Businesses that build products on top of large language models (LLMs) usually pay per “token”, the small chunks of text a model reads and writes. Alongside price, one of the most important numbers for any AI-powered app is time to first token: how long a user waits before the answer starts to appear. Model routing platforms such as OrcaRouter, which let developers send requests to many different AI models through a single connection, track both figures across real-world traffic.
That data reveals a free performance upgrade sitting in a lot of production config files. Three pairs of models listed on OrcaRouter charge exactly the same price per token, and in each pair one is between 1.3 and 2.6 seconds faster to first token. Nothing on a standard pricing page shows this. Because OrcaRouter measures price and latency across live traffic, the pairs sit side by side, which is the only way an LLM pricing comparison catches a gap that the price column alone cannot show.
The three pairs (figures as of 9 September 2026)
In the table below, “p50” means the median, the typical wait time across all requests measured.
| Price per 1M tokens (input / output) | Model | p50 time to first token | Gap |
|---|---|---|---|
| $5 / $25 | Claude Opus 5 | 2.52 s | — |
| $5 / $25 | Claude Opus 4.8 | 4.38 s | +1.86 s |
| $10 / $50 | Claude Fable 5.1 | 4.14 s | — |
| $10 / $50 | Claude Fable 5 | 6.70 s | +2.56 s |
| $2 / $6 | Qwen3.8 Max (0902) | 2.82 s | — |
| $2 / $6 | Qwen3.8 Max | 4.09 s | +1.27 s |
In all three, the newer release is the faster one, and the price is identical to the cent. There is no trade being made, no “pay less, wait longer”, just a version string that somebody has not revisited.
Why this happens
Providers usually hold a tier’s price flat across releases and improve what sits behind it. Serving infrastructure gets better, the model gets distilled or re-quantised, and routing improves. The price is a product decision; the latency is an engineering outcome. They move independently, and the price is the one that gets published.
This means the default state of any long-lived integration is drift: a team picks the fastest option in the tier at integration time, the provider ships a successor, and the config still names the predecessor. Nothing breaks, nothing alerts, and the pricing page still shows the number that was budgeted for.
What 1.9 seconds is worth
For a batch job, nothing. Throughput and cost per piece are the metrics that matter, and when the first token arrives is irrelevant.
For anything with a person in front of it, quite a lot, and not in a linear way. A response that starts in 2.5 seconds reads as the system working. One that starts in 4.4 seconds reads as the system having stalled, and the user starts reaching for another tab. The gap is long enough to break attention and too short to do anything else with.
The real cost depends on the shape of a session rather than a single request. A drafting loop where someone asks, reads, adjusts and asks again runs dozens of round trips in a working session. At fifty round trips, 1.86 seconds adds up to more than a minute and a half of dead time, and the frustration is worse than the arithmetic because it lands fifty separate times.
Checking your own config
Ten minutes, once a quarter, is enough to catch exactly this:
- List every model string your code actually sends. Not just the primary model, but the fallback route and the batch route too, since those are the ones nobody revisits.
- For each, look up its current price and median time to first token, and note the date.
- Check whether a newer release exists in the same tier at the same price. That is the whole exercise. A same-price, faster successor needs no further analysis to justify a test.
- Check quality separately before switching. Faster and same-price does not automatically mean better at your task. It means the swap is free to test, not free to assume.
That last step matters. These pairs are not always drop-in equivalents on output quality, and a speed gain that costs accuracy is not a gain. What the data shows is that the test is worth running, because the cost side is already settled.
Two honest limits on these numbers
They are 7-day rolling windows, and they drift. Claude Opus 5’s median moved from 3.09 s to 2.82 s to 2.52 s across three consecutive readings. The ordering within each pair stayed stable across those readings; the decimals did not. The direction is the finding, and each specific gap is a reading with a date on it.
They reflect one platform’s traffic, not yours. Latency is a property of a route under load: prompt lengths, concurrency, region and time of day all play a part. A median measured on someone else’s traffic tells you which model in each pair to test first. It does not tell you exactly what you will see.
What a stale model string costs beyond latency
Latency is the most visible cost of not revisiting a model version, and usually the smallest one.
Price. Elsewhere in OrcaRouter’s listings, the same version mismatch shows up as money rather than milliseconds. As of 9 September 2026, Gemini 3.5 Flash was listed at $1.50 input / $9.00 output per million tokens, while Gemini 3.6 Flash was listed at $0.75 / $3.75: the older version at roughly double the price. That is a 2x saving sitting in a single config line.
Capability. Successors generally score better on published benchmarks, so a stale model string also means missing whatever the newer release improved. That improvement still has to be tested on your own task, but it cannot be evaluated if nobody knows the successor exists.
Deprecation risk. Older releases eventually stop being served. A pipeline pinned to a version nobody is maintaining is a scheduled outage with no date on it yet, and discovering that during an incident is far worse than discovering it during a quarterly review.
All three have the same fix, and it is the ten-minute audit rather than anything clever: list the model strings your code actually sends, look each one up, and check whether a same-family successor exists. It often stays undone because nothing breaks: no error, no alert, no warning on a dashboard. It is the most boring kind of technical debt, and one of the cheapest to clear.
The takeaway
Three price points on OrcaRouter each contain two models priced identically, with the newer release 1.27 to 2.56 seconds faster to first token. That is not a trade-off to weigh. It is a config line that has gone stale, and it stays invisible on pricing pages because price and latency are published by different systems and move independently. Spending ten minutes a quarter listing the model strings you actually send, and checking whether a same-price successor exists, is some of the cheapest performance work available.
Sourcing note: Input and output rates are the list prices shown on OrcaRouter as of 9 September 2026. Median time-to-first-token figures are OrcaRouter production data from the same date, not a controlled benchmark; they reflect its traffic mix, regions and live provider load, and each is a 7-day rolling window that changes daily. The figures say nothing about output quality, which should be tested separately before switching models.
Changes made
- New headline that makes it clear the topic is AI models, for a general news audience.
- New opening paragraph explaining what tokens, time to first token and OrcaRouter are, so non-technical readers can follow.
- Brand voice removed: “Elsewhere in this catalogue”, “on this platform” and “this one included” were rewritten in a neutral third-person voice (“OrcaRouter’s listings”, “on OrcaRouter”).
- Data dates made clear: the table heading and sourcing note now say “as of 9 September 2026” (UK date format) instead of “read 2026-09-09”.
- Source and date added for the Gemini 3.5 Flash / Gemini 3.6 Flash pricing claim.
- Table made clearer: column headings now explain “input / output” prices and “p50” (median) is explained above the table.
- Anchor text fixed: “llm pricing comparison” changed to “LLM pricing comparison”.
- Minor wording: “the user’s hand moves toward the tab” changed to “the user starts reaching for another tab”, plus small readability edits. No figures were changed.
Publisher guidelines / note (London Daily News)
- Maximum 3 links per article: 1 do-follow and 2 no-follow. This article has 2 links.
- Do-follow: “LLM pricing comparison” → https://www.orcarouter.ai/compare. No-follow: “OrcaRouter” → https://www.orcarouter.ai/.
- Category: Technology. Submit through the Tech account.
- All prices and latency figures are dated 9 September 2026 and change daily. If the client wants fresher numbers, they should update the table before submission.
- The editors may still adjust the headline or wording before publishing.
Feature image by Pexels
