What did BenchLM's Token Price Index find for the twelve months to August 2026?

Tracking 143 models with live API pricing, BenchLM found frontier-tier pricing, its top 10 ranked models, up 36.4% year over year, while mid-tier and budget pricing across the remaining 133 models fell 35.8% over the identical window. The two trends moved in opposite directions on the same twelve-month clock rather than the market moving uniformly in one direction.

Why are frontier model prices rising while the rest of the market gets cheaper?

The deflation that ran from 2023 through 2025 came from efficiency gains and competition forcing cuts on models whose capability was not changing much between generations. The frontier tier's 2026 price increases track a different cause: each new frontier generation delivers a real capability jump, longer agentic task horizons and deeper multi-step reasoning, and providers are pricing that jump as a premium rather than absorbing it the way they absorbed efficiency gains on fixed-capability models.

How big is the frontier premium in dollar terms right now?

OpenAI's GPT-5.6 Sol prices at $5.00 per million input tokens and $30.00 per million output as of July 29, 2026, roughly double its own tier's January 2026 pricing. Anthropic's Claude Fable 5 runs $10.00 and $50.00. One tier down, Claude Opus 5 runs $5.00 and $25.00, Gemini 3.1 Pro runs $2.00 and $12.00, and GPT-5.6 Luna runs $1.00 and $6.00, all cheaper year over year rather than more expensive.

Does this mean the 'AI got cheaper' story from the last few years was wrong?

No, it describes a different slice of the market than the frontier tier. The cheapest tracked models have fallen roughly 214x from GPT-4's March 2023 launch price of $30 per million input tokens over the following 40 months, and open-weight models carry a median blended price around $0.53 per million tokens against roughly $2.81 to $3.00 for proprietary models. That deflation is real and it's most of the 143 models BenchLM tracks. It just doesn't apply to the 10 models at the top of the index, which moved the other way starting in January 2026.

What's the actual risk for a team that defaults every request to a frontier model?

A workload that stays flat in volume still gets 36.4% more expensive over twelve months purely from the provider's own repricing, with no change in traffic or usage pattern. The inverse holds for a default that already routes simple requests to a mid or budget-tier model: the same flat workload gets roughly 35.8% cheaper over the same period, for the same reason, in the other direction.

How does Nadir handle a market where the frontier and budget tiers are moving in opposite directions?

Nadir's classifier scores each prompt against live provider pricing on both sides of the split rather than a price ratio captured during a past cost review, and routes simple requests to whichever model on the falling side of the market can handle them while escalating only what genuinely needs frontier-level reasoning. The quality floor set per API key decides where that line sits. When a priced benchmark comparison is available, the response includes the nadir_metadata.benchmark_comparison.savings_usd field, so a repricing on either side shows up in your own metrics before it shows up on the invoice.

The Two-Speed Market

BenchLM's Token Price Index found the market splitting in two: frontier-tier pricing rose 36.4% year over year as mid-tier and budget pricing fell 35.8%. It tracked 143 models with live API pricing over the twelve months to August 2026. OpenAI's GPT-5.6 Sol, the current frontier flagship, prices at $5.00 per million input tokens and $30.00 per million output as of July 29, 2026, roughly double what the tier charged in January. One rung down, Google's Gemini 3.1 Pro runs $2.00 and $12.00, and DeepSeek's V4-Flash runs $0.14 and $0.28, both cheaper than a year ago. Same market, same twelve months, a 72-point spread between the two curves. A routing setup that defaults to "the frontier model, because it's safest" is compounding a 36% premium every quarter that default goes unreviewed, on the same traffic, with zero growth. Here's the split, why it's happening now, and what it means for a stack that still defaults up instead of routing.

The market split into two prices, and only one of them is falling.

BenchLM's Token Price Index tracked 143 models with live API pricing over the twelve months to August 2026 and found two trend lines moving in opposite directions on the same chart. Frontier-tier pricing, BenchLM's top 10 ranked models, rose 36.4% year over year. Mid-tier and budget pricing, the other 133 models, fell 35.8% over the identical window. Same market, same twelve months, a nearly 72-point spread between the two curves.

For three years, the standard line on LLM pricing was one number: prices keep falling, and routing to a cheaper model is free money with no expiration date. That is still true for most of the market. It stopped being true for the top of it in January 2026, and most routing setups built on the old, single-direction story have not been updated to reflect the split.

Why the top of the market got a repricing, not a discount

OpenAI's GPT-5.6 Sol, the frontier flagship as of July 29, 2026, prices at $5.00 per million input tokens and $30.00 per million output, roughly double what the equivalent frontier tier charged in January 2026. Anthropic's Claude Fable 5 runs $10.00 input and $50.00 output. Neither number reflects a shortage or a subsidy ending. It reflects a genuine capability jump, longer agentic task horizons, deeper multi-step reasoning, priced the way a new capability tier usually gets priced: at a premium, not a discount.

