XMarts room

Winter 2026

Carrot Labs

AI Cost Management: Track and attribute AI spend across every provider

Visit website

What they do

tracks and attributes AI spend across providers, customers and features

How close to us

5, very close

covers the AI cost part of W2 for small teams; checks only, no stuck or failed work view.

Wedges

  • W2 Seeing whether your AI agents actually produce direct
  • W1 Trusted 24/7 coding for small startup teams adjacent
  • W7 Find and fix margin leaks adjacent
L1 Company brain (knows the organisation) 1, light
L2 Agent team (runs agents, tools, approvals) 0, none
L3 Does and learns (finishes, checks, measures, updates the brain) 1, light

Who buys

Firm size
sme
Buyer type
developer
Role
engineering leads and finance leads at AI-heavy startups
UK
not known
Their own description (from YC)

The problem:
AI has quietly become one of the largest and fastest-growing line items for companies. Spend is scattered across a dozen provider consoles that each show one number at the end of the month and nothing about why it moved. There's no unified view, and no way to trace a dollar back to the customer, feature, team, or even the pull request that caused it.

Which customers are unprofitable once you subtract their AI costs? Which feature is quietly burning your Anthropic budget? Is this week's spike a runaway loop or real growth? How much did that AI-written PR actually cost to ship?
Today teams reverse-engineer answers by exporting CSVs from each provider and stitching them together in a spreadsheet, and by the time they do, the money is already spent.

What SuperPenguin does: SuperPenguin tracks AI spend across 14 providers (OpenAI, Anthropic, Google Gemini, Deepgram, ElevenLabs, AWS Bedrock, Azure, Modal, Cursor, OpenRouter, and more).

* Zero-code setup: connect an API key and costs sync automatically with model-level breakdowns, trends, and forecasts
* Per-request attribution: add two lines with our Python or TypeScript SDK to tag every AI call by customer, feature, team, or environment (or any other metadata).
* AI coding cost per PR: connect Cursor to see engineering spend attributed to each pull request, repo, and developer, so you know what it actually costs to ship.
* Alerts on budget thresholds and spend anomalies, delivered to Slack, email, or Discord.

Most teams are set up in under five minutes. We help companies see where their AI money goes and whether it's worth it.

Full teardown

Carrot Labs (YC Winter 2026): teardown vs XMarts

In one line: Carrot Labs sells SuperPenguin, a tool that shows what a company spends on AI, split by customer, feature, team and pull request [site] [yc]. Threat: 2 because it covers the "what it costs" part of W2 for small teams, but it is two people, early, and it checks only: it does not show stuck or failed work [site] [yc].

1. Fact file

Item Finding Source
What they say "Know where your AI Spend goes." Sub-line: "Your providers tell you the total. We tell you what it was for." [site]
Founded, team, funding YC Winter 2026, San Francisco, 2 people, status Active, no open jobs [yc]. Funding: one directory lists $500K total, one lists a $125k seed round. Neither is confirmed by the company (UNVERIFIED) [press]. [yc] [press]
Founders Christopher Acker (founder, ex-lead of AI at Skylo, a satellite network) and Yuta Baba (ex-senior data scientist at Snowflake, built bookings forecasts). What YC likely saw: two people from data and ML who know cost and forecasting (inference). [yc]
Customers One named case study: Channel3, "billions of tokens a day". Quote from its CTO, George Lawrence: "the central place for managing our AI spend". No savings number, no date. No other logos. [site]
Price Public, usage based. Free up to $2K of managed AI spend. Growth $30 a month (up to $5K). Pro $200 a month (up to $20K). Enterprise custom from $20K+. No sales model found in job ads (no jobs listed). [site] [yc]
Path Listed in the YC directory as "Anyware" before "Carrot Labs" (no date given) [yc]. The YC launch post pitches something else: custom models tuned to your tasks that keep improving [yc]. The company site now sells only AI spend tracking under the SuperPenguin name [site]. The YC page does not explain how the two fit (UNVERIFIED which came first). Blog posts run from 26 Jun 2026 to 18 Aug 2026 [blog]. [yc] [site] [blog]

