The week the labs stopped selling intelligence
Meta Superintelligence Labs finally shipped. And its answer wasn't a smarter model — it was a cheaper one.
That's not just Meta's story. It was the whole market's week — Musk and Altman shipped into the same seven days, and each one led with the invoice, not the benchmark. Read the week's launches together and they describe one reversal: the frontier labs stopped competing on how smart their models are, and started competing on what a finished task costs.
Why did all three pivot at once? The answer lives on the demand side. Bloomberg's read on the week was blunt: three prominent labs released models whose biggest selling point "may not be what they can do but how little they charge to do it." The "tokenmaxxing" era — burning tokens to look productive — is reversing hard now that finance teams have read the bill: Uber capped employee AI tooling at $1,500/month after burning its annual AI budget in roughly four months, and Lindy's CEO moved 100% of traffic off Claude to DeepSeek, calling it "a matter of survival for the business." Supply didn't decide to compete on price — a cost-conscious buyer forced it to. Enterprises pushed back on spend first; the labs repriced second.
3
labs launched on cost, not capability, in one 7-day window
$14.3B
Meta's talent bet a year before this week's answer
~4x
fewer tokens per coding task — Grok 4.5 vs. Opus 4.8
54%
more token-efficient — OpenAI's own headline stat for GPT-5.6
What a million output tokens costs, this week
per million output tokens
Meta Superintelligence Labs finally answered the question
Jun 2025
$14.3B for a 49% stake in Scale AI to install Alexandr Wang as Chief AI Officer; 50+ researchers poached from OpenAI, Google and Apple.
Aug 2025
Reorg into four teams; a reported hiring freeze follows within a day.
Oct 2025
~600 AI layoffs, four months after the hiring spree.
Nov 2025
Yann LeCun resigns over the closed, product-first direction.
Apr 2026
Muse Spark 1.0 ships into a closed partner program — no public API.
Jul 2026
Muse Spark 1.1 + the Meta Model API: the first time in company history Meta has charged for a model.
A year of near-silence, and one question: what did all that money buy? This week, MSL answered — $1.25/$4.25 per million tokens, roughly a quarter of flagship rates, with an API that speaks both OpenAI's and Anthropic's SDK formats natively, so switching is a base-URL change. On independent coding evals it still trails Opus 4.8 and GPT-5.5. They didn't build a moat, they built an off-ramp from everyone else's pricing.
Musk's launch quote wasn't "smarter"
“It is an Opus-class model, but faster, more token-efficient and lower cost.”
Grok 4.5 shipped at $2/$6 per million tokens, trained on Cursor developer-session data. The most self-promotional founder in tech had the chance to say "smartest model in the world" and instead said cheaper per task — roughly 4x fewer output tokens than Opus 4.8 on the same coding benchmark, about $0.49 per completed task by independent benchmarking. "Opus-class" concedes the capability race is close enough that parity is the claim; efficiency is the differentiator. One caveat: Cursor disclosed a codebase snapshot accidentally ended up in training data, so treat Cursor-specific benchmarks with some suspicion — the pricing, though, is real.
OpenAI led its biggest launch with a cost number
“Every enterprise now is thinking about spend and the value they're getting in exchange for AI, and this is what we really want to do.”
GPT-5.6 went generally available this week. The headline stat Altman gave CNBC wasn't a benchmark score — it was "54% more token efficient on agentic coding." That's not a founder talking to developers, that's a founder talking directly to the CFO — because the CFO just joined the model-selection meeting. The whole family (Sol, Terra, Luna) is priced like a company that knows what question its buyers are asking.
One reversal, three launches. Twelve months ago, every model launch bragged about being smarter. This week, none of them did. Capability is commoditizing — the frontier models sit within a few benchmark points and a few dollars of each other — so competition migrated to the invoice. MSL's year-in-the-making answer, Musk's launch quote, and Altman's cost stat are the same fact seen from strategy, marketing, and pricing, with enterprise buyers holding the pen.
For anyone building AI products, the practical consequence is sharp: the scarce skill is no longer picking the smartest model. It's knowing what a finished task should cost — and routing to hit that number. Model choice used to be a technical decision made once. It's now an economic decision made continuously, and the teams that treat it that way will quietly run 3–4x cheaper than the teams that don't.
Built something like this? I'm always happy to compare notes.
aniket.kgp25@gmail.com →