The Frontier Premium Is Collapsing. Here Is What to Build When a $0.10 Open Model Matches a $10 Closed One.

Look at the price sheet for August 2026. DeepSeek V4 Flash serves input tokens at $0.14 a million. DeepSeek V4 Pro sits at $0.435 in and $0.87 out after a 75 percent cut. Kimi K3, a 2.8 trillion parameter model that tops the coding leaderboards, is a free download. GLM 5.2 prices at a fraction of anything a US closed lab charges. Meta's Muse Glimmer, a 30B model distilled from the most powerful thing Meta has ever built, runs free on a laptop under clean Apache 2.0. On the other side of the ledger, GPT-5.5-pro charges $180 a million output tokens and Claude Sonnet 5 sits at $2 a million input. And for a growing share of real workloads, the quality gap between the $0.10-class open model and the $10-class closed one is now within noise on task-level evals. That sentence is the whole opportunity. The rest of this post is what to do about it.

What the frontier premium actually is

The frontier premium is the price multiple closed labs charge over open equivalents for capability you can measure. Two years ago the premium bought something real. Closed models were 12 to 18 months ahead, and if your product needed that capability, you paid, because there was nothing else to buy. Today the premium buys weeks of lead time on general benchmarks. On your specific task, after light fine-tuning, it often buys nothing at all. Bridgewater proved the point in public: a tuned open model beat frontier models on financial reasoning at a fraction of the cost. Not matched. Beat. The premium is still priced like it buys a capability gap. Increasingly it buys a logo.

Why this is structural, not a sale

It would be easy to read the current prices as a promotional war that ends when someone blinks. The mechanics say otherwise. Nine open releases landed in twelve days this month. Distillation now compresses frontier capability into small free packages within weeks of that capability existing. Chinese labs are pricing tokens at 18 cents against a $4 US closed average, and they are profitable enough at that price to keep shipping. Meta is commoditizing the complement because free intelligence serves its ads business. And the closed labs are already cutting, with Opus 5 launching at half its predecessor's price, which tells you they see the same curve. Each mechanism feeds the others: cheaper open models pull demand, demand funds more open releases, distillation resets the floor, closed labs cut to compete, and the premium ratchets down another notch. Venture money has noticed. VCs are retreating from vague AI pitches toward what they call expensive, defensible problems. That is the market pricing in the collapse before most founders have.

The question that replaces 'which model is best'

Here is the mental shift that separates founders who profit from this from founders who just watch it. Stop asking which model is best. Start asking what the cheapest capability is that clears your quality bar. Most founders never define the bar. They default to the premium model because it feels safe, and then they burn margin on quality their customers literally cannot perceive. If you cannot state the accuracy, latency, and failure tolerance your product actually requires, you are not making a model decision, you are making a brand decision with your gross margin. The eval harness that encodes your bar is the highest-ROI artifact in your company right now, because it is the only thing that can tell you the day the $0.10 model clears it.

What to build now

Start with products that have fat inference loops. Monitoring everything, retrying everything, pre-computing everything a user might ask for. At frontier prices these architectures were economically insane, because every speculative call was a dime you probably wasted. At open prices the same loops are gross-margin positive, and the products they enable feel like magic compared to call-once competitors. Next, high-volume, low-stakes automation: triage, enrichment, extraction, classification at scales where the premium model's cost was the entire blocker. There are categories of boring, valuable work that only pencil out below a cent per call, and they just penciled out. Then local-first and privacy-first products. Glimmer-class models on device mean zero marginal inference cost and no data leaving the machine, which unlocks whole categories that compliance kept closed: healthcare notes, legal review, personal agents that see everything you do. Finally, vertical fine-tunes. Your domain data on a cheap open base beats the general frontier on your task, Bridgewater style, and the collapsed base price makes the fine-tune economics trivial. The moat is your data and your eval, not the model.

What not to build

Two traps. First, anything whose pitch is arbitraging the premium itself, reselling cheaper inference as the product. The premium keeps collapsing under you, and the labs' own price cuts eat your wedge on their schedule, not yours. Second, anything that needs the frontier delta to justify its price. That delta is measured in weeks now. If your product only makes sense while the closed model is meaningfully smarter, you are building on land the tide has already reached.

Run the margin math this week

Take your current inference COGS. Remodel it at open-model prices for the 80 percent of calls that do not need the frontier, and keep the premium model only for the 20 percent that measurably do. That is a router architecture, and it is a weekend of work if you already have an eval harness. Most AI products that run this exercise discover their blended COGS drops 60 to 90 percent. That number is either margin expansion or price-cut ammunition against competitors still paying full freight. Either way it is yours to spend.

The honest caveats

Some workloads genuinely still need the frontier. The hardest reasoning, the longest horizons, the tasks where a wrong answer is expensive. The premium is collapsing, not collapsed, and pretending otherwise is how you ship a worse product. Open models also arrive with sharp edges. The K3 hallucination-rate lesson was a reminder that leaderboard rank and production reliability are different measurements, which means your eval bar matters more in this world, not less. And read the licenses. K3 gates revenue above a threshold. Glimmer is clean Apache 2.0. Those are different dependencies, and only one of them belongs in a product you plan to sell.

The frontier premium was a tax on not measuring. The founders who build eval harnesses, define their quality bar, and route every call to the cheapest model that clears it get to keep the tax as margin. Everyone else keeps paying it out of habit.