AI-First
As of yesterday, June 15, 2026, your Claude agents no longer draw from the same pool as your conversations. Anthropic has separated billing for interactive usage (standard chat on claude.ai) and programmatic usage (Claude Code, GitHub Actions, SDK scripts, third-party applications). The change barely made headlines, but it will affect the bill of every SMB that has started automating with Claude.
I work with SMBs that use Claude Code daily. Some have seen their programmatic credit drain in three days after the switch. This article lays out the numbers, explains the mechanics, and offers a concrete action plan to stay in control.
- ⚠️ Billing split is live: since June 15, chat and agents have two separate budgets.
- 📊 Programmatic credits: $20 (Pro), $100 (Max 5x), $200 (Max 20x) per month.
- 💡 Full API pricing: agents are billed at the API rate, not the subsidized subscription rate.
- 🎯 SMB action: separate workloads, cap agents, measure before you automate.
What Changed on June 15, 2026
Why did Anthropic separate the budgets now?
On May 14, 2026, Anthropic officially announced the overhaul in a blog post, picked up by The Decoder (via infosia.fr). The announcement boiled down to one sentence: starting June 15, programmatic usage would have separate monthly credits from standard chat.
Until that date, a Pro subscriber at $20/month used the same quota for chatting with Claude on claude.ai and for running agents via the SDK, Claude Code, or GitHub Actions. The flat-rate model held up as long as programmatic usage remained marginal. With the rise of autonomous agents and tools like OpenClaw, a single flat plan became unsustainable for Anthropic.
The timing is no coincidence. According to Bloomberg (April 2026), Anthropic filed a confidential S-1 with the SEC, with an estimated valuation of ~$965 billion. When you are preparing for an IPO, subsidizing programmatic usage at a loss is not part of the plan.
What counts as "programmatic"?
The scope is precise: the Claude Agent SDK, the claude -p command-line tool, GitHub Actions, and all third-party applications built on the Anthropic ecosystem. If your request goes through any interface other than claude.ai or the mobile app, it draws from the programmatic credit.
In practice, if you have an agent running overnight to sort your emails or generate reports, it pulls from a budget that is entirely separate from your daytime conversations.
Interactive vs. Programmatic Usage: Two Budgets, Two Logics
How does the billing model differ between chat and agents?
Interactive usage (claude.ai, mobile app, desktop app) stays under the same regime as before: a quota included in the subscription, with limits that vary by plan. No surprises there.
Programmatic usage works differently. Each plan receives a monthly dollar credit, consumed at the full API rate. A Pro subscriber gets $20 in programmatic credit. A Max 5x subscriber gets $100. A Max 20x subscriber gets $200. For teams, Team Standard gives $20 per seat, Team Premium gives $100 per seat.
The nuance most people miss: these credits are billed at the full API rate, not the subsidized subscription rate. Lydia Hallie, a developer at Anthropic, stated on X that "you don't pay extra." That is technically true (the credit is included in the subscription). But the volumes that previously passed "for free" under a single flat plan are now drawn from a much tighter budget.
How many tokens does $20 actually buy?
Let's look at a concrete example. Based on current pricing compiled by plateya.fr, Sonnet 4.6 costs $3 per million input tokens and $15 per million output tokens. Opus 4.6 goes up to $5/$25.
With $20 in Pro credit on Sonnet 4.6, you can generate roughly 1.1 million output tokens (the most expensive side). For an agent analyzing documents, that is the equivalent of 5 to 10 long reports per month. For a Claude Code workflow generating code, that is 2 to 3 one-hour sessions.
If you use Opus 4.6 for complex reasoning tasks (and according to ethanolle.com, Opus remains the right choice for multi-step agents), your $20 drains even faster: roughly 800,000 output tokens.
An SMB owner who has deployed 3 automated agents will hit that ceiling within a week.
The Real Cost, Plan by Plan
Which plan should you choose to avoid blowing your agent budget?
Here is the full breakdown of the new structure. I added a "Sonnet token equivalent" column to make the credits tangible.
| Plan | Price/month | Programmatic credit | Sonnet tokens (output) | Trend |
|---|---|---|---|---|
| Pro | $20 | $20 | ~1.1M tokens | ↓ tight for agents |
| Max 5x | $100 | $100 | ~5.7M tokens | → workable if managed |
| Max 20x | $200 | $200 | ~11.4M tokens | ↑ comfortable for active SMBs |
| Team Standard | $30/seat | $20/seat | ~1.1M tokens/seat | ↓ insufficient for shared agents |
| Team Premium | on request | $100/seat | ~5.7M tokens/seat | → suited for tech teams |
SOURCE: claude.com/pricing + plateya.fr · Updated 06/2026
What this table reveals: the Pro plan is now a discovery plan, not a production plan for agents. If you have automated even a single daily workflow (email sorting, meeting summaries, competitive monitoring), the $20 credit will not last the month.
I have already written about the true cost of LLMs and about at what volume an SMB should move from Pro to Team. The June 15 split makes both of those analyses even more relevant: the visible cost (the subscription) masks the real cost (the API consumption of your agents).
Claude for Small Business: The Dedicated SMB Offer in the Works
When will Claude for Small Business be available in Europe?
