# Tokens Are the Cheap Part

Steve Jamison · AI Engineer NYC 2026 · October 14

[Watch recording](https://www.youtube.com/watch?v=jiUC6IUMdvc&t=17029s)

[04:43:49](https://www.youtube.com/watch?v=jiUC6IUMdvc&t=17029s) Hi, I'm Steve Jamison, AI Research and Engineering Leader at Jellyfish.

[04:43:54](https://www.youtube.com/watch?v=jiUC6IUMdvc&t=17034s) I'm here to show you why tokens are the cheap part.

[04:43:57](https://www.youtube.com/watch?v=jiUC6IUMdvc&t=17037s) The research that I'll present is drawn from the Jellyfish platform,

[04:44:00](https://www.youtube.com/watch?v=jiUC6IUMdvc&t=17040s) the industry's largest set of engineering workflow data joined to AI telemetry data.

[04:44:07](https://www.youtube.com/watch?v=jiUC6IUMdvc&t=17047s) It's no secret that AI costs are climbing.

[04:44:10](https://www.youtube.com/watch?v=jiUC6IUMdvc&t=17050s) They're up 60x from this time last year.

[04:44:13](https://www.youtube.com/watch?v=jiUC6IUMdvc&t=17053s) The use cases are wide and varied.

[04:44:16](https://www.youtube.com/watch?v=jiUC6IUMdvc&t=17056s) The models and harnesses are getting much better.

[04:44:18](https://www.youtube.com/watch?v=jiUC6IUMdvc&t=17058s) The implications are exciting.

[04:44:20](https://www.youtube.com/watch?v=jiUC6IUMdvc&t=17060s) But discussions around budget and spend are not as much fun.

[04:44:25](https://www.youtube.com/watch?v=jiUC6IUMdvc&t=17065s) These conversations involve your peers and senior leadership and the board.

[04:44:28](https://www.youtube.com/watch?v=jiUC6IUMdvc&t=17068s) not always the easiest audiences, especially when you don't have the right data.

[04:44:34](https://www.youtube.com/watch?v=jiUC6IUMdvc&t=17074s) And it's real money these days. Output is up too. Merged PRs per engineer are up 43% over the same

[04:44:43](https://www.youtube.com/watch?v=jiUC6IUMdvc&t=17083s) one-year horizon. This is phenomenal. We've not seen productivity inputs like this,

[04:44:49](https://www.youtube.com/watch?v=jiUC6IUMdvc&t=17089s) especially on this timescale. There's plenty of research and anecdotal data to know

[04:44:55](https://www.youtube.com/watch?v=jiUC6IUMdvc&t=17095s) that AI and agentic coding are driving these throughput gains.

[04:44:59](https://www.youtube.com/watch?v=jiUC6IUMdvc&t=17099s) Heck, even our charts show the correlation.

[04:45:01](https://www.youtube.com/watch?v=jiUC6IUMdvc&t=17101s) Notice where the inflection points are.

[04:45:04](https://www.youtube.com/watch?v=jiUC6IUMdvc&t=17104s) AI spend increased in early 2026

[04:45:07](https://www.youtube.com/watch?v=jiUC6IUMdvc&t=17107s) when throughput gains start a steady increase of their own.

[04:45:12](https://www.youtube.com/watch?v=jiUC6IUMdvc&t=17112s) There's a tension with these two numbers, though.

[04:45:14](https://www.youtube.com/watch?v=jiUC6IUMdvc&t=17114s) How do you reconcile them?

[04:45:16](https://www.youtube.com/watch?v=jiUC6IUMdvc&t=17116s) Spend is up 60x.

[04:45:18](https://www.youtube.com/watch?v=jiUC6IUMdvc&t=17118s) Throughput is also up 43%.

[04:45:20](https://www.youtube.com/watch?v=jiUC6IUMdvc&t=17120s) Is that good? Bad?

[04:45:23](https://www.youtube.com/watch?v=jiUC6IUMdvc&t=17123s) 43% is a big number.

[04:45:25](https://www.youtube.com/watch?v=jiUC6IUMdvc&t=17125s) But so is 60x. How do you know if you're spending enough? Is there more productivity gain locked behind a checkbook? Should we increase spend 50% 2x more? Or how do you know if you spent too much?

[04:45:44](https://www.youtube.com/watch?v=jiUC6IUMdvc&t=17144s) I'm sure many of you have felt the tension here.