2. Their proposition, reverse engineered

Part What they do
Who buys Engineering and finance leads at companies that use many AI providers [yc]. Named customer is a startup (Channel3) [site]. Size: sme (inference from the free tier and the $30 plan) [site].
The pain AI spend sits in a dozen provider consoles, each showing one number. Teams "reverse-engineer answers" from CSV files in a spreadsheet [yc]. Questions they pose: which customers lose money after AI cost, which feature burns the Anthropic budget, is a spike a runaway loop or real growth [yc].
The promise Know what each dollar of AI spend was for, set up in under five minutes [yc].
How it works Connect a provider API key and costs sync with model-level breakdowns and forecasts [yc] [site]. Optional SDK (Python or TypeScript) tags each call by customer, feature, team or environment. It is not a proxy and adds under 10ms, they claim [site]. A free Mac app plus GitHub connect gives cost per merged pull request for Cursor, Claude Code and Codex [site]. Alerts go to Slack, Discord or email [site].
The output A dashboard: spend by provider, model, customer, feature, team, repo and PR; forecasts; budget and spike alerts [site]. A page example shows $18.40 for one PR and 85% of agent spend tied to PRs (an example, not a customer result) [site].
Proof style Scale and coverage, not results. They claim 14 providers; a third party repeats a launch claim of one trillion tokens a month, and says that is a scale claim, not proof of savings [press]. The one case study has no savings figure [site].
Who sells it Self-serve. Free plan needs no credit card [site]. Enterprise by contact (inference: no sales team is visible) [site] [yc].
Price model Tiered by "managed spend" seen each month, not by seats [site]. Cheap monthly fee next to the spend it tracks.
The story Their enemy is the provider bill that hides the "why". Their belief: AI FinOps is like cloud FinOps in 2012, and cost per outcome is the measure that matters [blog].

3. Positioning

  • Category they claim: "AI spend intelligence" and AI cost management [site] [yc].
  • Category a YC partner would put them in: AI FinOps for startups, a young cousin of cloud cost tools (inference). The YC site lists them under FinOps [yc].
  • What they want to be known for: the central place that explains AI spend, down to the pull request.
  • Where they sit on the map: layers L1/L2/L3 = 1/0/1, buyer = sme, developer (engineering lead adopts, finance reads).

4. Three layers vs ours

Layer Them (score + evidence) Us (Marton's logs show)
L1 Company brain 1. Keeps spend data and the tags you add (customer, feature, team). Optional prompt and outcome capture on Pro, off by default [site]. No lasting memory of how the company works that other work reads. A vault of about 1,400 pages, 112 call transcripts and about 8,700 customer quotes, searchable by meaning.
L2 Agent orchestration 0. Nothing runs or routes agents. It reads bills and sessions [site]. About 5,600 jobs run by agents from three makers, each checked by a different AI, then automatic tests.
L3 Execution and learning 1. Checks only. It alerts, forecasts and links cost to merged PRs, but does not do the work or change how the next run goes [site]. Their "optimization reports" are advice (UNVERIFIED depth) [press]. A heartbeat reports what is stuck, what silently stopped and what it costs. About 160 failure records, each with an automatic check. Only about 15% of jobs reached the main code.

5. Same and different

Topic Same or different Them Us
Core belief Same Token volume is not the goal. Ask what you got for it [blog]. Same (brief, W2). Marton's logs show a cost meter and an incident about runaway spend.
Cost per merged PR Same Cost per merged PR across Cursor, Claude Code, Codex [site]. Marton's logs show jobs tracked as merged or not (864 merged of 5,641), tied to a cost ledger.
What is measured Different Money spent, by dimension [site]. Marton's logs show named failure classes: stuck, silently stopped, wont-merge, plus cost.
Data source Different Provider billing, optional SDK, Mac session collector. Metadata only, not in the request path [site]. Run records from our own queue, gates and heartbeat. Mac and launchd bound.
Buyer Similar Small AI-heavy teams, self-serve [site]. W2 targets teams running agents in loops (brief). No customer yet.
Owner of the result Different The buyer reads a dashboard. Nothing is fixed for them [site]. The target is a view plus action. Not shipped to anyone outside Marton.
Proof Different One case study, no number [site]. No customer number. Own-business logs only.
Delivery Different Software, 5-minute setup [yc]. System plus two founders. Not packaged.
Price Different Public, $0 to $200 a month, then custom [site]. Not set yet.

