message or reading a Google Doc, are free.

Read that back against BYOK. The meter is priced primarily off token volume. Once you bring your own key, token volume is a cost you are already paying in full, somewhere else. Charging half a token-denominated meter for tokens you bought is not a charge for orchestration. It is a charge for orchestration that happens to scale with the one variable Gumloop no longer pays for.

If the fee were for orchestration, it would look like orchestration: a flat charge per run, or per active workflow, or per seat. Something that tracks their costs rather than yours. Instead it tracks your token consumption, at half rate, forever. That is the actual objection, and it is a design objection, not an accusation.

Why the variance is the real problem, not the 50%

Fifty percent of a small, predictable number is a small, predictable number. Nobody would write an article about that.

The trouble is that the number is neither small nor predictable, and Gumloop is unusually candid about this. From their credits documentation:

"Agent credit costs are variable. The same agent might cost 2 credits for a quick question and 200 for a deep research task."

That is a 100x spread on the same agent, published by the vendor, in their own docs. We want to be clear that we think publishing it is to their credit. Most vendors in this category would bury that.

Their docs also name the specific failure mode, and it is a good piece of technical writing:

"There's a hidden cost to watch for: when a tool returns data, the model has to read the entire response to use it. A tool that returns a large JSON payload... can generate thousands of input tokens just from the result. Sometimes analyzing the response costs more than the tool call itself."

Now put the two things together. The meter has a 100x range you cannot forecast. BYOK halves it. Half of an unforecastable number is still an unforecastable number. You have taken on an API key, a provider account, a second invoice, and a Pro subscription, and in exchange the thing you could not predict got 50% smaller and stayed exactly as unpredictable.

That is the argument in one line. Not the 50%. The variance that the 50% does not touch.

One user's report, and what it is worth

On Gumloop's own forum, a thread dated 12 January 2026 titled "GitHub MCP burning hundreds of credits?" has a user reporting that an agent searching a GitHub repo via MCP to generate a markdown file "cost me 900 credits." They say an earlier test "burned everything" on their first account, and they made a second one.

Gumloop's free tier includes 5,000 credits per month, read from their pricing page on 22 July 2026. If that run is representative, it is about five runs before a free account is finished for the month.

What that datapoint is worth, stated honestly: it is one user, on a forum, describing their own experience. We did not reproduce it. We do not know their workflow, their model, their repo size, or whether they had a loop bug. It is a plausible worst case, not a benchmark, and anyone who quotes it as "Gumloop costs 900 credits per run" is misusing it. We include it because it is dated, specific, and consistent with what Gumloop's own docs say can happen. It is corroboration of a documented variance problem, not proof of a typical cost.

Everyone offers BYOK now. That is the point.

If you take one thing from this piece, take this one, because it is the thing that reframes the whole category.

Bring-your-own-key is not a differentiator any more. It is table stakes. Look at what the market actually shipped:

Zapier's position is worth reading closely, because it is the clearest. Per Zapier's own help article on AI model tier pricing, dated 15 June 2026, AI steps are priced by model tier at Standard 1x, Advanced 3x and Premium 5x, new steps default to Advanced (3x), and you can "connect your own AI account (1x)" instead. The market leader documents bring-your-own-key as a first-class 1x option against its own 3x default. That is to Zapier's credit and we will say so. But Zapier still counts every task afterwards, including 2 tasks per successful MCP tool call per their MCP usage doc dated 4 June 2026.

Make announced the same shape on 13 August 2025, in their own community post moving from operations to credits: "Make AI Agents: Available on all plans with Make's AI Provider. Building with your own connection will require the Pro plan or higher."

So three of the four take your key and keep the meter running. Gumloop takes your key and keeps half the meter running. The differences between them are matters of degree.

The exception is n8n self-hosted, and it is a real exception. n8n's own docs state that when self-hosting, "You'll bring your own API key, so no credit limits apply." That is genuine, uncapped BYO-key with nothing counted, and the Community Edition is free under the Sustainable Use License, carried in n8n's own repository. We will come back to this, because it beats us.

