IWhat $1,462.37 Bought Black Matter VC in June 2026
In June 2026, running a small swarm of AI agents at Black Matter VC cost $1,462.37 total — $1,022.82 spent directly with model providers (OpenAI, Anthropic, Perplexity), and $439.55 on the tooling that keeps them fed (Replit, Firecrawl). For that money, the swarm did the work of an analyst and a research assistant combined: it watched every feed and account on my research watchlist and filtered it down to what mattered, every single day, with no human touching the routine passes.
Here's what that actually means, and why it exists: it's watching social posts, watching my own work, watching what's going on inside the studio, keeping up with the latest in tech news, and scouring the web — all to keep a live, always-current internal feed of what's happening in tech and at Black Matter VC. That feed is also the point of the whole exercise: it's collecting real, recent, actual data so that when I sit down to publish an article, it's grounded in what's actually happening right now, not a guess. That's the objective, full stop — a swarm that makes what I write genuinely useful.
Nobody publishing about "AI token costs" right now is showing you that trade with real receipts attached. So here's the full breakdown below: what each agent costs, which model does which job, and exactly where the money goes.
IINobody publishing about "AI token costs" is actually running the meter
Search "AI token costs" right now and look at what comes back. OpenAI's own pricing page. A glossary explainer. A token calculator. And a Reddit thread quoting Sam Altman, who told an enterprise crowd at OpenAI's "Intelligence at Work" event that costs went from a non-issue to a "huge issue" in about five months — one customer telling him, "My company spent my entire 2026 budget in Q1. Can you make this more efficient?"
Every result is a rate card or a doom-thread. Not one is written by someone actually running the meter, day in and day out.
I do. I run a small swarm of AI agents in production for Black Matter VC — a dozen or so worker roles, each on its own schedule, every run logged to an audit trail with a timestamp, a status, and a short note on what it did. So instead of reading you a pricing page, I pulled June 2026's actual bill — the OpenAI usage export, the Claude API token log, and the full spend breakdown across every AI tool the studio pays for — and matched it against what each agent actually does. Here's what a live multi-agent operation really costs, tool by tool and model by model — and where the money actually goes.
If you want the companion piece on the tools themselves, I already spent $1,300 testing every AI agent I could get my hands on. This one is about the bill — my actual June 2026 bill, not a hypothetical one.
IIIWhat the swarm actually is, in plain terms
Strip the branding and it's a handful of worker types, each doing one job well and each tied to a specific tool in the June bill. A scout watches for signal — pulling posts and feeds from a watchlist using Firecrawl's scraping API and Perplexity's search, then filtering the noise down to a few things worth keeping. A curator organizes what survives into clusters and drafts, leaning on OpenAI's text-embedding-3-large to group related material before anything gets written. An observer audits the runs — a quality gate that pulls in outside reading through Perplexity and flags anything weak or stale. A listener sweeps social accounts for what's resonating, and an engineer probes the code side and ships small fixes, working inside Replit's AI coding environment and, for the harder reasoning, Claude directly through Anthropic's Max plan.
Sitting above those is a weekly synthesizer that reads a stretch of collected posts, plus the house-style rules re-sent as cached context on almost every call, and turns the whole thing into a long essay. It's the one heavy job in the swarm, and it runs on Claude Opus — the frontier model, reserved for the one task that's actually worth the price. A few other agents only fire occasionally: a fact check here, a format pass for a specific channel there. Everything runs on a schedule, logs a row when it finishes, and pings me on Slack if it breaks. No humans in the loop for the routine passes. If you want the full architecture, I wrote up how I've structured this whole operation for AI in 2026 separately — here I only care about what it costs to keep the lights on.
IVThe real June numbers: fixed subscriptions vs. metered API usage
Here's the part that used to be a limitation and now isn't: I pulled the actual statement. Start with what's fixed. Perplexity's subscription runs $150/month, and it's what the scout and observer lean on for outside search whenever a topic needs checking. Anthropic's Max plan is $200/month flat, and it's what the engineer runs on day to day inside Claude. I also run OpenAI's Codex plan at $100/month, which powers my OpenClaw setup — not a swarm worker at all, but a personal assistant that handles the general upkeep and admin work around the company I'd rather not do myself. That's $450 a month in pure flat-fee subscriptions, paid whether the swarm has a light week or a heavy one.
