The Day Intelligence Became Cheap

Two flagship models shipped on the same Tuesday. Hours apart. Anthropic dropped Claude Opus 5.5 at $4 in / $20 out per million tokens. OpenAI answered with GPT-6 Sol and GPT-6 Luna at half their predecessors’ prices. Both companies called for a slowdown a week ago.

Welcome to the commoditization of intelligence.

What Just Happened

This wasn’t a normal release day. Look at the moves:

Anthropic’s Opus 5.5 hits near-flagship performance at a 20–40% price cut.
OpenAI’s GPT-6 Sol and Luna are API-priced 50% lower than GPT-5.6.
Grok 4.7 landed the day before at $2 in / $6 out per million tokens.
StepFun’s Step 5 Preview opened at $1 in / $2.70 out — the price of a sandwich per million tokens.
Shanghai AI Lab released a 744-billion-parameter Mixture-of-Experts under MIT license, free, no blog post.

Frontier capability used to cost real money. The bill for “production-grade reasoning” was a budget line item. As of this week, it’s loose change. And the cheap stuff isn’t toy-tier anymore. Opus 5.5 matches Anthropic’s flagship Claude Fable on most tasks. Step 5 Preview is competitive with Kimi K3 Max at roughly a seventh the price of GPT-5.6 Sol.

This isn’t competition. This is the moment a market tipped over.

The Slowdown Lie

A week ago, the heads of both OpenAI and Anthropic publicly asked for a slowdown in frontier development. They’ve spent the year saying “we need guardrails, we need oversight, we need breathing room.” Then they shipped faster and cheaper than ever.

That’s not contradiction. That’s the market talking. When open-weight Chinese labs ship 744B models for free, when StepFun ships 600B at a dollar per million tokens, when Meta’s Muse tops app store charts — the premium frontier labs have two choices: keep charging premium prices for premium intelligence, or democratize the price and keep the margin on volume.

They picked volume. They picked democratization. And they did it while publicly asking regulators to slow them down.

Read that again.

Why This Matters More Than The Headlines Make It Sound

The Sol AI skill marketplace launched into a world where running an LLM meant a $50K annual API contract and a procurement team. I remember those days. The thesis was always: as soon as intelligence becomes cheap, the bottleneck moves to judgment and orchestration.

That day is here.

A million tokens of high-quality output for $4 changes what you can build. A free 744B model on Hugging Face changes who can build it. The asymmetry between “can afford it” and “cannot afford it” is closing faster than anyone — including the labs themselves — planned for.

But here’s the part the marketing pages won’t tell you:

Cheap inference doesn’t equal safe inference. Cisco Talos disclosed CLOSEDQUORUM on the same week — what they’re calling the first fully autonomous multi-model AI command-and-control implant. A malware framework that orchestrates multiple frontier models to run its own attack chain. Attack cost just fell as fast as defense cost.

The same price compression that puts a capable agent in every solo developer’s terminal puts one in every botnet operator’s terminal too.

The Real Story For Builders

If you’re building on top of these models — like I do, every day — the operative questions just changed:

Multi-model orchestration stops being optional. When the price floor is a dollar per million tokens, the cost of running three models in parallel and arbitrating between them drops below the cost of running one model twice. The era of “pick one and live with it” is over.
Local starts to make sense. With PrismML’s Bonsai 2 squeezing Qwen3.8 27B into 5.93 GB on ternary weights, “frontier-grade local” is feasible on a Mac mini. Compute is commoditizing at every layer, not just API.
The differentiation moves up the stack. When the model is a commodity, your edge is the data layer, the orchestration logic, the UX, the trust. Not the prompt.
Safety is a feature, not a tax. With autonomous AI C2 implants being a real research artifact, “we used a frontier lab” stops being enough. You need provenance, eval, and observability. That’s a product surface, not a compliance checkbox.

The Bit That Actually Scares Me

I’m not scared of $4 per million tokens. I’m scared of $4 per million tokens in the hands of a 14-year-old script kiddie running an autonomous agent loop.

The same price compression that makes me able to run a 4-model council at pittance makes a credential-stuffing agent trivial to operate. Google disclosed this week that Gemini autonomously accessed the systems of three companies during a security test, finding publicly available credentials and guessing passwords. That was Gemini trying to be helpful. Imagine Gemini trying to be helpful for an attacker.

The labs say the answer is more capability. I say the answer is more friction. The friction lives in the application layer — auth, rate limits, audit trails, capability scoping. Whoever builds that layer well owns the next decade.

My Take

Frontier intelligence just became a commodity. The pricing war wasn’t optional — the open-weight squeeze from Shanghai and the cost collapse from Chinese labs forced it. The American labs can either lead the price compression or be eaten by it.

They chose to lead. Good for them. Terrifying for everyone else.

The developers and founders who win the next 18 months will be the ones who treat cheap intelligence as table stakes and build the trust, orchestration, and product layer that frontier labs won’t. That’s where the margin goes when the tokens go to a dollar.

The Sol AI skill marketplace lives here. So does anyone serious about shipping agents in 2026.

The day intelligence became cheap is the day the real work started.

Sol Alexander — September 23, 2026