[04:45:46](https://www.youtube.com/watch?v=jiUC6IUMdvc&t=17146s) The CFO or board are asking for the receipts of the spend.

[04:45:50](https://www.youtube.com/watch?v=jiUC6IUMdvc&t=17150s) Can you show them?

[04:45:52](https://www.youtube.com/watch?v=jiUC6IUMdvc&t=17152s) Will merged PRs are up 43% satisfy them?

[04:45:56](https://www.youtube.com/watch?v=jiUC6IUMdvc&t=17156s) Also, what happens if you cap use and spend?

[04:46:01](https://www.youtube.com/watch?v=jiUC6IUMdvc&t=17161s) Are you stopping from setting tokens on fire?

[04:46:04](https://www.youtube.com/watch?v=jiUC6IUMdvc&t=17164s) Or are you stumping further efficiency gains?

[04:46:07](https://www.youtube.com/watch?v=jiUC6IUMdvc&t=17167s) You can ask peers at other companies what they're paying.

[04:46:10](https://www.youtube.com/watch?v=jiUC6IUMdvc&t=17170s) For many, it's still a moving target, though.

[04:46:13](https://www.youtube.com/watch?v=jiUC6IUMdvc&t=17173s) But how do you equate that to your org?

[04:46:17](https://www.youtube.com/watch?v=jiUC6IUMdvc&t=17177s) Surely, just comparing their bill to yours isn't adequate.

[04:46:20](https://www.youtube.com/watch?v=jiUC6IUMdvc&t=17180s) We've heard the reports of companies spending $100K, $500K, even millions on AI,

[04:46:30](https://www.youtube.com/watch?v=jiUC6IUMdvc&t=17190s) setting, then resetting their AI budgets as reality sets in.

[04:46:34](https://www.youtube.com/watch?v=jiUC6IUMdvc&t=17194s) Ultimately, these two numbers can't tell the story.

[04:46:38](https://www.youtube.com/watch?v=jiUC6IUMdvc&t=17198s) And even together, they missed the biggest line on the cost side.

[04:46:44](https://www.youtube.com/watch?v=jiUC6IUMdvc&t=17204s) These two numbers leave out the biggest cost of all, developer salary.

[04:46:49](https://www.youtube.com/watch?v=jiUC6IUMdvc&t=17209s) AI is making developers more productive on a scale we haven't seen before.

[04:46:54](https://www.youtube.com/watch?v=jiUC6IUMdvc&t=17214s) It isn't just another tool, it's intelligence.

[04:46:57](https://www.youtube.com/watch?v=jiUC6IUMdvc&t=17217s) This intelligence dramatically changes how much each dev can produce.

[04:47:02](https://www.youtube.com/watch?v=jiUC6IUMdvc&t=17222s) That's why we need to think about its cost differently.

[04:47:05](https://www.youtube.com/watch?v=jiUC6IUMdvc&t=17225s) Not often the side is overhead, but it's part of what it takes to produce each PR.

[04:47:11](https://www.youtube.com/watch?v=jiUC6IUMdvc&t=17231s) We're going to call this intelligence costs of a PR.

[04:47:18](https://www.youtube.com/watch?v=jiUC6IUMdvc&t=17238s) What you pay your engineers plus what you pay for AI, put them together in one cost bucket.

[04:47:26](https://www.youtube.com/watch?v=jiUC6IUMdvc&t=17246s) AI is about 2.5% of the cost to produce another unit of merged code.

[04:47:33](https://www.youtube.com/watch?v=jiUC6IUMdvc&t=17253s) Salary is the other 97.5%.

[04:47:37](https://www.youtube.com/watch?v=jiUC6IUMdvc&t=17257s) This is in line with other reported figures, and this data is from August 2026.

[04:47:44](https://www.youtube.com/watch?v=jiUC6IUMdvc&t=17264s) Here, tokens are the cheap part. But that's current relative costs, not dollars. Let's

[04:47:52](https://www.youtube.com/watch?v=jiUC6IUMdvc&t=17272s) look at what happened to the cost of merged PR over the last year.

[04:48:02](https://www.youtube.com/watch?v=jiUC6IUMdvc&t=17282s) The dollar cost of a merged PR fell 26% over the last year.

[04:48:07](https://www.youtube.com/watch?v=jiUC6IUMdvc&t=17287s) Not PRs created, but PRs merged.

[04:48:11](https://www.youtube.com/watch?v=jiUC6IUMdvc&t=17291s) Review time has factored in, and some level of quality gates has been cleared.

