AI Tokens Just Got 60% Cheaper — And That Should Terrify The $600 Billion Capex Party

Honestly, the most interesting AI story this week isn't a flashy model launch. It's a number. The price of intelligence is collapsing faster than anyone in the C-suite wants to admit.

Silicon Data's LLM Token Expenditure Index fell to $0.9665 per million tokens on August 31 — down more than half from its May peak of $2.0651. Seeking Alpha's Damir Tokic puts the full decline at nearly 60% if you count from the earlier peak around $2.20. Either way, that's not a blip. That's a cliff.

And it lands right in the middle of the biggest infrastructure buildout in tech history. Alphabet, Amazon, Meta and Microsoft are on track to pour roughly $600 billion into AI in 2026. GPUs, data centers, networking, energy. All of it premised on demand staying high.

Here's the uncomfortable nuance most hot takes skip: a falling token price doesn't mean people are consuming less AI. It means intelligence is getting cheaper to consume. The index is usage-weighted — it tracks the effective price companies actually pay, not total token volume. So what we're really watching is commoditization in real time.

Why is it collapsing? Competition, plain and simple. Businesses are routing routine work — classification, summarization, extraction — to smaller models that cost a fraction of a frontier API. Open-weight models got good enough that the expensive stuff is now reserved for jobs that truly need it. Reuters reported back in June that the share of open-model tokens on one major routing platform jumped from 34% in January to 65% in June. DeepSeek and the Chinese labs piled on the pricing pressure. And model routers like OpenRouter — which Stripe just bought for over $7 billion — make it trivial to shop each request to the cheapest model that can do the job.

I've been saying for a while that the token market was heading toward this. The surprise isn't that it happened. It's that it happened this fast, while capex plans were still being signed at bubble prices.

To be fair, this could go two ways. Either the price compression chokes hyperscaler margins and forces a stall or cut in AI capex — which is what the bearish camp argues — or it's the moment AI actually becomes a real market, with real price discovery, and the $600 billion bet turns into computing's biggest economic payoff. I genuinely don't know which one wins. But the fact that we're even having this conversation in 2026 means the free-lunch era of AI pricing is over.

Meanwhile, GPT-6 Astra is a security nightmare wrapped in a math genius

OpenAI started rolling out GPT-6 Astra this week, and the benchmark sheet is genuinely absurd. It aced three of the toughest tests in the field. FrontierMath Tier 4 — the set of 50 problems that take human mathematicians weeks each — at 98%. ARC-AGI-3, which measures learning new tasks, at 99.9%. And a perfect score on ExploitBench, which measures how well a model can find and exploit software vulnerabilities.

That last one is the kicker. OpenAI itself classified Astra as "critical" risk under its internal safety framework — the designation reserved for models that can hack many well-protected systems without human input. They delayed the release for weeks just to build guardrails against it.

There's also a genuinely clever bit of engineering underneath. Most LLMs use a "compaction" mechanism that compresses context data when it overflows and just throws away the low-priority stuff. Astra instead archives the original information in a searchable form, so nothing gets discarded. That's the kind of design choice that explains why it's scoring like this.

At launch it's only available through a program called Daybreak for cybersecurity research, with OpenAI handing out $1 billion in credits to government agencies, utilities and essential service operators — training included, some of it with MS-ISAC. ChatGPT, Codex and API access roll out in the coming days.

Look, a model this capable is going to be a double-edged sword no matter how many guardrails you bolt on. I'm not calling it a disaster — the math advances alone (it solved multiple Erdős problems pre-release) are remarkable. But "critical risk" isn't a marketing label, and the fact that OpenAI says it out loud tells you how carefully you should treat anything this model claims to be able to do autonomously.

Hugging Face ships Transformers 2.0 — open source keeps pulling away

On the open side, Hugging Face dropped Transformers 2.0 this week after a year-long sprint with over 150 contributors. Two hundred-plus new pretrained models, 30 additional languages (total coverage now at 120, including Swahili, Burmese and Yoruba), and up to a 40% speedup on common inference workloads.

The part I actually care about is the unified quantization API — 8-bit inference with less than 2% accuracy loss. That's the kind of thing that cuts cloud compute bills by up to 30% for startups, which is exactly what makes open models dangerous to the premium-priced incumbents. Meta's Llama 3 apparently already cites Transformers 2.0 as a key reference implementation. When your "competitor" is literally building on your open stack, the moat conversation gets awkward fast.

The local LLM crowd is done waiting

Two stories this week from the tinkerer corner: someone got a Raspberry Pi to boot straight into a local LLM instead of Linux, and another dev ditched cloud AI for a local LLM running in Proxmox. Both are hobby-grade, sure. But together they're a signal. Every time token prices collapse or a frontier model turns out to be a security liability, a few more people decide their data doesn't need to leave the house. That's not a trend that's going to reverse, even if the economics don't always pencil out.

Quick add-on note: UGREEN also shipped a local AI NAS (the HomeAgent HA100) that Android Central is calling the most well-thought-out smart home product they've used — on-device AI for camera monitoring without a cloud subscription. It's early days, but the "AI runs at home" product category is finally getting real hardware instead of vaporware.

From my perspective, the through-line this week is uncomfortable for the industry: intelligence is getting cheaper, the most powerful model ever built is also the riskiest, and the open ecosystem keeps making premium pricing harder to justify. The $600 billion question isn't whether AI works anymore. It's whether anyone's capex plan accounted for the price of intelligence doing a 60% dive in a single quarter.

If you're building on top of all this, the smart play is boring: keep your stack portable, don't marry one provider, and assume the price you're paying for tokens today is not the price you'll pay in six months. It'll be lower. Much lower.

---

Found this useful? I keep an eye on similar cost-reduction and local-AI angles — check out Engineering Reference.