Agents Are Moving Into Your Texts. WhatsApp Just Started Charging For Replies.

On September 17, a startup called Photon held a funeral for mobile apps in a San Francisco church. There were speeches, panels with Vercel, Stripe, and OpenAI, and an actual coffin for app icons.

The joke had a business behind it. On October 1, TechCrunch reported that Photon raised $4.5 million in seed funding to help developers build AI agents that live inside iMessage, WhatsApp, SMS, email, and voice. The company says it has signed up more than 40,000 developers and grown revenue 10x in four months.

On the same day, a quieter change took effect. Meta started charging businesses for WhatsApp service messages, the replies a business sends inside an open conversation. Those replies had been free since November 2024.

Put those two stories side by side and you get a real question for early founders: should your next product be a conversation instead of an app? And if so, what does each reply cost you now?

Photon's thesis is simple. People are tired of installing apps. They already spend their day in a handful of messaging apps, so put the product there.

The idea came from the founders' own pain. CEO Daniel Tian and CTO Ryan Zhu built consumer apps as students, often at hackathons. Every time they shipped one, they hit the same wall: nobody found it. Then Tian built a bot that texted his friends back as him over iMessage. When the pair open sourced it on GitHub, thousands of developers picked it up.

Today Photon sells a unified API, a channel framework, a command-line tool, and monitoring so developers can run one agent across many messaging channels. The open source version still accounts for 98% of usage. The hosted version, launched in April, adds 99.95% uptime and SOC 2 Type II and HIPAA compliance. Anything beyond 10 users requires a paid tier. The company reports under 3% churn and says message volume grew 5x in the last month.

For a first-time founder, the pull is obvious.

Distribution. You skip the app store, the install, the sign-up screen, and the push notification permission prompt. A user texts a number and they are in.

Retention by default. Your product sits in the same thread list as the user's friends and family. You do not need to win a spot on their home screen.

Less to build. No mobile UI, no app review, no separate iOS and Android builds. For an AI product where the core value is a conversation, a chat thread is the interface.

Fast learning. Every message is a transcript. You can read exactly what users asked, where the agent failed, and where they gave up.

These are real advantages, especially for products like scheduling, reminders, customer support, personal finance check-ins, or concierge services where the user mostly asks and the product mostly answers.

Here is the part founders tend to skip. Messaging channels are not free pipes. Each has its own owner, rules, and price list, and those can change.

Meta's developer documentation lays out the new WhatsApp timeline:

• August 1, 2026: Meta began charging for messages powered by its own Meta Business Agent, at $2.00 per million tokens. Meta estimates a typical message uses 20,000 to 25,000 tokens, or roughly 4 to 5 cents.
• October 1, 2026: Meta began charging per message for service messages, meaning any free-form reply sent by a person or a third-party AI agent inside the 24-hour customer service window. Service rates match each country's utility and authentication rates.
• Also October 1: utility messages sent inside that open window, like order confirmations, are now charged again. They had been free since July 2025.

Trade coverage, including from the messaging provider YCloud, reports a free allowance of 1,000 service messages per business phone number per month, which resets and does not roll over.

Meta's own worked example shows the shift. A short conversation that started with a cart reminder, included two AI answers, a human handoff, and an order confirmation cost 3 charges in August. The same conversation costs 5 charges after October 1.

If you are building an AI agent on WhatsApp with your own model, you now pay twice: once to your AI provider and once to Meta for delivery. Meta's documentation spells this out, and its own comparison for Brazil estimates 10,000 lower-complexity AI replies at about $268 all in, and 10,000 higher-complexity replies at about $968.

Here is a simple way to model your own product. These numbers are an illustrative example, not real rates for your market.

• An active user sends 10 messages a week.
• Your agent sends about 12 replies a week to that user.
• That is roughly 50 replies per user per month.
• If delivery costs 0.5 cents per reply and your AI costs 2 cents per reply, each active user costs about $1.25 a month in variable costs.

Now compare that to what the user pays or what they are worth to you. A $5 a month consumer subscription still works. A free product monetized by occasional affiliate revenue may not.