Anthropic is preparing a dedicated offering for smaller organizations: Claude for Small Business. The program includes 15 preconfigured workflows and 7 native connectors (HubSpot, Canva, DocuSign, Google Workspace, Microsoft 365, among others). The EU rollout is planned for late 2026.
Today, a 15-person SMB has to choose between Pro (too limited for agents), Max (expensive for collective use), or Team (which assumes structured IT management). Claude for Small Business promises a middle ground with ready-to-use workflows.
Should you wait for this offering before automating?
No. And I say this with conviction: waiting means losing time and money. SMBs that have already set up agents (even simple ones) have a head start. The June 15 split does not change the value of automation; it changes how you budget for it.
What I recommend to my clients is to start now with a clear, measurable use case, and to actually calculate the ROI before scaling up. A well-built agent that saves you 2 hours a day is easily worth $100/month in Max 5x credit.
According to McKinsey, generative AI could increase productivity by 0.1 to 0.6 percentage points of GDP per year. For an SMB, that is the difference between a team of 10 that produces like a team of 15, with the same headcount.
How to Cap Your Agent Spend as an SMB
How can you cut your Claude bill without giving up automation?
Here are the five levers I have been using with my clients since June 15.
First lever: audit your agents. List every automated workflow, its model (Sonnet, Opus, Haiku), and how often it runs. Many SMBs have agents running "by default" that nobody checks to see whether they still deliver value.
Second lever: switch to Haiku 4.5 for simple tasks. At $1/$5 per million tokens, Haiku costs 5x less than Sonnet for sorting, classification, and extraction tasks. Your $20 of Pro credit then covers ~4 million output tokens instead of 1.1 million.
Third lever: batch your calls. An agent making 50 short requests consumes more (context overhead) than an agent making 5 long, well-structured requests.
Fourth lever: separate your workloads. If a team member uses Claude Code for coding in the morning and chats with Claude in the afternoon, the two counters no longer cannibalize each other. That is actually an upside of the split.
Fifth lever: measure before you automate. My mantra: if you cannot estimate how many tokens an agent will consume per month, you are not ready to deploy it. Run a 7-day test, calculate the actual consumption, then decide on the right plan.
Does the IPO context change things for users?
Anthropic's confidential S-1 filed with the SEC signals that pricing will continue to evolve. A company on the road to an IPO at ~$965 billion needs to show a rising revenue-per-user trajectory. For SMBs, that means one thing: prices are unlikely to come down.
Better to build your automations with a clear budget envelope now rather than counting on price cuts. If you have integrated Claude into your processes (and if you have not, GoLive Software can help), protect your margin by capping agents, not by removing them.
"The June 15 split does not penalize automation. It penalizes automation without measurement."
Vincent Roye, June 2026
My clear-cut take: this separation is healthy. It forces companies to treat their AI agents as identified cost items, not as a free bonus bundled with the subscription. SMBs that measure, cap, and optimize will come out ahead. Those that let agents run unchecked will find an unpleasant bill waiting for them at the end of July.
Frequently Asked Questions
How does Claude bill agents in 2026?
Since June 15, 2026, Anthropic assigns a separate monthly credit for programmatic usage (agents, SDK, CLI). This credit is consumed at the full API rate: $3/$15 per million tokens for Sonnet 4.6, $5/$25 for Opus 4.6, $1/$5 for Haiku 4.5. The credit varies by plan: $20 for Pro, $100 for Max 5x, $200 for Max 20x.
Are Claude Code and GitHub Actions billed separately?
Yes. Claude Code, GitHub Actions, the claude -p SDK, and all third-party applications draw from the programmatic budget. They no longer consume your interactive chat quota. If you use Claude Code for coding and claude.ai for conversation, the two counters are fully independent.
How can SMBs reduce their Claude bill?
Three quick wins: switch simple tasks to Haiku 4.5 (5x cheaper than Sonnet), batch your agent calls to reduce context overhead, and audit your workflows to cut any that no longer deliver value. For an SMB running 3 to 5 active agents, the Max 5x plan at $100/month offers a good balance between budget and capacity.
What is the price of Claude for Small Business?
Anthropic has not yet published pricing for Claude for Small Business. The offering will include 15 preconfigured workflows and 7 connectors (HubSpot, Canva, DocuSign, Google Workspace, M365). The EU rollout is planned for late 2026. For SMBs that cannot wait, Team Standard ($30/seat plus $20 in programmatic credit per seat) remains the most accessible option.
Is the June 15 split permanent or temporary?
Anthropic has communicated no end date. The context (IPO preparation, annualized revenue growth of ~$30 billion according to Bloomberg) suggests this pricing segmentation is structural. Credits or thresholds may evolve, but the principle of separate budgets is here to stay.
Vidéos YouTube
Articles & ressources
- New Claude Pricing: Programmatic Usage Moves to API Rates · infosia.fr
- Pricing | Claude · claude.com
- Claude AI Pricing 2026: Plans, Rates, and Full Comparison · ethanolle.com
- Claude AI Guide 2026: Models, Code, Pricing, Security · plisio.net
- Claude Pricing 2026: Pro, Max, Cowork, Code & API · plateya.fr