Gumloop is good, and it is cheaper than we are

This section is not a courtesy. Skipping it would make everything above less believable, not more.

Gumloop Pro is $37 a month on their pricing page, and that is the genuine monthly price, not an annual number dressed up as monthly. We are $39. They are cheaper than us, and cheaper again if you pay annually, which brings their price to $30 a month for the same tier. Their Pro tier includes unlimited seats, which is a genuine break from how this category prices and is straightforwardly better than most of it. It includes hosted MCP server capability. Both their free and Pro tiers advertise unlimited agents and unlimited flows. All of that is read off their pricing page on 22 July 2026.

They raised a $50M Series B led by Benchmark, announced 12 March 2026 per TechCrunch, with SiliconANGLE reporting participation from Shopify Ventures and Y Combinator and total outside funding of $70M. They are a serious, well-capitalised, well-run company building an AI-native product, and our own internal research names them the competitor to watch. Not the competitor to dismiss. The one to watch.

They are also more honest in their documentation than most of this category manages to be. The "2 credits or 200" line and the "analyzing the response costs more than the tool call" warning are both things a less scrupulous vendor would never have written down. We built a section of this article out of their candour, which is a slightly uncomfortable thing to admit, and worth admitting anyway.

If you want an AI-native workflow builder, with unlimited seats, from a company with Benchmark's money behind it, for two dollars less than we charge, Gumloop is a completely reasonable purchase and we are not going to insult your judgment by suggesting otherwise.

And n8n self-hosted beats us more badly than Gumloop does. BYO-key with nothing metered, no limits, free software, for the price of a small VPS, which is less than $39. If you are comfortable running a container and keeping it patched, we do not win on price and we do not win on the metering argument either. We lose both. What we offer against n8n is that there is no server to run, nothing to upgrade, no Docker, no maintenance, MCP support out of the box, and a founder who answers his email. That is a hosting-and-time argument, not a pricing argument, and we would rather name the fight we are actually in.

What we do differently

One thing. We are going to say it and stop, because it is the only claim we have that survives the section above.

We do not meter.

You bring your key, you pay Anthropic or OpenAI directly at their prices with nothing added by us, and then we count nothing. Not tokens. Not runs. Not tool calls. Not MCP calls. There is no allowance to exhaust, no percentage retained, and no multiplier. Pro is $39 a month, flat. There is a Founding option at $249 once, hard-capped at 300 seats and enforced at checkout. There is a free beta tier.

The free tier bounds how many Pilots and Systems run at the same time, three and one, and that is a concurrency ceiling, not a volume meter. An active Pilot on the free plan runs an unlimited number of times, as often as its trigger fires, forever. Runs are unlimited on every plan including free. The only limits that exist at all are structural fair-use ones, disclosed in our terms, and none of them counts volume.

That is it. That is the entire argument. Everything else about us is arguable and some of it is worse than the alternatives.

Two more things you are owed before you weigh that. We are pre-launch and small. No customer counts, no revenue figures, no testimonials, because we do not have any and we are not going to invent them. And a flat price means the founder absorbs the variance that the metered vendors push onto you, which is the point, and is also a thing you should factor into how durable you think we are.

The claim we are not going to make

One more, in the spirit of the rest.

There is a third-party review of Gumloop dated 21 April 2026 claiming Gumloop's pricing is a free tier with 1,000 credits, Starter at $97/month and Pro at $297/month. Those numbers are wrong. They contradict Gumloop's own pricing page on every single point, including the free-tier credit count, which is 5,000 and not 1,000.

We are flagging it because it would be useful to us if it were true, and it is not. Gumloop's page says $37, which is two dollars below ours. If you find that article, or one of the many posts recycling from it, you have learned something about the article and nothing about Gumloop.

