Growth Unhinged is proudly supported by beehiiiv
On July 16th at 1pm ET, beehiiv is unveiling the next chapter for audience-led businesses.
For years, creators and brands have been forced to stitch together bloated stacks of tools just to publish content, grow an audience, and make money online. I’ve lived this firsthand: newsletters in one platform, websites in another, podcasts somewhere else, analytics scattered everywhere.
beehiiv thinks there’s a better way. Now, they’re ready to show it off at their Summer Release Event. See you there?
I recently talked to a Series B founder who told me their Anthropic bill would soon jump from $400,000 to $1.4 million per year.
This is at a software company with 150 employees and only covers internal use of AI.
The company is Pylon, a support platform built for B2B. Co-founder and CEO Marty Kausas went viral after posting about how they face a 3.5x overnight cost jump from Anthropic.
This cost jump didn’t have to do with Pylon’s AI usage exploding. It was because they were about to pass 150 seats, which is when Anthropic forces customers into their Enterprise tier according to Marty. Enterprise seats no longer include subsidized usage meaning that every token is billed at Anthropic’s standard API rates. Given Pylon’s AI consumption, this was a wakeup call.
The Information reported on Anthropic’s new enterprise pricing in April 2026
Pylon is far from the only company seeing truly eye-watering AI bills and wondering what to do about it. As much as there’s talk about AI costs coming down – OpenAI is reportedly considering “drastic price cuts” – frontier models continue to get more expensive.
Anthropic’s splashy new model, Fable 5, is finally available (again) to the public. Fable 5 will soon move off subscription plans; however, and require usage credits. The sticker price of this usage – $10 per 1M input tokens, $50 per 1M output tokens – is about twice as expensive as Opus 4.8. Fable 5 also produces longer responses on the same prompts. On a cost per task basis, Fable 5 is estimated to cost about 3-5x Opus 4.8 and 10x Sonnet 4.6. Ouch.
The TL;DR: AI bills are likely to keep going up, not down. Today’s post unpacks the looming AI cost crisis and what GTM teams can do about it — without killing AI adoption.
A huge thank you to Marty Kausas (co-founder and CEO at Pylon), John McCauley (CFO at Vanta), Dan Zhang (CFO at ClickUp), and Sonalee Parekh (CFO at SentinelOne) for sharing their insights.
We panicked our way into this crisis
This is a pretty strange piece to write because the AI cost debacle is a completely invented problem.
We as an industry created this problem.
We panicked and pushed AI adoption over all other things. And “we” means panicked investors, which trickled down to founders, then management teams, and finally scared employees who didn’t want to get fired because they didn’t use AI.
If you’re anything like me, you spent nights and weekends learning Claude Code or Cowork possibly because everyone else was doing it. You set up all the context that Claude needs to deliver reliable outputs. You turned your GTM expertise into custom Claude skills.
Now that AI finally (for the first time!) feels like it’s working for GTM, there’s a new crisis. It’s gotten insanely expensive to keep running the AI workflows you worked so hard to build.
The whiplash is exhausting. But the rent (LLM tokens) is still due.
Start here to manage your AI costs
There are some lighter lift things that folks can do to manage runaway AI costs. Start here if you haven’t already and then keep reading for longer-term fixes:
Set smarter defaults for your org like Sonnet instead of Opus. (Coinbase’s CEO, for example, says they’re now defaulting to open weight models.)
Consider a model router or harness to route requests to cheaper models. Examples include OpenRouter and LiteLLM.
Split different functions to use different providers to avoid being pushed to an enterprise plan without token subsidies (as Pylon experienced).
Pair AI reasoning with automation wherever possible as it’s more reliable and is a fraction of the cost. (monday.com is a good case study.)
Test on a small scale or cap output tokens to avoid unknowingly burning through credits on big tasks.
Compress context files and connections so they don’t waste tokens.
Batch usage for low urgency, high-intensity tasks (Anthropic’s Batch API is 50% off).
Keep chat sessions short and regularly start fresh sessions.
Simply adjusting the default model within Claude can make a big cost difference.
Vanta CFO John McCauley told me that they found Opus 4.7 was similarly priced per 1M tokens compared to Opus 4.6, but consumed far more tokens per task. They decided to set the default to Sonnet 4.6 for everyday tasks and have blocked usage of Fable. (I do this myself and haven’t seen any difference in quality.)
Step 1: Show people what they spend on AI
Like many startups, Pylon pushed employees to embrace AI. Then the token bill arrived.
Marty’s first step was to track AI spend at the individual and team-level. The results were pretty shocking.
Pylon’s monthly Anthropic costs by team (this doesn’t include OpenAI or other AI vendors)
Marty himself had unknowingly spent $4,000 in just a few days after having Claude Code run a deep-dive analysis of all closed-lost deals.
Engineering spent $3,107 per person per month on Claude alone. After accounting for Codex, Cursor, and Devin usage, Marty believes the true figure is $5,000 (!) per month. This still felt worth it, he told me, because AI coding is non-negotiable these days.
