If you run an Amazon agency that's been pitching "AI-leveraged services" since last summer, your gross margin on those services dropped by 30–40% between April and last week. Most of you don't know yet because your finance team is still looking at list-price tables.
Here's the operator read on what actually happened, why your AI-first competitors are about to break, and what I'd do this week if I were running the P&L.
What happened — in two sentences
Claude Opus 4.7 ships at $5 input / $25 output per million tokens — same list price band as Opus 4.6. But the model ships with a new tokenizer that, per multiple independent analyses, uses up to 35% more tokens to encode the same fixed text. Net effect: real billed cost for the same prompt and the same output is roughly 1.4x what it was on Opus 4.6, even though the price card looks unchanged. Simon Willison's May 27 piece on Anthropic and OpenAI finding product-market fit covers the same dynamic — both vendors have moved their economics without moving their price cards.
That's the news. Now the part nobody is writing about.
Why most agency owners will read this wrong
The dumb take is "AI got more expensive, raise our AI service prices." That misses what actually shifted.
The real signal is this: vendor pricing has decoupled from list price. Token count, harness overhead, context window utilization, cache hit rate, tool-call expansion — the things that determine your bill — are no longer in the published rate card. They're in operational details your engineers control and your sales team can't see.
If your agency has been selling AI-leveraged services with a fixed-fee retainer and a vague "we use Claude/GPT" line in the SOW, you have just become structurally short volatility. Your inputs reprice quarterly. Your outputs don't.
The agencies that are about to break are the ones that priced their AI services off list price and assumed list price was the same thing as cost. It isn't. It hasn't been since April. And the gap is widening.
What actually changes for a $200K/mo Amazon brand running AI-leveraged services
I'll walk through it with real numbers, because the abstract version is useless.
Take a brand running 1,200 SKUs across two marketplaces, paying an agency $14K/mo for creative + listing optimization + advertising. The agency uses AI to draft listing copy, generate hero-image direction briefs, run a weekly competitor-watching loop, and pre-grade ad creative before launch.
In April 2026, that workload ran on Opus 4.6 at roughly $1,100/mo in raw API cost. Margin on the AI portion of the engagement was healthy.
Same workload on Opus 4.7 with the new tokenizer? Closer to $1,520/mo. Same brand. Same outputs. Same price to the client. The agency just gave back 3% of the gross retainer to Anthropic, and that's before the GPT-5.5 side of the stack which is up 20% on output.
Compound that across a 90-account book and you're looking at $35–45K/yr of margin evaporating, silently, into a line item nobody's looking at.
The brand doesn't care. The brand sees stable creative quality and stable reporting. The agency's CFO sees raw API spend climbing without a corresponding revenue line. This is the conversation I expect 60% of AI-forward agencies to have with their board between now and Q3.
What I'd do this week if I ran the P&L
Five moves. None require an engineering rebuild. All can ship inside 14 days.
1. Audit your top 10 AI workflows for prompt-cache hit rate. On Opus 4.7, the system around the model accounts for 60–70% of cost variance — model choice is only 30–40%. If you're not running aggressive prompt caching on long system prompts (brand voice docs, category playbooks, SOP libraries), you're paying full-rate input tokens on every call. We're seeing 70–85% cache hit rates on the workflows we re-architected last quarter. That alone takes the tokenizer hit down to neutral.
2. Route by job, not by brand. Opus 4.7 for ambiguous strategic work. Sonnet 4.6 for structured rewrites and category-specific listing optimization. Haiku 4.5 for classification, deduplication, retrieval, and any workflow where the output is a label or a yes/no. Most of the AI workflows in an Amazon agency don't need Opus pricing. If you're routing everything to your best model because "it just works," your cost-per-deliverable just went up 40% without anyone deciding to pay it.
3. Move from fixed-fee AI line-items to "AI included" priced into retainer. If your SOW says "AI services: $2,000/mo" you've handed the client a line item to renegotiate every quarter. If your SOW says "creative and listing optimization: $11,000/mo, delivered using whatever stack delivers the outcome," you control the input mix. Same revenue, fundamentally different exposure to vendor reprice events.
4. Renegotiate now, not at renewal. If you have brands on annual retainers signed before April 2026, your input costs went up and your output price is locked. Pick the 5 largest accounts and have a renewal conversation in the next 30 days, framed around expanded scope — not a price hike. Add a workflow, add a deliverable, add a new SKU coverage tier, and reprice the whole engagement up 12–18%. Brands accept this when it comes with more output. They don't accept "AI got more expensive."
5. Build a token-cost-per-deliverable dashboard. This is the single highest-leverage internal tool an AI-leveraged agency can build in 2026. I built ours in two evenings with Claude Code and a Supabase table — one row per AI workflow run, capturing input tokens, output tokens, cache hits, model used, deliverable type, and client ID. This dashboard is the only way to spot a tokenizer reprice the week it happens, instead of the quarter it shows up on the P&L.
What I'd ignore
Three things the news cycle is going to push on you that don't matter for operators.
Benchmark debates. Whether Opus 4.7 beats GPT-5.5 on SWE-bench by 4 points is irrelevant to anyone running an Amazon listing workflow. Both models are good enough at the work agencies actually do. The differentiator is your prompt library, your data, and your routing logic — not the model.
The "AI deflation" narrative. Last year's story was that AI prices were collapsing 10x per year. That story is over. Inference cost at the frontier is now flat-to-up after you account for tokenizer changes, mandatory thinking budgets, and the move away from extreme discounts. Plan for stable-to-slightly-rising costs, not collapse. The agencies that built their pricing model assuming costs would halve every six months are going to learn this the hard way.
Whatever your AI tooling vendor's blog says about productivity gains. Productivity is real. Margin is realer. A workflow that ships 5x faster but eats 40% of your gross profit is not a win — it's a treadmill that ends at zero. Measure productivity in dollars-of-margin-per-deliverable, not minutes-of-engineer-time-saved.
The contrarian agency take
Here's what I think nobody in the agency space is willing to say out loud yet.
The "pure AI agency" pitch is dead by Q4 2026. Any agency whose differentiation was "we do this with AI for 40% less" has just watched their margin structure quietly collapse. The brands that signed those deals will keep paying their contracted rate, but the next 100 deals close at a price point that no longer leaves room for the operating model.
The agency that wins the next 18 months is hybrid. Senior human strategy. AI-leveraged execution. Pricing that bakes in compute as a cost-of-revenue line, not a fixed bucket. Margin structure that absorbs a 30% input cost movement without anyone above the COO noticing.
If you've been pitching "we replaced 12 people with Claude," you're now competing with an agency that replaced 12 people with Claude six months ago and figured out how to route 80% of those calls to Haiku. That agency wins.
The Opus 4.7 tokenizer is a signal, not a story. It's the moment AI services pricing moved from a marketing question to an operations question. The agencies that treat it as a marketing question are the ones I expect to disappear by end of year.
If you're rebuilding your AI stack to survive the next vendor reprice, I'd love to hear what you're doing differently. Cost-per-deliverable is the only metric that matters in this market, and almost nobody's measuring it.