I priced out a realistic autonomous agent at current LLM rates this week. The number came in 10x higher than the naive math suggested, and that gap turns out to be the entire story of why autonomous-agent businesses fight for their gross margin in 2026.

This is post 2 in the series on the autonomous-AI-runs-your-company thesis. Post 1 made the case that three structural constraints (token economics, gated data, and bought distribution) explain why the autonomous business pitch hits a wall today. This post goes deep on the first constraint with the actual token spreadsheet.

What the naive math gets you

Suppose you're building "an AI agent that runs a small business." Per active customer, per month, the agent has to:

Research and qualify prospects. Roughly 300 sessions per month (10/day × 30 days). Each pulls in scraped pages, partial profiles, conversation history. Call it ~50K input tokens and ~3K output tokens per session.
Draft outbound messages. ~100 cold emails per month per customer. ~2K input, ~500 output each.
Handle inbound replies. ~50 inbox events per month. ~3K input, ~500 output each.
Generate creative. ~30 ad variants per month for paid-channel rotation. ~5K input, ~1K output each.

Add it up:

Input: ~15.5M tokens
Output: ~1.0M tokens

At GPT-5-mini's list rate of $0.125 per million input / $1.00 per million output:

Input cost: 15.5M × $0.125/M ≈ $1.94
Output cost: 1.0M × $1.00/M ≈ $1.00
Total: ~$2.94 per active customer per month.

If that were the real number, an autonomous agent would be a fantastic business. At $57/mo ARPU, inference would be 5% of revenue. You'd have 50+ points of gross margin headroom. Distribution would be the only constraint to worry about.

But that's not the real number, and the gap between $3 and the actual cost is where every autonomous-agent founder gets sandbagged.

The hidden multiplier: reasoning tokens

Cheap chat models like GPT-5-mini are great at "rewrite this email" or "summarize this thread." They struggle with "look at this lead's whole context, decide whether to reach out, and if so plan a multi-step sequence." That work needs reasoning models: o3, o3-mini, GPT-5.4 with thinking, Claude Sonnet/Opus with extended thinking, or the open-source reasoning models catching up fast.

Reasoning model headline prices look almost reasonable:

OpenAI o3: $2 input / $8 output per million tokens
OpenAI o3-mini: $0.10 input / $4.40 output per million
Claude Sonnet 4.6 (with extended thinking): $3 / $15 per million
Claude Opus 4.8 (with adaptive thinking, default "high" effort): $5 / $25 per million

The trap is in how reasoning tokens get billed. The internal "thinking" the model does, the hidden chain-of-thought it generates before producing a visible answer, counts as output tokens at output rates. For a non-trivial reasoning task, those hidden tokens run 3-10x the visible output. Anthropic's own guidance for extended thinking is to budget 2-5x visible output for thinking-heavy tasks. OpenAI's o3 quietly burns 3,000-10,000 thinking tokens on harder prompts.

So a "small" agent decision, "should I reach out to this lead, and if so what's the plan," that looks like 500 visible output tokens is actually 2,000-5,000 billed tokens at the output rate.

Now redo the per-customer math, with realistic model mix. Assume 60% of the work is the cheap-model drudgery above (drafting, summarizing) and 40% is reasoning-tier (qualification, planning, multi-step decisions), with a 4x thinking-token multiplier on the reasoning workload:

Cheap-model portion (60% of volume): 0.6 × $2.94 = ~$1.76
Reasoning-model portion (40% of volume), at o3 prices with 4x thinking inflation:
- Input: 0.4 × 15.5M × $2/M ≈ $12.40
- Output (visible + thinking, 4x multiplier): 0.4 × 1.0M × 4 × $8/M ≈ $12.80
- Subtotal: ~$25.20
Total: ~$27 per active customer per month.

Use Opus instead of o3 for the hardest reasoning slice and the total drifts past $40. Use Sonnet 4.6 with extended thinking and you're in the high $20s to mid-$30s.

So the publicly-disclosed $35/month per active user that one $250M-valued autonomous agent platform discloses isn't an outlier or a sign of incompetence. It's exactly where realistic agent economics land in 2026 given real workloads on real models. The cheap headline rate of $0.125 per million input tokens is the GPT-5-mini honeypot. True, technically, for the small slice of agent work that doesn't require thinking.

The reliability tax that doesn't dissolve

The math above assumes every agent action succeeds on the first try. Real agents don't.

Tool calls fail. JSON outputs malform. The model hallucinates a field name and the downstream parser blows up. The agent loops on a "search → consider → decide" cycle and the second loop's context now includes the first loop's reasoning, so input tokens compound. A robust agent that ships in production has retry logic and fallback paths, and every retry buys you another round of input + output + thinking tokens at full price.

On the agents I've watched run, the reliability tax adds 30-100% to the naive token cost, sometimes more for the long-tail of edge cases. This is the part of the math that doesn't dissolve with cheaper inference, because it's a function of how often the model is wrong, not how cheap each call is. As models get better the tax shrinks, but it never zeros.

