"I have zero interest in buying market share through price cuts. I'd rather focus on customers and how they extract maximum value from the models."
That's Paul Smith, Anthropic's Chief Business Officer, in an interview that reportedly landed days before Anthropic shipped Fable 5.1 and cut its cache-read price from $1.00 to $0.25 per million tokens — a 75% reduction that Anthropic itself estimates at 25% savings on typical workloads and up to 45% on context-heavy agentic tasks.
So here's the question the headline won't ask: is a cut you refuse to call a cut still a cut?
Base rates held. Cache reads fell 75%. The premium label and the discounted invoice are now two separate documents — believe the invoice, not the interview.
Last week, the open thread was whether report-only labs and providers would actually cut per-token prices, or just talk about capability. This week it's confirmed — three frontier labs compressed prices inside a 72-hour window, and the industry benchmark broke a threshold it had never broken before.
On September 1, Silicon Data's LLM Token Expenditure Index reportedly hit $0.97 per million tokens — the first sub-$1 reading since the index launched, down 8.6% over seven days and less than half its summer peak. That's the floor. Now watch how Anthropic stood on it while insisting it wasn't moving.
The trick is a dual-track ledger. Smith's "no price war" claim is technically true on one set of numbers, and the Fable 5.1 cut is technically true on another — and they never touch.
Base input and output for Fable 5.1 held at $10 and $50. That's the number underwriters see. It preserves the premium sticker, the "we don't discount" posture, the margin story a pre-IPO Anthropic needs while it lines up a reported $15 billion credit facility led by Morgan Stanley.
The cache-read cut is the number you actually pay. It targets the repeat-read pattern — the same context loaded again and again across a long-running agent loop — which, per E1, is exactly where agentic spend concentrates.
So the "premium" claim and the "discount" are both live at once, because they describe different lines on the invoice. Smith can say "this is not about price cuts" with a straight face, because the line he's defending genuinely didn't move. The line that moved is the one he didn't mention.
That's not a contradiction. It's positioning. And it tells you precisely where Anthropic thinks its stickiest customers live: in long-horizon agent workflows where cache reads dominate, and where a 45% effective cut buys retention without touching the headline rate.
Which answers 08/30's open thread directly. The question was whether Anthropic would respond once its stickiest asset started commoditizing. The answer is yes — with a cache cut dressed as a value story.
Now widen the frame, because the reason Anthropic can't just hold the line is what's happening one tier down.
On September 2, Google shipped Gemini 3.8 Flash at $0.75 / $3.75 per million tokens — introductory pricing through the end of 2026, doubling to $1.50 / $7.50 in January. Even at the 2027 rate, that's reportedly less than half of Opus 5's $5 / $25.
The old defense of the premium was capability: pay more, get materially better. Google's benchmark table is quietly detonating that argument on specific workloads.
Per Google's published benchmarks, Gemini 3.8 Flash beats Opus 5 on the financial-analysis agent (61.4% vs 58.6%), on legal workflow (10.0% vs 6.7%), and on hard biological research (56.5% vs 49.4%). On finance, law, and biology, the Flash-tier model is no longer behind the flagship — at roughly one-seventh the price.
That is the exact profile of Anthropic's Cowork ambition — finance teams, legal departments, researchers. Smith himself said 98% of enterprise users aren't software engineers. Google just published benchmarks showing a workhorse model matching flagship tier on the workloads those 98% actually run.
But the premium isn't dead everywhere — and this is the part that explains Anthropic's whole play. On Terminal-bench 4.0, Gemini 3.8 Flash scores 19.1% against Opus 5's 51.8%. On computer use, 59.0% vs 75.4%.
So the capability moat hasn't collapsed. It's narrowed — to long-horizon autonomous coding and computer use, the exact domains where cache-heavy agent loops run longest and cost most. Which is precisely where Anthropic aimed its cache cut.
Read those two facts together and the strategy snaps into focus. Anthropic is conceding the general-capability premium is gone, retreating to the two workloads it still wins, and defending them not with a lower sticker but with a lower effective rate on the traffic pattern those workloads generate. Premium label, discounted invoice, same week.
The other half of the ledger is margin protection you'll never see priced. Anthropic gated Mythos 5.1 behind its Cyber and Life Sciences Verification Programs; Google gated 3.8 Flash Cyber behind its Fairwind Program; OpenAI's Astra reportedly keeps its most consequential capabilities off the public grid. Commoditize the volume tier, wall off the top, charge what the market bears where nobody's watching the rate card. Silicon Data reports roughly 95% of enterprise AI usage runs on frontier models; the split between a commodity tier and a gated one is not directly established — high-volume low-margin below, high-value restricted above.
For a company underwriting an IPO, that's the ideal shape: volume metrics from the commodity track, margin story from the gated track, and a "we don't do price wars" quote to narrate the whole thing.
Here's the procurement playbook, because none of this changes unless you change how you buy.
Don't buy the rhetoric. "No price war" is a positioning statement aimed at underwriters, not a description of your bill. The label "premium" and the invoice are now two separate documents — stop treating one as evidence about the other.
Don't buy the sticker either. Base input/output rates are the number labs want you to anchor on precisely because they're the number they're not cutting. The cut lives in cache reads, dynamic tiers, and promo windows.
Instead, audit where your tokens actually go. If you run long agent loops, cache reads may dominate your spend — in which case Fable 5.1's 75% cache cut is a real 25–45% reduction you should capture, regardless of what the headline base rate says.
Then re-underwrite against the effective rate, not the rack rate. Model Gemini 3.8 Flash at its 2027 price of $1.50 / $7.50, not the promo $0.75 — the discount expires and your budget shouldn't assume it won't. Map your workloads to the benchmarks that match them: if you're doing finance, legal, or biology, the flagship premium may be indefensible; if you're doing long-horizon coding or computer use, it may still be worth paying.
The blunt version: the index is at $0.97 and falling, the capability moat has narrowed to two workloads, and the one lab loudly refusing to compete on price just shipped one of the week's most aggressive effective cuts.
Believe the invoice, not the interview.