A founder told me last week, "I had to raise my agent's price three times in the last year because Claude kept getting cheaper and I kept feeling guilty charging the same." That's not a complaint. That's the most exciting sentence in agent economics right now.

Anthropic shipped Opus 4.7 on April 16. Same pricing as 4.6: $5 per million input tokens, $25 per million output tokens. But the model itself is materially better at the work indie agent builders care about — sustained agentic tasks, screen understanding, long-running workflows. Combine that with the broader trend (per-task agent cost falling from $0.50 in 2024 to $0.05 in 2026 according to multiple platform benchmarks), and the math on "ship an agent, charge €500/month" finally works without holding your breath.

Most builders haven't noticed yet. They're still costing their agents like it's 2024.

What Actually Changed With 4.7

Anthropic's launch post called Opus 4.7 "a fundamental shift toward asynchronous agentic workflows." Translated from press-release-ese, three things matter to operators.

Self-verification. 4.7 can re-read its own outputs and catch its own errors before reporting back. The practical impact: fewer "the agent ran but produced garbage" failures. The cost of a failed agent run is way higher than the cost of a successful one — refunds, support tickets, churn — so reducing failure rate is reducing operating cost in disguise.

Cross-session context. 4.7 carries context across sessions to manage complex multi-day projects. For an agent that, say, runs a weekly onboarding sequence for new customers, this turns a hand-coded state machine into a single prompt. The build time drops dramatically.

Screen understanding at 92%+. Through CUA-style computer use, 4.7 plus the Anthropic computer-use API can drive UIs reliably. The relevance for indie operators: any workflow that requires the agent to log into a legacy SaaS and click around — which is most B2B back-office work — is now sellable.

The Margin Math That Finally Works

Let's walk through what selling a workflow agent at €500/month actually looks like in 2026 numbers.

Average agent run cost. A typical back-office agent run — pull a document, classify it, route it, log the result — consumes roughly 8,000 input tokens and 1,500 output tokens on Opus 4.7. That's $0.04 per run in raw model cost. Tooling and infrastructure overhead adds another $0.02–$0.03. Call it $0.07 per run, fully loaded.

Volume per customer. A small-business customer paying €500/month is doing perhaps 800–1,200 runs (8 hours of human work replaced × 20 working days). At $0.07/run, that's $56–$84 in cost per month per customer.

Gross margin. €500 minus €75 in delivery cost equals €425 gross margin per customer per month. That's 85% — better than typical SaaS, dramatically better than what these numbers looked like 18 months ago when Opus 3 was charging 5× the rate.

What scales. Add a Stripe processing fee, an email tool, an n8n hosting line — call it €25/month all-in overhead. Add maybe €20/month for monitoring and support. Net contribution per customer: ~€380/month. At 30 customers, you're netting €11,400/month before tax. That's a real business, run from a laptop.

Where the Cheaper Models Unlock New Categories

Some agent categories were uneconomical at 2024 prices. They're now in-scope.

High-volume document processing. Invoice OCR with line-item extraction, contract clause auditing, insurance claim categorization. These run thousands of documents per customer per month — at 2024 prices the model cost ate the revenue. At 2026 prices, this is one of the highest-margin verticals in the agent economy.

Always-on monitoring agents. An agent that watches a competitor's website and alerts the customer to changes. Or watches inventory levels and orders restocks. These need to run continuously — economically suicidal at $0.50/run. At $0.05/run they're a no-brainer at €299/month.

Multi-step research agents. Sales prospecting agents that take a prospect name, find the company, find the role, find recent news, draft the outbound — these chain 5–10 model calls per run. Cost in 2024: $5. Cost in 2026: $0.50. Margin at €0.99 per enriched lead is finally there.

Voice agents with screen control. Combining real-time voice with browser-driven actions used to require throwing the kitchen sink at it. 4.7's screen-understanding capabilities plus reduced compute cost mean an indie operator can ship a voice receptionist for a dental office at €499/month with a margin that survives growth.

What This Doesn't Solve

Cheaper models don't fix the things that have always killed indie agent businesses. Worth saying.

Distribution is still hard. Your agent costs €0.07/run to deliver. Your CAC is €200–€800 per customer. The model getting cheaper doesn't move that number. Marketing, sales motion, and brand still matter more than tokens.

Reliability still costs engineering time. A cheap model run that fails 5% of the time isn't actually cheap — the support burden eats the margin. Investing in evals, monitoring, and retry logic is still the difference between a real business and a perpetually-on-fire side project.

Specialization still beats price. A horizontal "AI assistant" agent has to charge €19/month to compete. A vertical "AI receptionist for German dental practices" agent charges €499/month and the buyer doesn't blink. The model getting cheaper doesn't help the horizontal player — it just compresses their margin faster.

Compliance still applies. Cheaper compute doesn't change GDPR, the EU AI Act, or the boring contractual stuff. Bake those costs in from day one or they catch up to you at scale.

The Window That Just Opened

There's a recurring pattern in technology economics: a price drop happens, the existing incumbents are too busy defending margin to chase it, and a wave of new entrants captures the new category. The 4.7 release is one of those drops. Existing horizontal AI tools are stuck — they can't lower prices without crushing their existing revenue line. Indie operators starting today get to price into the new economics.

The math is now permissive. A solo operator can build a 30-customer book at €500/month each in 6–12 months. The agent runs reliably enough on 4.7 that you sleep through the weekends. The unit economics survive a downturn. The margin compounds.

The question is no longer whether the models are cheap enough. They are. The question is whether you ship before the next builder figures out the same workflow you're sitting on.