What we could not verify

• Gumloop's BYOK documentation carries no date. We read it on 22 July 2026. We cannot tell you when it was published or last revised, and we are not going to imply a date we do not have. The same applies to their credits docs and their tokens-and-costs material.
• The 900-credit forum report is one user's account, undated beyond the thread date of 12 January 2026, and unreproduced by us. Treat it as a plausible worst case that matches the vendor's documented variance, not as a typical cost.
• We have not audited Gumloop's cost structure, so our claim that a token-denominated meter is the wrong instrument for an orchestration fee is an argument from their published pricing design, not from their books. There may be a cost we cannot see that makes 50% correct. We doubt it. We cannot prove it.

So what should you do

If you are already on Gumloop Pro and your workflows are stable and modest, turn BYOK on. Half a meter is better than a whole one and the fact that we find the design objectionable does not make the discount fake. Take it.

If you are evaluating, do the boring thing that beats every comparison post including this one. Take your three highest-volume workflows. Count the model calls. Note which ones return big JSON payloads into a model's context, because Gumloop's own docs tell you that is where the credits go. Then price that shape against a 20,000-credit allowance at half rate, and ask yourself honestly whether you can predict the answer within a factor of ten. If you can, Gumloop at $37 with unlimited seats is a good product and a fair price and you should buy it.

If you cannot, that is not a failure of your arithmetic. It is the thing their docs already told you: 2 credits, or 200. BYOK makes it 1, or 100.

We built the other option, which is that nobody counts. That is the only reason we are in this conversation, and we would rather you picked us for that reason or not at all.

Sources

Gumloop pages in this list were re-read on 22 July 2026. Other vendor pages were read on 16 July 2026 and were not re-fetched this pass. Vendor documentation pages in this category are generally undated, which is noted where it applies.

• Gumloop BYOK, "50% off AI model credits", Pro-plan gate, 30 credits per image and 15 with BYOK: https://docs.gumloop.com/core-concepts/ai_models (undated vendor doc)
• Gumloop credit variance, "The same agent might cost 2 credits... or 200": https://docs.gumloop.com/core-concepts/credits (undated vendor doc)
• Gumloop credits as tokens analyzed plus tool calls, output 2-4x input, internal tools free, "analyzing the response costs more than the tool call itself": https://university.gumloop.com/ai-fundamentals/tokens-and-costs (undated vendor doc)
• Gumloop pricing, free tier 5,000 credits, Pro $37/mo with 20,000+ credits, unlimited seats, MCP server hosting: https://www.gumloop.com/pricing. Confirmed live via the toggle on 22 July 2026: $37/mo is the true monthly price, and $30/mo is the annual-equivalent price for the same tier.
• Gumloop forum, 900 credits on one GitHub-MCP run, 12 January 2026, single user report: https://forum.gumloop.com/t/github-mcp-burning-hundreds-of-credits/3553
• Gumloop $50M Series B led by Benchmark, 12 March 2026: https://techcrunch.com/2026/03/12/gumloop-lands-50m-from-benchmark-to-turn-every-employee-into-an-ai-agent-builder/
• Gumloop $70M total outside funding, Shopify Ventures and Y Combinator participating, 13 March 2026: https://siliconangle.com/2026/03/13/gumloop-reels-50m-ai-automation-platform/
• Zapier AI model tiers, Standard 1x / Advanced 3x / Premium 5x, default Advanced, "connect your own AI account (1x)", 15 June 2026, confirmed unchanged 22 July 2026: https://help.zapier.com/hc/en-us/articles/46425475442829-AI-by-Zapier-model-tier-pricing
• Zapier MCP, each successful tool call = 2 tasks, 4 June 2026: https://help.zapier.com/hc/en-us/articles/45645738385805-How-Zapier-MCP-usage-works
• Make operations to credits, own-connection gated to Pro and above, 13 August 2025: https://community.make.com/t/introducing-credits-a-new-system-of-billing/89480
• n8n Sustainable Use License, free self-hosted Community Edition: https://github.com/n8n-io/n8n

Pricing in this category changes without notice, and undated vendor docs can change without any signal at all. Every number above is a snapshot with a date on when we looked, not a promise about today. Re-check before you buy anything, including ours.