That is a different mechanism from the deflation that defined 2023 through 2025, which came from efficiency gains and competition forcing cuts on models whose capability was not moving much generation to generation. When the capability itself jumps, providers reprice upward instead of down. Both dynamics are running in the same market right now, on different rows of the same price list.

The three tiers, side by side

Tier	Model	Input ($/M)	Output ($/M)	12-month direction
Frontier	GPT-5.6 Sol	$5.00	$30.00	+100% since Jan 2026
Frontier	Claude Fable 5	$10.00	$50.00	up
Mid	Gemini 3.1 Pro	$2.00	$12.00	down
Mid	Claude Opus 5	$5.00	$25.00	down
Mid	GPT-5.6 Luna	$1.00	$6.00	down
Budget	DeepSeek V4-Flash	$0.14	$0.28	down
Budget	GPT-4.1 Nano	$0.10	$0.40	down

Frontier pricing up 36.4% year over year while mid and budget tiers fall 35.8%, BenchLM Token Price Index, August 2026

The gap between the cheapest and most expensive model BenchLM tracks now runs to roughly 4,773x on a blended per-million-token basis, as of August 6, 2026, the widest that spread has been since BenchLM started publishing the index. A market that spreads out this far is not a market where one number, "AI is getting cheaper," tells you anything useful about where your own traffic sits.

What "prices fell in aggregate" was hiding

The 40-month, 214x price collapse on the cheapest tracked models since GPT-4's March 2023 launch at $30 per million input tokens is real, and it is the reason the overall narrative has been "costs keep dropping" for three years running. But an aggregate built from 143 models, most of them budget and mid-tier, absorbs a 36.4% frontier increase without the average line even flinching. The average is not wrong. It is just describing a different 133 models than the ones a team defaulting to the frontier tier is actually paying for.

Open-weight models carry the deflation the furthest: a median blended price around $0.53 per million tokens against roughly $2.81 to $3.00 for proprietary models, an 81% to 82% gap that has only widened as budget-tier proprietary pricing chases it down. That's the pull driving the mid and budget curve toward zero. It has no bearing on what happens at the top of the market, where a harder task and a longer context window are now underwritten by a price increase instead of absorbed by one.

The default that quietly got 36% more expensive

Here's the part that matters for a routing decision made months ago and never revisited. A team that standardized on a frontier model as the safe, one-size-fits-all default, "just use the best model for everything so we don't have to think about it," is not looking at flat pricing while they wait to decide whether routing is worth building. They are on a curve that added 36.4% to the same workload, at the same volume, with zero growth in traffic, purely from the provider's own repricing over the last twelve months. That's not a cost of doing more. It's the cost of standing still.

The inverse holds for a setup that already routes simple, well-defined requests to a mid or budget-tier model by default and escalates only when a task needs it. That default rides the 35.8% deflation curve down, on the same twelve-month clock, again with zero change in behavior. Two teams running an identical workload, one default apart, are now 72 points of relative cost away from each other, and the gap is compounding every quarter both defaults stay unreviewed.

Three questions for your own stack

Do you know, right now, whether your default model sits on the tier that went up 36% or the one that went down 36%? Most teams can name their default model. Few have checked which side of this split it landed on in the last twelve months specifically.

Is your routing threshold pinned to a price ratio you measured before this divergence started? A cascade tuned against last year's frontier-to-budget gap is tuned against a gap that has since widened by roughly 72 percentage points. The threshold that made sense in August 2025 is escalating too eagerly, or not eagerly enough, today.

Would a repricing on either side show up in your dashboard before it shows up on the invoice? A market moving in two directions at once needs a routing layer that reads both, not a table someone wrote down when the market still moved in one.

Where Nadir fits

Nadir's classifier scores each prompt against live provider pricing on both sides of this split, a frontier tier that's getting more expensive and a mid and budget tier that's getting cheaper, rather than a price ratio someone captured during a cost review before the divergence widened. Simple requests route down to whichever model on the falling side of the market can handle them; only the requests that need frontier-level reasoning escalate to the rising side, and the quality floor you set per API key decides where that line sits, not a hardcoded model name.

The integration is two lines: point the base URL at Nadir, set model="auto". When a priced benchmark comparison is available, the response includes the nadir_metadata.benchmark_comparison.savings_usd field, so when either tier reprices, frontier up or budget down, you see the shift in your own metrics instead of finding out from the bill.

Conclusion

The story was never just "AI got cheaper." As of August 2026, it's "AI got cheaper for most of the market and meaningfully more expensive for the top of it, at the same time." A static default, in either direction, now rides one side of that split by accident. A routing layer that checks live pricing on every request rides whichever side actually fits the task, on purpose, and that difference compounds every quarter the split keeps widening.