Model Leaderboard: The Top 10 Read as a Price-Performance Story, Not a Ranking

Divide $165.00 by $0.38 and you get 434.2 — the cost ratio between the third row and the second row of the OrcaRouter model leaderboard top ten, using the blended $/1M figure the board prints for each. The two rows are 1.9 composite points apart. That single division is the most useful thing on the page, because it turns a ranking into a purchase decision: what is 1.9 points worth to you, at 434 times the price?

Rankings are usually read down the left column and abandoned there. Read across instead — composite, blended price, latency, success rate — and the top ten stops being a ladder and becomes a set of distinct trade-offs. Two of the ten rungs are worth paying for at any volume. Two more are worth paying for at a specific volume. The rest are a story about what happens when a table sorts on quality and prices on something else.

The 434x, computed rather than asserted

The blended figure is the average of the input and output rates per million tokens, and it is the right number to divide because most workloads read and write. At that rate, the arithmetic is unforgiving:

The composite column is our own score, built from the evidence families the board publishes, and it is not a third-party benchmark — the board says so itself, and we repeat it every time the number appears in this article. The blended price comes from the provider’s published input and output rates, with $0 markup added by the gateway.

Run the right-hand column and the ranking collapses into three groups. There is a cheap tier at $30 to $38 per hundred million blended tokens: ranks 2, 9 and 10. There is a mid tier at $261 to $1,500: ranks 1, 4, 5 and 8. And there is a top tier at $3,000 to $16,500: ranks 3, 6 and 7. The composite column does not have those groups. It runs 83.0 down to 69.1 as a single smooth line.

What 1.9 composite points is worth

Here is the second piece of arithmetic, and it comes from the board’s own normalisation: 50 is the board average, every 15 points is one standard deviation. So the 1.9 points between rank 2 and rank 3 is 1.9 divided by 15 — about 0.13 standard deviations. Thirteen hundredths of a standard deviation is the entire measured advantage of the $165.00 row over the $0.38 row.

Cost per composite point makes the same point in a single column. At 100 million blended tokens a month, rank 2’s $38.00 buys 82.4 points — roughly 46 cents per point. Rank 3’s $16,500.00 buys 80.5 points — about $205.00 per point. That is a 445-fold difference in what a point costs, for two rows that sit next to each other on the board. Rank 1, at $1,200.00 and 83.0, costs about $14.46 per point — thirty-one times rank 2’s unit cost for 0.6 points above it.

The compressed scale at the top is not a defect; it is what a standardised composite looks like when the field is genuinely close. But anyone reading the top three as three tiers is reading the left column and ignoring the middle one. On this board, the first three rows are one tier of quality and three tiers of price.

The latency column narrows the gap further

Latency is measured as p50 in milliseconds, and the top ten spread across 1483 ms at rank 2 to 6395 ms at rank 8 — a 4.3-fold span. Rank 2 reaches its median response at 1483 ms; rank 3 at 4400 ms; rank 8 at 6395 ms. Rank 1 sits at 3868 ms, about 2.6 times rank 2’s figure.

So the 434x price gap is not buying speed. The row that is roughly 434 times more expensive per blended token is also about three times slower at p50 — the expensive row answers later, not sooner. The three fastest rungs in the top ten are rank 2 at 1483 ms, rank 9 at 1533 ms and rank 6 at 1762 ms, and two of those three are in the cheap tier on price. If your product is interactive, the ranking and the latency column are pulling in the same direction on those rows.

Reliability, the board’s seven-day success rate, is the one column where the top ten barely separate at all: 99.00% at rank 4 through 99.97% at rank 2, a spread under one percentage point across ten rows. A sub-one-point spread on a seven-day rate is not a way to choose between rows. It is a way to confirm that all ten stayed up.

The two rows where rank and price disagree

Two rows in the top ten carry an unusually high price for their position, and it is worth naming why rather than dismissing them.

Rank 4, GPT-6.1 Sol, sits at $6.00 blended ($2.00 in, $10.00 out) and 74.0 composite, at 3321 ms and a seven-day success rate of 99.00% — the lowest of the ten. It is 6.5 points below rank 3 at 3.6% of rank 3’s price. That is not a row you buy for its rank. It is a row you buy because it is the cheapest member of the mid tier that still clears 74 points.

