If you are an Amazon brand owner paying an agency, two things happened in the last 10 days that change the math on what that agency is actually selling you. Most agency principals I have spoken to this week are reading both stories wrong. The operator implication matters more than the news.
What happened
On May 13, 2026, Anthropic announced that starting June 15, 2026, programmatic Claude usage — Claude Code, Agent SDK, GitHub Actions, third-party tools built on the Agent SDK — gets its own separate monthly credit pool denominated in dollars and billed at full API rates. Pro tier gets $20/mo in agent credits. Max 5x gets $100. Max 20x gets $200. Credits do not roll over. Interactive chat usage (web, desktop, terminal Claude Code in attended mode) still runs against the old subscription cap. Source: Anthropic via The New Stack.
Two days later, on May 14, OpenAI announced OpenAI Deployment Co., an in-house AI consulting firm, plus the acquisition of Tomoro, an applied AI engineering shop. Source: AI Business.
Why most brand owners will read this wrong
The dumb take, the one that has been ricocheting around LinkedIn this week: "AI is getting more expensive, agencies will pass it through, brand owners will pay more."
That is not the signal. The signal is the opposite. The platforms are repricing AI agent work so that the per-task economics now resemble per-task labor economics. And both Anthropic and OpenAI are saying — through their pricing structure and through their consulting acquisitions — that they expect the implementation layer to commoditize. They are entering it themselves.
Translation: the agency that was running 14 brands through one $200/month Max account, charging each brand $1,500/month for "AI-powered everything," and pocketing 90%+ gross margin on the AI line — that model dies on June 15. The same agency, after June 15, sees its Anthropic bill multiply 12x to 175x depending on workload, per the community-note analysis circulating since May 13.
What actually changes for someone running $200K/mo on Amazon
Concrete impact on the brands I have visibility into.
Agency pricing pressure. Three agencies I work alongside have already told their clients they are raising AI-tooling line items by 30-60% effective July 1. One agency stopped offering "always-on AI listing optimization" entirely. Expect your invoice to grow, or expect features to disappear quietly.
The 'free AI insights' deliverables go first. The unpaid weekly Rufus-retrievability scan, the auto-generated competitor monitoring, the AI keyword expansion — these were loss-leaders because compute was effectively unlimited. They are now line items with margin attached. If your agency was throwing in 6 AI-generated deliverables a month, expect 2 of them to vanish or move behind a paywall.
In-house automation suddenly looks more expensive too. If you built your own Claude Code workflow to pull SQP, score keywords, draft listing copy — your $200/month Max 20x account stops covering that on June 15. The actual cost of running my own listing-copy automation, which I measured this week against the new pricing: $0.34 per SKU pass with Sonnet 4.6, $1.18 with Opus 4.7. A 400-SKU monthly refresh is now $136-$472 in API spend, not free.
Margin disclosure becomes a real question. When AI compute was a flat $200/month for an agency, nobody asked what the margin was. When it is a metered line item, brand owners will ask. And they should.
What I would do this week if I were running a brand
Five things. None of them is "panic."
-
Email your agency and ask one question: "What does your Anthropic and OpenAI bill look like in May, and what will it look like in July?" Watch how they answer. The agencies that can answer specifically are the ones who understand their own cost structure. The ones that deflect are the ones whose margin model is about to break in public.
-
Audit which AI deliverables you actually use. Most brands I audit have 6-8 AI-generated reports landing in their inbox monthly. They open 1-2. The other 4-6 are about to get more expensive to produce. Tell your agency to kill the unused ones now, before the agency kills them quietly post-June-15 and charges you the same.
-
Renegotiate AI line items, do not absorb them. If your agency raises the AI tooling line by 30%, you are entitled to ask which deliverables that increase funds. Get a per-deliverable cost. The market is in transition; you have leverage right now you will not have in 90 days.
-
Bring one AI workflow in-house this quarter. Pick a workflow you understand — say, weekly SQP keyword extraction, or competitor-listing change monitoring. Run it on a metered Anthropic or OpenAI account where you see the bill. Two months of doing this will recalibrate what you think AI deliverables are worth.
-
Watch for the OpenAI Deployment Co. pitch. It will land in your inbox in Q3 if your brand does over $5M. The pitch will be: skip the agency, work with us directly. Take the meeting, but understand what it actually is — OpenAI is selling implementation hours to feed token consumption. The agency model is not dead. The undifferentiated agency model is dead.
What I would ignore
The "Anthropic is greedy" narrative on Twitter. They are not greedy; they are pricing into the actual cost of inference. Compute does not get cheaper at the rate prompts get longer. This was always coming.
The "Claude Code is dead, switch to Codex" posts. Codex pricing is moving in the same direction — OpenAI's Plus tier has separate weekly limits on Codex usage since their April update. The platforms are coordinating, not competing, on metering structure.
The vendors selling "AI cost optimization platforms" that hit your inbox this week. The actual fix is prompt-length discipline, batch processing, and using Haiku 4.5 for 70% of routine work. You do not need a platform for this. You need a one-page audit.
The contrarian operator take
Here is what nobody is saying out loud yet. The agencies hurt most by June 15 are the ones whose value prop was "we use AI." That was never a moat. It was a pricing arbitrage on flat-rate compute that the platform just closed.
The agencies that will be fine are the ones whose value prop was always judgment on top of tooling — which creative actually converts, which keyword cluster matters, which test to run. Those agencies were absorbing AI as COGS, not selling it as a feature. Their pricing barely moves.
If you cannot tell which kind of agency you are paying, you will find out on your July invoice.
I have been running my own Anthropic and OpenAI bills hot for 18 months across image generation, copy automation, and competitor monitoring. Happy to share the cost numbers — what each workflow actually costs to run at metered rates — if you ask. Email me.