Pylon’s support team was the second biggest spender at more than $11,000 per month across a 10 person team. Every ticket was triggering a Claude skill to run an investigation pulling together logs, databases, and account context. The base inference cost was $2 per ticket just to pull this data together. There were savings to be found.
There’s an emerging tech stack around AI spend visibility. Ramp, for instance, introduced AI spend intelligence in April. It’s hard to imagine these types of tools not becoming widely available.
Step 2: Figure out whether AI spend is a real problem
“Spend is climbing so fast we have to start asking whether it’s justified,” Marty told me. “A lot of it is. A lot of it isn’t. And most people (including me) aren’t conscious of what they’re spending. That has to change.”
The marketing team was spending $739 per person yet Marty hasn’t seen the results to justify that level of spend.
“AI written copy has been terrible,” Marty admitted. “It’s gotten better, but is very far from perfect and keeps repeating the same mistakes.”
The best marketing use case has been creative brainstorming (e.g. billboard headlines, positioning wording), although this is relatively low volume and low cost.
Sales hasn’t been a game changer, either, despite the team spending nearly $10,000 per month in AI tokens.
Pylon built an AI coach for SDRs via Claude, although it wasn’t built-into existing SDR tools and adoption has dropped off. Claude helps with list building (although not super effectively), drafting emails and sequence copy, and crawling closed-won deals to check attribution. AE’s use Claude mostly for meeting prep, competitor comparisons, and drafting follow-ups.
John McCauley from Vanta, who I mentioned earlier, recommends showing AI spend to the executive team and aligning on where you sit on the spectrum of ‘irrational exuberance’ (i.e. AI spend is purposely dislocated from value) and ‘trough of disillusionment’ (i.e. AI spend is viewed as mostly wasteful and needs to stop).
Where you sit will dictate the cost control strategy.
Step 3: Set token budgets per team
From there, it’s time to set token budgets at the team and even individual level.
These token budgets seem to range pretty widely across companies. Tesla caps AI spend at $200 per week. Uber said it caps AI spend at $1,500 per employee and per tool on a monthly basis.
Pylon is setting different token budgets for each team based on a combination of current spend and reflecting on what the upper limit should be.
Token budgets could backfire and hurt AI adoption. Most caps come with a workaround: employees can get more token access with manager approval.
“The goal is not to remove AI access, but to add a small amount of friction,” Marty said.
For teams that are prepared to ratchet up AI cost discipline, John McCauley shared five action items:
Tighter procurement rules with teeth on adopting new tools, including value rationalization.
Dynamic leaderboards of individual user consumption, available to all leaders.
Per-user token caps for consumption-based tools with exceptions for proven use cases.
Avoid the Ferrari for the daily commute when the Honda Civic will do the job. Make sure model economics and capabilities are well understood by users and set defaults for proper model selection.
You need an internal AI czar with authority to create an AI tooling vision in order to mitigate wasteful tool sprawl. Lock down the tools you plan to standardize on and block all the others.
Step 4: Once budgets exist, products can compete for that spend
ClickUp CFO Dan Zhang told me that companies need to do real zero-based budgeting that accounts for all spending, not just AI costs.
“Question all your existing spend, set it back to zero, and put AI in as an equal budget competitor, not a line add on top of headcount and vendors. What would you choose?”
Marty from Pylon had a similar perspective. Now that there’s an established budget that’s been created by Anthropic and OpenAI, teams should be able to swap in purpose-built tools that can compete on cost and quality.
When Maja Voje and I surveyed 200 GTM operators who use Claude Code and Cowork, we found that credit, cost, and usage limits were a near-constant source of frustration. Other limitations included losing context, no long-term memory, weak integrations with certain tools, and orchestration issues.
Purpose-built tools can optimize models under the hood and can more efficiently gather the right context for a given workflow. Aside from his own product, Marty mentioned Unify as an example (they’re starting to trial Unify’s new AI product for the SDR team). Perhaps the pendulum swings back to the application layer?
Thanks for reading Growth Unhinged! To support my work, please subscribe, share, or upgrade to paid. Paid subscribers unlock quarterly analyst reports, the full archive of 200+ editions, and savings on top GTM tools.
Step 5: Treat AI spend like any other strategic investment
FOMO – or, perhaps, hysteria – has given AI special budget treatment. Soon it’ll be treated like every other investment.
SentinelOne CFO Sonalee Parekh said she thinks the largest risk is uncontrolled AI consumption without a clear business outcome.
Her advice: “Assign budgets, measure value delivered by use case, and maintain visibility into consumption across teams and applications.”
FOMO got us into the AI cost crisis. What gets us out of it: building a system for deciding when AI is worth the cost.
What else is on my mind:
Implications for AI pricing: As enterprises scrutinize AI costs, where does AI pricing go? My prediction is one of two directions: dual-track monetization with platform + tokens (see: Clay) or outcome-based pricing (see: SFDC’s new pay-per-resolution pricing model).
LLM switching costs: Companies that want to optimize LLM costs will naturally arbitrage AI use across multiple vendors including open-source and local models. How AI companies will (probably) fight back: introducing products that create lock-in. An interesting recent example is Claude Tag — Anthropic’s new multi-player mode. Will proprietary vendors move fast enough to lock us in?