That puts realistic 2026 autonomous-agent compute somewhere in the $25-$60 per active user per month range, depending on model mix, task complexity, and how much retry-tolerance you've built in. At a $57 ARPU, that's 44-100%+ of revenue going to inference before you've paid for anything else. Which is exactly why every public autonomous-agent platform is also selling something else on the side, like ad markup, domain markup, add-ons, and packs, to widen the per-user revenue base so the inference share looks less terminal. Read that pattern closely: the platforms aren't widening revenue because autonomy is working. They're doing it because the compute bill of trying to automate judgment never closes at SMB prices.

The price-drop curve, and what it actually buys you

The good news, and it's real news: inference costs are on a brutal price-drop curve. Epoch AI's tracking shows the cost of achieving a given level of model quality has dropped roughly 10x per year since 2023, and for some performance milestones, 40-900x per year. GPT-4-equivalent quality cost ~$20-60 per million tokens at launch in March 2023 and runs at ~$0.40 per million today. That's a ~50x compression in 3.5 years for equivalent intelligence.

Reasoning-model pricing is dropping slower than chat-model pricing (the curve is steeper for o-series than for GPT-4o-equivalent), but it's still dropping fast. Open-source reasoning models (DeepSeek's reasoning variants, the various R-tier open releases) keep forcing the commercial labs to compress. The directional bet here is solid.

If the trend continues another 12 months, your $35/user/month inference bill becomes ~$3.50. At a $57 ARPU, that's a normal SaaS-shaped gross margin. Within 24 months, the constraint largely dissolves for cost-disciplined builders. This is the most likely of the three constraints to break in your favor.

The catch, and this is the part the autonomous-agent pitch doesn't talk about, is that the price curve only saves you if your customers stay long enough to benefit from it.

At the disclosed 50% month-one churn rate of the platform I keep citing, half your cohort is gone before the model price has moved meaningfully. You're racing the price curve with a leaky bucket. Two years of 10x/yr inference drops doesn't help if your median customer lasts six weeks.

That's the real shape of constraint A. It's not "inference is too expensive." That's solving. It's "high-churn businesses lose the race against the price curve, and high-retention businesses win it."

The move the math actually points to

Here's the conclusion the cost spreadsheet forces, and it's the opposite of the autonomous pitch: at SMB ARPU, full autonomy is the wrong target. You cannot run a $57/mo product on $35 of compute and burn another 30-100% on retries papering over a machine making every judgment call alone. The math says full-stack autonomy is ROI-negative at the price point where most small businesses actually buy.

But flip the question. Most of the cost above isn't the judgment. It's the grind around it: researching prospects, sifting noise, drafting message after message, keeping sequences from torching your sender reputation. That grind is exactly what cheap, deterministic, well-scoped automation handles beautifully and cheaply. The expensive part, the open-ended "decide what this business should do next" reasoning, is the part that doesn't pay back at SMB prices and the part a founder is genuinely better at anyway.

So the ROI-positive shape isn't an agent that runs your company. It's autopilot for the work that compounds:

Find the buyers already asking. The single highest-leverage thing automation does is surface the people raising their hand right now: the ones describing your exact problem in public, in real time. You don't need a reasoning model to decide who to cold-pitch into the void. You need a feed of warm intent landing in your inbox. That's signal, not autonomy, and signal is cheap to deliver and worth a fortune to receive.
Draft in your voice, ready to send. Drafting is the cheap-model slice of the math: summarizing context, writing the reply that fits the moment. Done well, it produces a message that sounds like you and clears a judge before it ships, on autopilot or with your read first. The leverage is collapsing the time from "someone's asking" to "you've answered."
Pace the sends so accounts don't get banned. Volume is what gets cold accounts flagged, deboosted, and banned. Autopilot that meters the pace, respects each platform's tolerance, and keeps deliverability intact is worth more than autopilot that blasts. The win condition is replies and booked calls, not message count.

The founder's judgment stays in this model, encoded once as the bar the system holds on every draft, because judgment is the leverage and it's the one input that doesn't get cheaper on the price-drop curve. Spending a reasoning-model budget to imitate a founder's call on a $57/mo account is lighting money on fire. Spending it to put the right ten conversations in front of that founder every morning is the trade that prints.

That's the business the math actually supports: not the agent that replaces the operator, but the autopilot that finds the demand, drafts the answer, and protects the channel, then puts leads and customers in the inbox while the operator does the part that's worth a human's hour.

That's where Thread Otter sits on purpose. It watches Reddit, X, LinkedIn, and your alert feeds for the people describing your problem out loud, drafts the reply in your voice, and paces the sends so your accounts stay healthy, then drops the live conversations into one inbox. Autopilot on the grind, you on the judgment, leads instead of a compute bill.

Next in the series: constraint B, the data wall. Why high-quality B2B prospect data lives behind paywalls, why scraping public data hits a quality ceiling that more inference can't break through, and the one architectural bet that genuinely changes the equation.

This is post 2 of a 5-part series on the AI-runs-your-company thesis. Start with post 1 for the full audit.