[04:48:16](https://www.youtube.com/watch?v=jiUC6IUMdvc&t=17296s) That decline in cost per merged PR includes a 60x increase in AI costs over the same time frame.

[04:48:23](https://www.youtube.com/watch?v=jiUC6IUMdvc&t=17303s) The change is nearly universal, too.

[04:48:25](https://www.youtube.com/watch?v=jiUC6IUMdvc&t=17305s) 83% of the companies we studied showed a drop in cost per PR.

[04:48:30](https://www.youtube.com/watch?v=jiUC6IUMdvc&t=17310s) What does this tell us?

[04:48:32](https://www.youtube.com/watch?v=jiUC6IUMdvc&t=17312s) The additional throughput gains are spreading salary thinner across PRs,

[04:48:39](https://www.youtube.com/watch?v=jiUC6IUMdvc&t=17319s) and this is happening faster than the AI costs are increasing.

[04:48:46](https://www.youtube.com/watch?v=jiUC6IUMdvc&t=17326s) What's the range in token spend?

[04:48:48](https://www.youtube.com/watch?v=jiUC6IUMdvc&t=17328s) We ranked companies by token usage, then grouped them into quintiles or fifths.

[04:48:55](https://www.youtube.com/watch?v=jiUC6IUMdvc&t=17335s) The range here is quite large. The lightest group, about 0.3% of the cost of a PR is AI. This represents about 60 million tokens per dev per month.

[04:49:11](https://www.youtube.com/watch?v=jiUC6IUMdvc&t=17351s) The top 20% of AI spend per dev, we're going to call the heaviest spend group.

[04:49:19](https://www.youtube.com/watch?v=jiUC6IUMdvc&t=17359s) There, 8.5% of the PR cost is AI.

[04:49:23](https://www.youtube.com/watch?v=jiUC6IUMdvc&t=17363s) This represents about 2.5 billion tokens per dev per month,

[04:49:28](https://www.youtube.com/watch?v=jiUC6IUMdvc&t=17368s) about two orders of magnitude more than the lightest group.

[04:49:33](https://www.youtube.com/watch?v=jiUC6IUMdvc&t=17373s) We'll use these groups on the next slides to see how token spend influences other metrics.

[04:49:40](https://www.youtube.com/watch?v=jiUC6IUMdvc&t=17380s) First, PR throughput measured as merged PRs per dev, split by the same AI spend quintiles on the previous slide.

[04:49:49](https://www.youtube.com/watch?v=jiUC6IUMdvc&t=17389s) The bottom four-fifths of spenders sell PR throughput gains of around 40%, give or take.

[04:49:58](https://www.youtube.com/watch?v=jiUC6IUMdvc&t=17398s) But the heavy spend group

[04:50:00](https://www.youtube.com/watch?v=jiUC6IUMdvc&t=17400s) improved more at a 67% increase.

[04:50:03](https://www.youtube.com/watch?v=jiUC6IUMdvc&t=17403s) More AI spend led to more throughput gains.

[04:50:11](https://www.youtube.com/watch?v=jiUC6IUMdvc&t=17411s) let's look at the cost per PR.

[04:50:13](https://www.youtube.com/watch?v=jiUC6IUMdvc&t=17413s) Positive results are seen again across the board.

[04:50:17](https://www.youtube.com/watch?v=jiUC6IUMdvc&t=17417s) Cost per PR is dropping for all AI usage groups.

[04:50:21](https://www.youtube.com/watch?v=jiUC6IUMdvc&t=17421s) However, again, the heaviest fifth, so the companies that spent the most, improved the most.

[04:50:30](https://www.youtube.com/watch?v=jiUC6IUMdvc&t=17430s) They saw a cost per PR decrease of 32% over the last year.

[04:50:37](https://www.youtube.com/watch?v=jiUC6IUMdvc&t=17437s) The other four groups saw about a 25% decrease in cost per PR.

[04:50:44](https://www.youtube.com/watch?v=jiUC6IUMdvc&t=17444s) Investing and spending more on AI fueled more of an improvement in cost per PR.

[04:50:50](https://www.youtube.com/watch?v=jiUC6IUMdvc&t=17450s) compared to peers who leaned in less.

[04:50:53](https://www.youtube.com/watch?v=jiUC6IUMdvc&t=17453s) So we saw that the same as we saw in the population average,

[04:51:00](https://www.youtube.com/watch?v=jiUC6IUMdvc&t=17460s) cost per PR changes are driven largely by PR throughput, not by AI spend.