6. Wedge overlap

Wedge Fit Why
W2 Seeing whether your AI agents actually produce direct They cover the cost half and a first output measure (cost per merged PR, spend that never reached a PR) for small teams [site]. They do not show what is stuck or what silently failed (UNVERIFIED that none exists).
W1 Trusted 24/7 coding for small startup teams adjacent Same buyer, other job: they measure what coding agents cost; W1 would run them [site]. Their buyers are likely W1's buyers (inference).
W7 Find and fix margin leaks adjacent Their "which customers are unprofitable after AI cost" question is a margin question for AI software firms [yc]. Narrow slice, no fix step.

7. Whitespace (where they stop)

  • Layer gap: they read bills and sessions; they do not see run health (stuck, failed, silent) or do any work [site]. To own it we need the failure view working on a second team's agents, not only ours.
  • Number gap: their one case study has no savings or output number [site]. Cost per merged PR is the closest they get. To own it we need a before/after number such as unmerged spend down, or rework down, on a real team.
  • Buyer gap: not open. They sell to the same small AI-heavy teams. Do not claim it.
  • Price gap: not open. Free to $200 a month is cheap [site]. Do not claim it.
  • Speed gap: not open. They claim setup in under five minutes [yc]. Do not claim it.

8. For the YC application

  • Proven today: Marton's logs show a heartbeat that reports what is stuck, what silently stopped and what it costs. Marton's logs show about 160 written failure records, each with an automatic check, and one named cost incident (a single launch authorised 156,000 agent turns and used 39% of a weekly budget in 14 hours). SuperPenguin's public pages show spend and cost per merged PR, not failed or stuck runs [site].
  • Draft answer to "how are you different from SuperPenguin?": DRAFT. "SuperPenguin tells you what your agents cost. We tell you whether they produced, what got stuck and what failed quietly, because we have run that fleet every day for five months." Founders say it their own way.
  • Future claim (needs proof): "We cut wasted agent work on a real team." Needs one outside team's agents watched for 4+ weeks, with a before/after number such as spend that never reached a merged change.
  • What we understand that they don't: cost alone does not tell you if agents are producing; the useful signal is named failure classes (needs proof outside our own system; Marton's logs show the classes on our own fleet).
  • Copyability (M3): medium. They already show spend that never reached a merged PR and have session data from the Mac app [site], so a "wasted work" view is close for them. A run-health view with stuck and silent-failure classes needs run-level data they avoid by design (cost metadata only, not in the request path) [site] (inference). They are two people with a fast blog and a self-serve motion, so they could ship it (inference, no jobs listed [yc]). Not hard, so M3 can not score above 3 against them alone.
  • Why now (M4): None found in this teardown. Their own posts cite price changes by model makers [blog], which is general AI progress, not a dated change that helps us more than them.
  • Do not claim: "nobody tracks agent cost" (they do, across 14 providers). "Cheaper or faster to set up than SuperPenguin". "Our system fixes stuck work for other teams" (nothing outside Marton's own business yet). "Cost per PR" as our edge. Marton's logs also show only about 15% of jobs reached main, so our own merged-PR cost would look poor next to their metric.

M2 (what YC funds now) is NOT judged here.

9. Ideas worth taking

  1. One plain number per PR, with a tag showing how sure the match is (Verified, Exact, Fuzzy) [site]. A buyer understands it at once.
  2. A free tier with no card and a 5-minute start. It gets the first user and the first data fast [site].
  3. Write as the category maker: "AI FinOps is having its 2012 moment" frames a new job with a known history [blog]. Use the shape for agent run health.

10. Unknowns

  • How many paying customers they have, and the real funding. Look for a Product Hunt or YC launch thread, or ask a user.
  • Whether the "custom models" launch post is the old direction or a second product. Look at the YC launch page comments and the X account.
  • Whether they sell in the UK or Europe. Look for UK customers or data location terms.
  • Whether they plan run-health features (stuck, failed work). Look at their changelog and docs.
  • Founding dates and the Anyware to Carrot Labs rename date.

Sources

Technical detail
Slug
carrot-labs
Lens hash
179096c479
Pass
deep
Stale
no