The Ten Cent Token: Inside September 2026's LLM Pricing Collapse

Between September 12 and 25, 2026, industry trackers counted more than 20 distinct model releases, with September 23 standing out as the single busiest release day of the month at nine tracked pricing updates. Pull those releases into one leaderboard and the spread from top to bottom is 119 fold: Claude Fable 5.1 priced at $11.90 per million tokens blended against Meta's Muse Spark 1.3 landing at roughly $0.10 per million, a price point the tracker described plainly as a new floor. GPT-6 Luna matched that floor within days at $0.10 input and $0.50 output per million tokens. A gap that large between models nominally competing in the same market is not a rounding difference in margin, it is a sign that "LLM API pricing" stopped describing one market some time ago and started describing several, and the practical question for any engineering team is which of those markets your actual workload sits in.

The leaderboard, as it stood by month end

Why the spread is the story, not any single price point

A single cheap model launching is not news, cheap small models have existed for years. What changed this cycle is the simultaneous presence of a genuine frontier tier still pricing above ten dollars per million blended tokens and a usable, branded model from a major lab sitting at one tenth of one cent per thousand tokens, with both ends of that range shipping and updating in the same two week window. That simultaneity is what makes "what should we budget for LLM API costs" an unanswerable question without first answering "what is the task actually asking the model to do." A workload that needs Claude Fable 5.1 level reasoning does not get materially cheaper by switching vendors within that tier, the frontier tier is priced the way it is because the capability gap above the mid tier is real. A workload that a $0.10 per million token model handles adequately was probably never a good candidate for frontier pricing in the first place, and the 119x gap is the market finally pricing that difference explicitly instead of burying it inside one blended sticker price.

Time of day pricing is the detail worth copying

DeepSeek V4.1 Flash's peak and off peak pricing structure, $0.30 and $1.20 during peak hours against $0.15 and $0.60 off peak, is a structure almost no Western provider offers yet, and it is directly actionable for any team running batch or non latency sensitive workloads. If your pipeline does not need a response inside a user facing request window, scheduling that work into off peak hours against a provider offering time of day pricing is free cost reduction that requires zero model or infrastructure change, only a scheduling change. Teams evaluating self hosted inference specifically to escape API pricing volatility should check whether a simpler scheduling shift against an existing API relationship gets most of the same benefit before committing to the operational overhead of running inference themselves.

What this means for the build versus buy decision

The honest framing is that this spread makes the decision easier, not harder, once you separate it by workload tier rather than trying to pick one answer for your whole stack. Frontier reasoning workloads stay on API access to frontier models almost by definition, self hosting a model at Claude Fable 5.1's capability tier is not a realistic substitution for most teams regardless of the price gap, because the gap exists precisely because that capability is hard to replicate. High volume, latency tolerant, moderate complexity workloads are where the math genuinely shifts: a $0.10 per million token API price is low enough that self hosting an equivalent open weight model rarely wins on pure unit economics anymore unless your volume is large enough that even that low price adds up to real spend, or unless you have specific data residency, latency, or customization requirements the API path cannot satisfy regardless of price.

When this applies to your stack

If your team is still budgeting LLM costs against a single blended assumption, one number that is supposed to represent "what we pay per token" across every workload, this pricing cycle is the moment to stop. A 119 fold spread between the frontier tier and the new floor means a single number cannot represent your actual spend profile without hiding which workloads are overpaying and which are underserved by a model too cheap for the task.

Contra Collective builds and cost models production LLM infrastructure for enterprise clients, including the workload tiering and vendor selection work that turns a pricing spread like this one into an actual budget rather than a single misleading average. If your team cannot currently break out API spend by workload tier rather than by vendor alone, that breakdown is worth building before your next budget cycle, not after.

FAQ

What caused the $0.10 per million token price floor specifically? Meta's Muse Spark 1.3 hit that price point this cycle, and GPT-6 Luna matched it within days at $0.10 input and $0.50 output per million tokens, establishing it as a floor other providers are now pricing against rather than an isolated promotional rate.

Does a lower price always mean a materially weaker model? Not necessarily for every task, but the 119 fold spread exists because the frontier tier and the floor tier are solving genuinely different problems. The right comparison is always task fit, not price alone, since a model at either end of the spread can be the wrong choice for a workload it was not built for.

Is DeepSeek V4.1 Flash's peak and off peak pricing common? No, it is unusual among major providers as of this pricing cycle. Most Western vendors still price on a flat rate regardless of time of day, which makes DeepSeek's structure a specific opportunity for teams with schedulable, non latency sensitive workloads.

Should this pricing spread change our self hosting plans? It depends entirely on which workload tier you are evaluating. At the floor end of the spread, API pricing is now low enough that self hosting rarely wins on unit economics alone; at the frontier end, self hosting an equivalent model was never realistic regardless of price. The spread mainly clarifies which tier a given workload actually belongs to.