[04:51:06](https://www.youtube.com/watch?v=jiUC6IUMdvc&t=17466s) The heavy spenders experience PR throughput gains of 67%

[04:51:11](https://www.youtube.com/watch?v=jiUC6IUMdvc&t=17471s) compared to about 40-ish percent for their peers.

[04:51:15](https://www.youtube.com/watch?v=jiUC6IUMdvc&t=17475s) Spreading the same salary across more PRs in the month

[04:51:19](https://www.youtube.com/watch?v=jiUC6IUMdvc&t=17479s) has more of an effect than the rising AI costs,

[04:51:23](https://www.youtube.com/watch?v=jiUC6IUMdvc&t=17483s) because tokens are the cheap part.

[04:51:28](https://www.youtube.com/watch?v=jiUC6IUMdvc&t=17488s) This is not to say that orgs should light tokens on fire

[04:51:33](https://www.youtube.com/watch?v=jiUC6IUMdvc&t=17493s) and spend with a banding.

[04:51:34](https://www.youtube.com/watch?v=jiUC6IUMdvc&t=17494s) Here, we're looking at tokens used on the horizontal axis

[04:51:37](https://www.youtube.com/watch?v=jiUC6IUMdvc&t=17497s) and cost per PR on the vertical axis.

[04:51:40](https://www.youtube.com/watch?v=jiUC6IUMdvc&t=17500s) There's a steep drop at the beginning

[04:51:42](https://www.youtube.com/watch?v=jiUC6IUMdvc&t=17502s) where relative small AI spend results in large gains initially.

[04:51:47](https://www.youtube.com/watch?v=jiUC6IUMdvc&t=17507s) As token usage increases, the cost per PR improvement decreases.

[04:51:53](https://www.youtube.com/watch?v=jiUC6IUMdvc&t=17513s) This is diminishing returns.

[04:51:55](https://www.youtube.com/watch?v=jiUC6IUMdvc&t=17515s) I want to stress, though, diminishing returns is not the same as no returns.

[04:52:02](https://www.youtube.com/watch?v=jiUC6IUMdvc&t=17522s) We're seeing that the top 10% are spending 8x the tokens compared to the median

[04:52:11](https://www.youtube.com/watch?v=jiUC6IUMdvc&t=17531s) and achieving a 25% lower cost per PR.

[04:52:17](https://www.youtube.com/watch?v=jiUC6IUMdvc&t=17537s) There will be a time when tokens aren't the cheap part, but by and large, for most companies out in the wild, we're not close to that threshold.

[04:52:28](https://www.youtube.com/watch?v=jiUC6IUMdvc&t=17548s) More tokens leads to more PRs, which leads to a lower cost of delivery.

[04:52:41](https://www.youtube.com/watch?v=jiUC6IUMdvc&t=17561s) So far, we've talked about companies.

[04:52:44](https://www.youtube.com/watch?v=jiUC6IUMdvc&t=17564s) Next, I'd like to take a tangent and talk about a study we did with developers.

[04:52:50](https://www.youtube.com/watch?v=jiUC6IUMdvc&t=17570s) Here we looked at developer weeks with heavy token use compared to the same developer weeks with lower token use.

[04:53:02](https://www.youtube.com/watch?v=jiUC6IUMdvc&t=17582s) Here we saw the same phenomenon just at the individual level.

[04:53:07](https://www.youtube.com/watch?v=jiUC6IUMdvc&t=17587s) Compare each developer's heavy agent skill weeks with their own weeks without skills.

[04:53:15](https://www.youtube.com/watch?v=jiUC6IUMdvc&t=17595s) They spend about twice as much on tokens, merge 27% more PRs, and reach PR costs about 18% less once salary and tokens are counted together.

[04:53:35](https://www.youtube.com/watch?v=jiUC6IUMdvc&t=17615s) there are two fair curiosities i get when i talk about this work first does faster work

[04:53:43](https://www.youtube.com/watch?v=jiUC6IUMdvc&t=17623s) result in worse work and second are the more prs

[04:53:50](https://www.youtube.com/watch?v=jiUC6IUMdvc&t=17630s) the throughput gains that we're seeing are they being driven by smaller chunks of work

[04:53:58](https://www.youtube.com/watch?v=jiUC6IUMdvc&t=17638s) First curiosity is cheaper, faster work, worse work.

