One thousand three hundred and eighty-five. That is how many times more expensive the priciest output token is than the cheapest one you can buy today — $180.00 per million against $0.13 — and both models will answer the same prompt. OrcaRouter passes these through at list price, which is what makes an honest llm pricing comparison possible: the numbers below are what you pay, not a rate card with a markup folded in.
Prices read 2026-09-09. The table covers the nineteen models whose prices were verifiable directly on their own pages that day; the catalogue is larger, and anything unreadable that day is left out.
Table of Contents
The ladder
| Model | Input /1M | Output /1M | Output ÷ input |
| Qwen3.7 Flash | $0.03 | $0.13 | 4.3x |
| GLM 5.3 Flash | $0.07 | $0.25 | 3.6x |
| DeepSeek V4 Flash | $0.24 | $0.73 | 3.0x |
| DeepSeek V4 Pro | $0.73 | $2.18 | 3.0x |
| Gemini 3.6 Flash | $0.75 | $3.75 | 5.0x |
| Muse Spark 1.2 | $1.25 | $4.25 | 3.4x |
| Gemini 3.5 Flash | $1.50 | $9.00 | 6.0x |
| Qwen3.8 Max | $2.00 | $6.00 | 3.0x |
| Grok 4.6 | $2.00 | $6.00 | 3.0x |
| GPT-5.6 Sol | $4.00 | $20.00 | 5.0x |
| Claude Opus 5 | $5.00 | $25.00 | 5.0x |
| Claude Opus 4.8 | $5.00 | $25.00 | 5.0x |
| Claude Fable 5.1 | $10.00 | $50.00 | 5.0x |
| Claude Fable 5 | $10.00 | $50.00 | 5.0x |
| GPT-6 Astra | $10.00 | $50.00 | 5.0x |
| GPT-5.5 Pro | $30.00 | $180.00 | 6.0 |
What the ladder does not tell you
A price table is where most comparisons stop, and it is where the three expensive mistakes start.
Per-million rates do not order per-piece cost. Models do not emit the same amount of text for the same brief. On EQ-Bench Creative Writing v3 — an independent benchmark, read 2026-09-08 — GPT-5.6 Sol produced 8,548 output tokens where Claude Opus 5 produced 6,003. At $20.00 and $25.00 per million respectively, one comparable piece costs $0.171 on Sol and $0.150 on Opus 5. The model with the cheaper rate produces the higher bill. Those lengths come from creative-writing prompts specifically, so treat the ratio as indicative rather than universal — but the inversion is real and per-million tables cannot show it.
**The output-to-input ratio decides what a model is cheap *for*.** Look at the right-hand column. DeepSeek’s models charge 3.0x their input price on output; Gemini 3.5 Flash and GPT-5.5 Pro charge 6.0x. For work that reads a lot and writes a little — summarising long documents, extraction, classification — the input price dominates and a 3x model’s advantage largely disappears. For generation, where output dominates, a 6x model’s premium doubles. Two models at the same headline price can be the cheaper choice on opposite workloads.
Price does not predict speed. OrcaRouter’s own production telemetry, read the same day: DeepSeek V4 Flash reaches first token in 435 ms at $0.24 / $0.73 — the fastest figure in the whole table, from near the bottom of the price ladder. Meanwhile Qwen3.7 Flash, the cheapest model here, sits at 9.79 s. “Cheap is slow” is not a rule; it is an assumption that this data contradicts in both directions.
The version trap
One row pair deserves calling out on its own, because it costs money for nothing.
Gemini 3.5 Flash is $1.50 / $9.00. Gemini 3.6 Flash is $0.75 / $3.75. Same family, same “Flash” tier, and the *lower* version number is the more expensive one — by exactly 2x on output. If you integrated Flash at some point and never revisited the model string, you are paying double for a model the vendor has already superseded.
Version numbers do not order price, and they do not order quality either. Claude Opus 4.8 and Claude Opus 5 are both $5.00 / $25.00, and their median time-to-first-token is 4.38 s and 2.52 s. Same money, nearly half the wait.
Reading the ladder for your own workload
Four questions, in order, and they narrow it fast.
1. Does your work read more than it writes? If yes, sort by input price and the ladder compresses dramatically — $0.03 to $30.00 is a 1,000x spread, but the models people actually shortlist cluster between $0.24 and $5.00. If no, sort by output price and the spread stays brutal.
2. How many calls a month? Under a few thousand, the entire table is a rounding error against one engineer-hour. Above a few hundred thousand, the difference between the $0.73 tier and the $25.00 tier is the difference between a line item and a headcount.
3. Is a person waiting? Then first-token latency belongs in the decision, and it is not correlated with price — see above. Two of the figures in that data (Grok 4.6 and Gemini 3.6 Flash, both at 10.00 s) sit at a reporting ceiling rather than being measured values, so treat those two as “not measured” rather than “slow”.
4. Does the length matter? If your prompts are open-ended, multiply expected output length by the output rate rather than comparing rates. That is the only figure your invoice agrees with.
A note on how fast this data moves
Two of the columns above are not stable, and pretending otherwise is how comparison pages go stale.
Prices change, usually downward, and a cut on a model you already use requires no work from you at all. Latency changes faster still: the time-to-first-token figures are 7-day rolling windows and move daily. Claude Opus 5’s median moved 3.09 s → 2.82 s → 2.52 s across three consecutive reads. What survives that drift is the ordering, not the decimals.
So take a date with every number, this table included, and re-read the ones that matter to you quarterly. A pricing comparison without a read date is not a comparison, it is a snapshot somebody forgot to timestamp.
What a router changes about this table
Worth being explicit about the structural point, because it is the reason these figures are comparable at all.
Every rate above is a list price passed through unchanged. That matters in a way that is easy to skip past: when a platform adds a margin, the margin is usually a percentage, and a percentage applied to a 1,385x spread does not preserve the ordering you are trying to read. A flat fee, or a markup that varies by tier, quietly re-orders the ladder. If you are comparing rates from a reseller, the first question is whether the numbers are the vendor’s or the reseller’s.
The second thing it changes is the cost of being wrong. Every argument in this article ends with “so measure it on your own prompts”, and that advice is only actionable if trying a second model is a configuration change rather than a procurement exercise. One key against the whole ladder is what makes a twenty-brief comparison an afternoon instead of a quarter — which is why the recommendation throughout is to test rather than to trust a table, including this one.
None of which is a reason to use any particular platform. It is a reason to check, before accepting any pricing comparison, whether the person publishing it is quoting the vendor’s price or their own.
The takeaway
The 1,385x headline is real, and it is also the least useful number in this article. What decides your bill is the interaction between four things a price table cannot show: how much text your model actually emits, whether your workload is read-heavy or write-heavy, whether the output-to-input ratio is 3x or 6x, and whether the version string you integrated two quarters ago has since been superseded by something cheaper. Get those right and the ladder becomes navigable; read only the per-million column and you can pick the cheaper rate and still pay more.