Run this for your actual country mix. Meta prices by the recipient's country, so a product with users in many markets has a blended cost, and Meta says rates may update up to quarterly.

Messaging-first does not mean one channel. Photon's pitch is that developers can work across many through one API. That matters because each channel has different constraints.

WhatsApp limits free-form replies to a 24-hour window that opens each time a user writes to you. Outside that window, you need pre-approved message templates, which are charged by category.

SMS in the US requires registration if you send texts from software on a standard 10-digit number. Microsoft's Azure documentation, like most providers, explains that businesses must complete brand registration and campaign registration before sending this kind of traffic. Carriers filter unregistered messages. Plan for days of approval time, not minutes.

iMessage is where many US consumers live, but it is Apple's platform, and you should read your provider's terms carefully about how they reach it.

The lesson: pick your first channel based on where your users already talk, then check its rules before writing code. If you want a structured way to work through channel costs and assumptions before you build, Foundra can help you turn them into a simple plan you can test.

Before you commit, answer these with your team.

1. Is the core job a conversation? If users mostly ask questions, get reminders, or confirm choices, chat works. If they need to compare options visually, edit documents, or browse catalogs, a chat thread gets painful fast.
2. Who pays per message, and is it covered? Write down your delivery cost, AI cost, and revenue per active user. If the margin is thin at 50 replies a month, it will break at 500.
3. What happens when the channel changes its price or rules? WhatsApp just did. Could you move users to SMS, email, or a web link without losing them?
4. Do you own the relationship? Capture an email or account outside the channel early, with clear consent. Otherwise your entire user base lives inside someone else's product.

Photon's own numbers show a useful pattern. Most usage is on the free open source version, and the company charges once a developer passes 10 users. That is a clean line between experimenting and running a business, and it is worth copying in your own pricing.

A few failure modes show up again and again.

Surprise bills. A viral week on a per-message channel can produce a large invoice before revenue catches up. Set spending alerts and per-user message caps from day one.

Compliance gaps. Healthcare, finance, and insurance use cases bring privacy rules. Photon advertises HIPAA compliance on its hosted tier for this reason. If you handle sensitive data, confirm what each layer of your stack covers.

Opt-in shortcuts. Texting people who did not ask to hear from you can get your number blocked and expose you to legal risk. Get clear opt-in and honor opt-out keywords.

You can test a messaging-first idea without a big build.

Days 1 to 3: Pick one channel and one job. Write the 10 most common questions a user would send and the ideal reply for each.

Days 4 to 7: Run it by hand or with a simple agent for 10 to 20 real users. Count messages per user per week.

Days 8 to 10: Plug those counts into the cost model above with your real market rates. Decide if the margin works.

Days 11 to 14: Ask users two questions: would they pay, and would they rather use a website or app for any part of this? Their answers tell you whether chat is the product or just the front door.

What is Photon? Photon is a startup that helps developers build AI agents that work over iMessage, WhatsApp, Telegram, SMS or RCS, email, and voice. It raised $4.5 million in seed funding, announced October 1, 2026.

What changed with WhatsApp pricing on October 1, 2026? Meta began charging per message for service messages, the free-form replies businesses send inside a 24-hour customer service window. It also resumed charging for utility messages sent inside that window.

Is there still a free allowance on WhatsApp? Trade coverage reports 1,000 free service messages per business phone number per month. Check Meta's current pricing page for your account before you plan around it.

Do I need to register to send business texts in the US? Yes, if you send texts from software on a standard 10-digit number. Most providers require brand and campaign registration first.

Should every startup skip the app? No. Messaging works best when the core job is a conversation. If users need to browse, compare, or edit, plan for a web or app experience alongside chat.

1. Photon held a funeral for mobile apps. Now it has $4.5M to help replace them with agents., TechCrunch, October 1, 2026
2. Pricing for non-template messages, Meta for Developers (WhatsApp Business Platform)
3. Pricing on the WhatsApp Business Platform, Meta for Developers
4. WhatsApp API Pricing Update: Effective October 1, 2026, YCloud
5. Ten-Digit Long Code (10DLC) Registration Guidelines, Microsoft Azure Communication Services