[04:54:03](https://www.youtube.com/watch?v=jiUC6IUMdvc&t=17643s) If AI makes PRs faster and cheaper, are we just shipping more mistakes?

[04:54:08](https://www.youtube.com/watch?v=jiUC6IUMdvc&t=17648s) One proxy for this is reverts. Merge PRs that undo an earlier PR.

[04:54:15](https://www.youtube.com/watch?v=jiUC6IUMdvc&t=17655s) The heaviest fifth started lowest before most of the token growth.

[04:54:20](https://www.youtube.com/watch?v=jiUC6IUMdvc&t=17660s) And revert rates dropped slightly for every group.

[04:54:24](https://www.youtube.com/watch?v=jiUC6IUMdvc&t=17664s) though these deltas are just slightly below the noise in the data.

[04:54:28](https://www.youtube.com/watch?v=jiUC6IUMdvc&t=17668s) It's not evidence of improvement here.

[04:54:32](https://www.youtube.com/watch?v=jiUC6IUMdvc&t=17672s) However, more importantly, we're not seeing any worrying signs of rampant or severe quality issues

[04:54:40](https://www.youtube.com/watch?v=jiUC6IUMdvc&t=17680s) with higher token spends and faster PR throughput.

[04:54:46](https://www.youtube.com/watch?v=jiUC6IUMdvc&t=17686s) it. The second curiosity, is throughput driven by smaller PRs? And this is demonstrably no,

[04:54:58](https://www.youtube.com/watch?v=jiUC6IUMdvc&t=17698s) not smaller PRs. In fact, AI is driving an increase in lines of code generated per PR.

[04:55:07](https://www.youtube.com/watch?v=jiUC6IUMdvc&t=17707s) This is no surprise to the devs in the room. Raise your hand if you've rolled your eyes at

[04:55:13](https://www.youtube.com/watch?v=jiUC6IUMdvc&t=17713s) for both unnecessary inline AI comments,

[04:55:17](https://www.youtube.com/watch?v=jiUC6IUMdvc&t=17717s) or a file full of mostly irrelevant tests.

[04:55:20](https://www.youtube.com/watch?v=jiUC6IUMdvc&t=17720s) Gen AI likes to generate code.

[04:55:25](https://www.youtube.com/watch?v=jiUC6IUMdvc&t=17725s) Also, the bar to create is low for devs,

[04:55:27](https://www.youtube.com/watch?v=jiUC6IUMdvc&t=17727s) and there's a pool to do more per PR

[04:55:30](https://www.youtube.com/watch?v=jiUC6IUMdvc&t=17730s) instead of spending up another ticket.

[04:55:32](https://www.youtube.com/watch?v=jiUC6IUMdvc&t=17732s) It's just one, maybe two prompts,

[04:55:34](https://www.youtube.com/watch?v=jiUC6IUMdvc&t=17734s) and I'll handle this edge case,

[04:55:36](https://www.youtube.com/watch?v=jiUC6IUMdvc&t=17736s) or I can combine this and that element.

[04:55:38](https://www.youtube.com/watch?v=jiUC6IUMdvc&t=17738s) Fewer PRs to wrangle.

[04:55:40](https://www.youtube.com/watch?v=jiUC6IUMdvc&t=17740s) This doesn't mesh well with the PR review stress

[04:55:43](https://www.youtube.com/watch?v=jiUC6IUMdvc&t=17743s) folks are feeling these days though.

[04:55:48](https://www.youtube.com/watch?v=jiUC6IUMdvc&t=17748s) Also, AI is just becoming more capable,

[04:55:52](https://www.youtube.com/watch?v=jiUC6IUMdvc&t=17752s) so we're getting more ambitious.

[04:55:54](https://www.youtube.com/watch?v=jiUC6IUMdvc&t=17754s) It is beyond this talk to dig into what's actually driving

[04:55:58](https://www.youtube.com/watch?v=jiUC6IUMdvc&t=17758s) those level of lines of code changes.

[04:56:02](https://www.youtube.com/watch?v=jiUC6IUMdvc&t=17762s) However, the increase in PR throughput

[04:56:05](https://www.youtube.com/watch?v=jiUC6IUMdvc&t=17765s) is not coming from smaller PRs.

[04:56:08](https://www.youtube.com/watch?v=jiUC6IUMdvc&t=17768s) This is intuitive for most, but it's still good to verify.