Everything else moves with usage. OpenAI billed $522.79 across 15 charges in June 2026 for the models actually doing the work — GPT-5.5 and GPT-5.1 for synthesis and curation, GPT-4o and GPT-4o mini for routine filtering, GPT-5.4 mini for light passes, text-embedding-3-large for clustering, and the GPT-image family for the handful of images the swarm generates. Perplexity's API credits added $130 on top of its subscription — real-time search calls the scout and observer make when a claim needs outside verification. And there's a separate $20.03 Anthropic console charge, distinct from the $200 Max subscription, that turns out to be a near-perfect match for the actual Claude Opus token log for June 2026: about 3.9 million input and output tokens for the month on that one key, which prices out to right around $20 at Anthropic's published Opus rates once you account for its cached-token discounts. Two completely different systems — the bank statement and the raw usage log — landing on the same number is about as good a gut check on a bill as you'll get. On top of all of that sits $439.55 that's AI-adjacent but not really the swarm's own upkeep: $283.55 for Replit, which I use for quick side projects I spin up on the fly, separate from my normal production stack and nothing the swarm itself runs on, and $156 for Firecrawl, which does more than feed the scout its watchlist — it's the engine behind deeper agentic research across several of my projects, on the theory that the more an agent actually knows about an industry, the better its output.
Add it up: $1,022.82 spent directly with model providers, $439.55 on the tooling around them, $1,462.37 total for the month. That's the number, and it's deliberately unglamorous — less than a single day of a junior analyst's salary, for a swarm that never clocks out. The interesting part isn't the total, though. It's how unevenly it's distributed.
aCost by agent role
Everyone assumes the bill is spread evenly across the swarm. It absolutely is not.
The agents that run constantly — the scout watching for signal, the curator organizing it, the observer auditing runs, the listener sweeping social — are the cheapest things I own. In the June 2026 logs they're almost entirely GPT-4o mini and GPT-5.4 mini calls: small read-and-filter jobs over short text, so however often they fire, they barely move the meter.
The agent that runs rarely is the expensive one, and the Claude Opus token log proves it. On a quiet day like June 16th, that API key processed 8,552 input tokens and produced 142 tokens of output — a rounding error. On June 8th, the same key processed 1,263,887 input tokens and produced 31,831 output tokens — roughly 150 times the input of the quiet day, in a single 24-hour window. The same kind of spike shows up again on June 22nd, 26th, 29th and 30th, roughly the cadence you'd expect from a job that fires about once a week: the synthesizer working through a batch of collected posts plus the house-style rules, then producing thousands of words of output in a single pass. One of those runs costs more than a month of the frequent little agents put together.
That's the whole lesson in one line: token cost tracks the size of the job, not how often the job runs. A frequent, tiny agent is cheap. A rare, greedy one is where your money goes. If you're budgeting, count the big-context jobs, not the ticks on the schedule.
bCost by model tier
The other half of the equation is which model each job runs on, and the gap is enormous. In my swarm, the routine, high-volume filtering runs on GPT-4o mini, at roughly fifteen cents per million input tokens and sixty cents per million output, with GPT-5.4 mini handling a smaller slice of the same kind of work. At those rates, a scout or listener pass costs a rounding error.
The mid-weight curation and drafting work — GPT-4o and GPT-5.1 — runs closer to $2.50 in / $10 out per million tokens for GPT-4o, with GPT-5-class pricing in roughly the same band. The one job that's genuinely reasoning-heavy — the weekly synthesis — runs on Claude Opus at $5 in / $25 out per million tokens, the most expensive model I touch, reserved for the one task actually worth it. Notice the pattern across every tier — output is billed at roughly five times input. That single fact matters more than which vendor you pick, and I'll come back to it. If you want the full list of what runs where, the exact stack I build all of this with is public.
VWhere the money actually goes (the part vendor pages don't tell you)
Here's what a pricing page will never tell you: the sticker price per token is almost never why a bill blows up. Three other things are.
The first is output length. Because output costs about five times input, a chatty agent that pads every answer is quietly the most expensive thing you own. Trimming what an agent is allowed to say back does more for the bill than shaving pennies off the input rate.
The second is retries. An agent that hits an error and just tries again — and again — burns real money going nowhere. This is the failure mode I designed hardest against. When one of my agents can't reach the outside sources it needs — a blocked feed, an auth wall, a dead endpoint — it doesn't sit there hammering a dead source on my dime. After a fixed number of failed tries it stops, logs the run as partial, and pings me on Slack. Those partial runs turn up in the log now and then, and that's fine: a partial run that stopped itself costs almost nothing. A runaway one that retries into the void is how a trivial bill becomes a horror story. The stop rule is the entire difference between the two.