Rank 6 and rank 7 share a price: $30.00 blended, from identical $10.00 input and $50.00 output rates. GPT-6 Astra scores 70.6 at 1762 ms; Claude Fable 5.1 scores 70.4 at 3890 ms. Same price, 0.2 points apart on our composite, and a 2.2-fold difference in p50 latency. When two rows are priced identically and ranked almost identically, the latency column is the only column left that separates them — and it separates them clearly.

The genuinely awkward pair is ranks 9 and 10. GPT-6 Luna at $0.30 and 69.5, GLM 5.3 Flash at $0.33 and 69.1. Three cents per million blended tokens and four tenths of a composite point, both inside the cheap tier alongside rank 2. Anyone who assumes the cheap end of a leaderboard is where the weak models live has the ordering upside down: two of the three cheapest rows on this board rank ninth and tenth out of twenty-five, not twenty-fourth and twenty-fifth.

Which rungs are worth paying for

Rank 2 is the one row on this board that needs no defence. DeepSeek V4.1 Flash — the model page is `deepseek/deepseek-v4.1-flash` — carries a composite of 82.4 at $0.38 blended, a p50 of 1483 ms and a seven-day success rate of 99.97%, the highest of the ten. It is 0.6 points behind rank 1 for 3.2% of rank 1’s per-token price, and the model is 2.6 times faster. If a workload can run on one row from this board, this is the row the arithmetic picks.

Rank 9 is the cheap-tier alternative, and its case is narrower than rank 2’s. GPT-6 Luna at $0.30 blended and 69.5 composite is eight cents per million cheaper than rank 2 and 12.9 composite points behind it. Under the board’s scale that gap is 0.86 standard deviations — the largest composite step between any two adjacent cheap rows, and cheap rows do not often give you a step that size. Paying $0.08 more per million to gain 12.9 points is the best unit trade in the top ten, which is exactly why rank 2 is the default rather than rank 9.

Ranks 1, 4, 5 and 8 are worth their price only against a specific requirement. Rank 1, Claude Opus 5.5 at $12.00 and 83.0, is the highest composite on the board at a p50 of 3868 ms; if you need the top of the composite scale and can absorb 31.6 times rank 2’s price, it is there, and it is the only row that buys the top of the scale at all. Rank 5, GLM 5.3 at $2.61 and 71.3, is the cheapest row above 70 points after rank 4, and rank 4 itself is the cheapest 74. Ranks 3, 6 and 7 cost $165.00, $30.00 and $30.00 per million blended tokens for composites of 80.5, 70.6 and 70.4 — one row that buys a high rank at an extreme price, and two that buy mid-tier scores at top-tier prices.

One honest note about rank 3, because the arithmetic here cuts the other way and the board does not hide it. GPT-5.4 Pro carries the highest score in the top ten on the independent intelligence index the board lists beside each row — 66.0, against 39.5 for rank 2 and 38.1 for rank 9. On that one third-party measure, the expensive row is 26.5 points ahead of the cheap one, and on our composite it is 1.9 points behind. That divergence is the blended method working as designed: a composite that weights human preference at 40% can rank a model highly for how it is judged rather than how it scores on a task suite. It also means the 434x is not 434x of nothing. It is 434x of a different opinion about what matters.

The takeaway

The top ten on this board is a price-performance table that happens to be sorted by composite. Rank 2 delivers 82.4 points at $0.38 per million blended tokens and the fastest p50 of the ten; rank 3 delivers 80.5 at $165.00 and 4400 ms. Paying the 434x buys 1.9 composite points — 0.13 standard deviations — and a later first response. Read that way, the question stops being which row is best and becomes which row crosses your quality floor at the lowest blended rate, which is a question with two or three defensible answers rather than ten. Rank 2 is one of them. Rank 9 is another at a lower floor. The expensive rows are answers to requirements the composite column does not measure, and that is worth knowing before the invoice rather than after.

Sourcing note: ranks, composite scores, seven-day success rates, p50 latency figures and blended $/1M prices are read from the OrcaRouter leaderboard on 2026-10-14, with the board itself updated 2026-10-13 [OURS]. Composite scores are our own, built from the board’s published evidence families; they are not a third-party benchmark. The independent intelligence index values quoted for three rows are third-party measurements the board lists as evidence [INDEPENDENT]. Input and output rates are the providers’ published prices with $0 markup [VENDOR]. Cost-per-100M and cost-per-point figures are our arithmetic on those public numbers. Latency and success-rate figures are rolling telemetry and will drift. All checked 2026-10-14.