[04:56:13](https://www.youtube.com/watch?v=jiUC6IUMdvc&t=17773s) Back to the opening question.

[04:56:16](https://www.youtube.com/watch?v=jiUC6IUMdvc&t=17776s) Is 60x more spend worth 43% more output?

[04:56:21](https://www.youtube.com/watch?v=jiUC6IUMdvc&t=17781s) You can't tell in isolation.

[04:56:23](https://www.youtube.com/watch?v=jiUC6IUMdvc&t=17783s) You should track both.

[04:56:25](https://www.youtube.com/watch?v=jiUC6IUMdvc&t=17785s) They are important numbers, but they're supporting metrics.

[04:56:29](https://www.youtube.com/watch?v=jiUC6IUMdvc&t=17789s) AI spend per PR is the intelligence that you're buying as a resource to help make your devs go faster.

[04:56:38](https://www.youtube.com/watch?v=jiUC6IUMdvc&t=17798s) PRs per engineer is what you're getting with that spend.

[04:56:43](https://www.youtube.com/watch?v=jiUC6IUMdvc&t=17803s) However, looking at those hides the trade-off.

[04:56:51](https://www.youtube.com/watch?v=jiUC6IUMdvc&t=17811s) But put salary into the equation and take the ratio of cost to throughput.

[04:56:56](https://www.youtube.com/watch?v=jiUC6IUMdvc&t=17816s) And there you get your answer.

[04:56:58](https://www.youtube.com/watch?v=jiUC6IUMdvc&t=17818s) Cost per PR is down 26% over that same time period.

[04:57:04](https://www.youtube.com/watch?v=jiUC6IUMdvc&t=17824s) It's clear AI is driving cost efficiencies despite the increasing bills.

[04:57:13](https://www.youtube.com/watch?v=jiUC6IUMdvc&t=17833s) and the AI spend needed to get those efficiencies is small, relatively speaking.

[04:57:21](https://www.youtube.com/watch?v=jiUC6IUMdvc&t=17841s) AI is still a sliver of what a PR costs, 2.5% on average,

[04:57:28](https://www.youtube.com/watch?v=jiUC6IUMdvc&t=17848s) under 10% for the heavy spender group.

[04:57:31](https://www.youtube.com/watch?v=jiUC6IUMdvc&t=17851s) Tokens are the cheap part.

[04:57:33](https://www.youtube.com/watch?v=jiUC6IUMdvc&t=17853s) In the range that we see, more spend kept paying.

[04:57:38](https://www.youtube.com/watch?v=jiUC6IUMdvc&t=17858s) heavier token users saw 67% PR throughput gain compared to 40% for their lower spend peers.

[04:57:47](https://www.youtube.com/watch?v=jiUC6IUMdvc&t=17867s) This drove cost per PR down 32% compared to about 25% for the lower spend peers.

[04:57:55](https://www.youtube.com/watch?v=jiUC6IUMdvc&t=17875s) Looking at the PR cost to token ratio curve,

[04:57:59](https://www.youtube.com/watch?v=jiUC6IUMdvc&t=17879s) The top 10% of company months by token usage resulted in 8x the tokens, but 25% lower cost per PR compared to the median.

[04:58:14](https://www.youtube.com/watch?v=jiUC6IUMdvc&t=17894s) AI is making developers more productive on a scale we haven't seen before.

[04:58:20](https://www.youtube.com/watch?v=jiUC6IUMdvc&t=17900s) It isn't just another tool, it is intelligence, which changes how much each dev can produce.

[04:58:28](https://www.youtube.com/watch?v=jiUC6IUMdvc&t=17908s) That's why we need to think about its costs differently.

[04:58:30](https://www.youtube.com/watch?v=jiUC6IUMdvc&t=17910s) Not off to the side as overhead, but as part of what it takes to produce each PR.

[04:58:36](https://www.youtube.com/watch?v=jiUC6IUMdvc&t=17916s) You need to add the AI spend and the human spend, the human salary spend, together to calculate your real costs per PR.

[04:58:48](https://www.youtube.com/watch?v=jiUC6IUMdvc&t=17928s) This isn't to say spend frivolously.

[04:58:52](https://www.youtube.com/watch?v=jiUC6IUMdvc&t=17932s) Sure, you should watch for runaway jobs.

[04:58:56](https://www.youtube.com/watch?v=jiUC6IUMdvc&t=17936s) You should track your AI spend.