The third is orchestration overhead. Multi-agent setups aren't twice the cost of a single agent — a production benchmark found they typically run five to ten times more once you count coordination, failure-handling and the shared context each agent has to re-read. That re-reading is the silent tax, and it's visible in my own June numbers: one heavily-used model sent 68.5 million cached tokens against 87.7 million total input tokens across the days I have full data on — seventy-eight percent of everything that model read in was repeated context, not new work. Every agent that needs the same background pays for it again, and the longer the context window gets, the more every single call costs — unless you do something about it.
VIHow this stacks up against the numbers everyone's quoting right now
Put a $1,462.37-a-month operation like mine next to the figures making headlines and the contrast is almost comic.
Goldman Sachs's bottom-up simulations peg a single software-development agent at about $13 a day, and a data-entry agent at up to $60 a day — one such agent, for a month, costs more than my entire swarm. At the top end, OpenClaw's Peter Steinberger disclosed his three-person team burned over $1.3 million on tokens in a single month, roughly 603 billion tokens in 30 days. Uber rolled Claude Code out to about 5,000 engineers, exhausted its full-year AI budget in four months, and had to cap spend at $1,500 per person per tool. And one unnamed company reportedly blew $500 million on Claude in a single month after forgetting to set usage limits on employee licenses.
My operation cost $1,462.37 total in June, and that's the point. Altman has said OpenAI's single heaviest user now runs about 100 billion tokens a month, up from 100,000 six years ago — and Goldman expects agent adoption to push global token demand up 24x by 2030. The curve is real. But none of those horror stories are a law of physics. They're what happens when nobody's watching the meter.
VIIWhat actually cut the bill — and what didn't move the needle
Here's what genuinely worked, in order of impact.
Routing routine work to cheap models. This is the biggest lever by a mile. Send the easy, high-volume jobs to a small model and reserve the frontier model for the handful of reasoning-heavy tasks. It's exactly what shows up in my own June bill: GPT-4o mini and GPT-5.4 mini handle the bulk of the swarm's daily call volume, while Claude Opus — the priciest model I touch — gets reserved for the one weekly job that actually needs it. It isn't a hack — peer-reviewed research on model routing (RouteLLM, from UC Berkeley and collaborators) showed it can cut cost by more than 85% while keeping about 95% of flagship quality, and production teams report 40–85% bill reductions with no visible drop in output.
Caching the repeated context. My agents re-read the same background every run — the watchlist, the house-style rules, the standing instructions. Prompt caching bills that repeated chunk at up to 90% off versus sending it fresh each time, and it shows up directly in the June numbers: one model alone ran nearly 78% cached tokens against its total input for the days I have full data on. For a swarm where every agent reloads the same context, that's most of the orchestration tax handed straight back.
Killing runaway retries. The stop rule from the last section is as much a cost lever as a reliability one: after a fixed number of failed tries, an agent quits, logs a partial run, and flags me instead of trying again into the void. Boring, and it quietly prevents the exact spiral that turns a trivial bill into a five-figure one.
Batching the non-urgent jobs. Nothing my swarm does needs to happen this second. Asynchronous batch processing takes a flat 50% off input and output on scheduled work — free money for anything that can wait an hour.
And what didn't move the needle: fiddling with prompt wording. Rewording a prompt to be "more efficient" is the tactic everyone reaches for first, and in my experience it's noise next to routing and caching. Aaron Levie made the same point on X — real savings require "a deep understanding of the underlying work being done in a non-abstract way", not clever phrasing. Cut the work, cut the model tier, or cut the output. Cutting adjectives doesn't do it.
VIIIThe honest verdict: what this means if you're about to run your own agents
Token cost is not a reason to avoid running agents. The unit economics, once you're watching them, are absurdly in your favor — a contained support ticket resolved by an agent costs about $0.46 versus $4.18 for a human, and a routine code review runs around $0.72 against roughly $48 of senior-engineer time.
But it is a reason not to run them unsupervised. The same research found only 41% of agent rollouts reach positive ROI inside a year, and about 19% never pay back at all — and every headline blowout in this piece traces to the same root cause: nobody set a limit and nobody watched the meter. There's even a name now for treating heavy spend as a badge of honor, after Jensen Huang said he'd be "deeply alarmed" if a $500k engineer weren't burning $250k of tokens a year. That's a fine story for a company printing money. It's a terrible default for everyone else.
My swarm ran for $1,462.37 in June — about the cost of one contractor's day rate — because someone reads the log, routes the cheap work to cheap models, caches the boring context, and stops the retries before they spiral. None of that is exotic. It's the difference between running the meter and being run over by it. If you're a fund, an ops team, or a founder sizing up whether to build this yourself, that's the whole decision — and it's the same instinct I brought when I gave four AI tools the same impossible job. Real numbers beat a pricing page every time. You just have to be the one keeping them.
— Michael Rouveure · 10 JUL 2026