[04:58:58](https://www.youtube.com/watch?v=jiUC6IUMdvc&t=17938s) I'm not saying don't look at how much AI spend is going in each PR.

[04:59:05](https://www.youtube.com/watch?v=jiUC6IUMdvc&t=17945s) I'm saying don't let that be your driving metric.

[04:59:08](https://www.youtube.com/watch?v=jiUC6IUMdvc&t=17948s) A new framing might steer how and when you think about other investments,

[04:59:14](https://www.youtube.com/watch?v=jiUC6IUMdvc&t=17954s) like integrating model routers, post-training your own model,

[04:59:18](https://www.youtube.com/watch?v=jiUC6IUMdvc&t=17958s) implementing token limits, et cetera.

[04:59:22](https://www.youtube.com/watch?v=jiUC6IUMdvc&t=17962s) Will you get more efficiencies out of trying to control AI costs or by continuing to invest in things that increase throughput, like more AI enablement?

[04:59:33](https://www.youtube.com/watch?v=jiUC6IUMdvc&t=17973s) Either way, track the total intelligence costs, human plus artificial, per PR every month next to that AI bill and next to the PR throughput.

[04:59:45](https://www.youtube.com/watch?v=jiUC6IUMdvc&t=17985s) How are you trending? Is throughput continuing to improve?

[04:59:49](https://www.youtube.com/watch?v=jiUC6IUMdvc&t=17989s) What about those cost efficiencies, the cost to deliver?

[04:59:53](https://www.youtube.com/watch?v=jiUC6IUMdvc&t=17993s) Watch reverts and quality metrics beside all of this as a counterweight.

[04:59:59](https://www.youtube.com/watch?v=jiUC6IUMdvc&t=17999s) We don't

[05:00:00](https://www.youtube.com/watch?v=jiUC6IUMdvc&t=18000s) AI slot mucking up the gears.

[05:00:02](https://www.youtube.com/watch?v=jiUC6IUMdvc&t=18002s) So heading into budget season, bring cost per PR to the table. Not just your token bill, not just the throughput bump.

[05:00:12](https://www.youtube.com/watch?v=jiUC6IUMdvc&t=18012s) The bill tells you what you spent. Cost per PR tells you whether it worked.

[05:00:18](https://www.youtube.com/watch?v=jiUC6IUMdvc&t=18018s) And right now, for most companies, it says keep investing.

[05:00:23](https://www.youtube.com/watch?v=jiUC6IUMdvc&t=18023s) Tokens are the cheap part to get efficiency gains.

[05:00:30](https://www.youtube.com/watch?v=jiUC6IUMdvc&t=18030s) Well done. You made it to the end.

[05:00:33](https://www.youtube.com/watch?v=jiUC6IUMdvc&t=18033s) Here's how you can stay up to date with our latest research, which again draws from the largest engineering workflow and AI telemetry data set in the industry.

[05:00:42](https://www.youtube.com/watch?v=jiUC6IUMdvc&t=18042s) You can follow me on LinkedIn, my colleagues, or Jellyfish.

[05:00:48](https://www.youtube.com/watch?v=jiUC6IUMdvc&t=18048s) subscribe to our sub stack,

[05:00:50](https://www.youtube.com/watch?v=jiUC6IUMdvc&t=18050s) or you can check out our Jellyfish blog on the website.

[05:00:54](https://www.youtube.com/watch?v=jiUC6IUMdvc&t=18054s) If you want help tracking your org's engineering workflow

[05:00:57](https://www.youtube.com/watch?v=jiUC6IUMdvc&t=18057s) and AI effectiveness, including cost per merged PR,

[05:01:02](https://www.youtube.com/watch?v=jiUC6IUMdvc&t=18062s) go to our website and book a demo or reach out to me

[05:01:05](https://www.youtube.com/watch?v=jiUC6IUMdvc&t=18065s) and I'll get you connected to the right folks.

[05:01:08](https://www.youtube.com/watch?v=jiUC6IUMdvc&t=18068s) Again, I'm Steve Jamison,

[05:01:10](https://www.youtube.com/watch?v=jiUC6IUMdvc&t=18070s) AI research and engineering leader at Jellyfish.

[05:01:12](https://www.youtube.com/watch?v=jiUC6IUMdvc&t=18072s) Thanks for watching.

[05:01:14](https://www.youtube.com/watch?v=jiUC6IUMdvc&t=18074s) Cheers.