There are two ways to call the same model. Through OrcaRouter, you can address the alias — `deepseek/deepseek-v4.1-flash` — which tracks whatever DeepSeek last rolled out, or the DeepSeek V4.1 Flash 1017 snapshot, which pins the October 17, 2026 checkpoint and never moves [OURS]. Both cost $0.15 in and $0.60 out per million and both carry a 1,048,576-token context. The choice between them is not about price or capability. It is about whether you can tolerate your model changing under you.
In the seven days ending 2026-10-26, the pinned snapshot moved 48,238,639,973 tokens on our gateway while the alias moved 22,119,239,520,045 [OURS]. The alias carries about four hundred and sixty times more. That ratio is the whole argument, and it cuts both ways.
What “pinned” actually buys
An alias is a promise that you get improvements. A snapshot is a promise that you get the same model twice. Those are different products sharing a price tag.
For a chat application, the alias is the right call. You want the better model on Tuesday without shipping a release.
For anything with an evaluation harness, the alias is a slow-motion audit failure. Your regression suite measures model A on Monday and model B on Friday and reports the difference as noise. The failure is quiet, which is what makes it expensive: nothing breaks, the scores just drift, and the drift gets attributed to the prompt, the data, or the engineer.
The snapshot removes that variable. If a number moves, it moved because you changed something.
The interesting part is that most traffic chose the alias anyway
Ninety-nine point eight percent of the traffic in this pair goes to the alias. That is not irrational — most calls are one-off questions where reproducibility has no value — but it means the 48 billion on the snapshot represents a specific kind of customer: one who has an evaluation harness, or a compliance requirement, or a production system where a silent model update is a change-management event.
Within that 48 billion, the serving profile is unusually clean:
| Metric (7 days) | Snapshot 1017 | Alias |
| p50 first token | 1,453 ms | 1,626 ms |
| p95 first token | 6,169 ms | 8,040 ms |
| Output throughput | 199.2 tok/s | 185.4 tok/s |
| Error rate | 0.0089% | 0.0355% |
| Tokens | 48,238,639,973 | 22,119,239,520,045 |
The snapshot is faster and more reliable than the alias on every measured line [OURS]. That is not a claim that the pinned checkpoint is better — it is a narrower workload on a fixed deployment, so the tail is naturally tighter. But it is a real operational advantage for a pinned card: when you freeze the model, you also freeze the thing you have already tuned your timeouts around.
An error rate of 0.0089% is roughly one failed request in eleven thousand. On a batch of a hundred thousand calls, expect about nine failures — a number you can retry your way out of rather than design around.
The third-party number and why it does not contradict ours
An independent public API-traffic ranking covering 2026-10-19 to 10-25 put V4.1-Flash-1017 first worldwide at 33.6 trillion tokens, a +801% jump, while the previous V4.1 Flash fell to 18 trillion as traffic migrated [INDEPENDENT].
Thirty-three point six trillion globally versus forty-eight point two billion through our gateway is a factor of about seven hundred. The two figures are not in conflict; they measure different denominators. That ranking aggregates public API traffic across many gateways and excludes official apps and private deployments. Ours is one gateway’s slice.
The direction, though, agrees: after October 17, traffic went to the pinned checkpoint, and the alias’s predecessor lost it.
What the snapshot does not pin
Three things still move even when the checkpoint does not.
Your provider’s serving stack. The weights are fixed; the hardware, batch sizes and routing are not. Latency will drift even on a pinned model.
The upstream. The card is served upstream under DeepSeek’s own `deepseek-flash` model name [OURS]. Pinning here means pinning the checkpoint the gateway routes to, not owning the weights.
Everything around the model. Your prompt, retrieval index, tool definitions, and data all continue to change. The snapshot only removes one variable from the experiment.
That last point is worth saying because “pinned” invites over-trust. It removes the largest uncontrolled variable in most LLM regressions. It does not remove the rest.
Moving to a new snapshot without a surprise
Pinning raises the question people actually ask: what happens when I want to move?
The answer is a procedure, not a decision. Four steps, in order.
Run the evaluation suite against the new snapshot while you are still on the old one. The suite has to be the same one you used to set your acceptance thresholds; a suite you wrote after seeing the new results is not a test.
Compare on your own cases, not on a leaderboard. The differences that matter are the ones on your prompts: a changed refusal boundary, a changed output format, a changed sense of when to stop.
Look at the distribution of outputs, not just pass rates. If forty percent of answers change in wording and two percent change in substance, the pass rate is unchanged and your users will notice. Sample a hundred outputs and read them.
Switch the model string behind a flag and keep the old checkpoint addressable. Because both cards are on the same key, the old snapshot remains callable after the switch, which makes a rollback a config change rather than a redeploy.
That last property is the reason to pin rather than to self-host a copy of the weights. Self-hosting gives you the same guarantee and a large new surface area — GPUs, serving stack, upgrades, capacity — to maintain. A dated checkpoint on a hosted catalogue gives you the guarantee for the cost of a string.
The one thing you cannot get from either is a promise that the old checkpoint stays available forever. Deprecation is a real risk with dated snapshots, and the mitigation is the same as with any dependency: know which model string your production traffic uses, and watch for it disappearing.
When to pay for it
You pay nothing extra — the rates are identical. The cost is opportunity: you do not get future improvements until you deliberately move to a new snapshot.
The rule that works: pin anything that has a regression suite, a compliance trail, or a customer-facing SLA; alias everything else. A migration then becomes a date you choose, with your tests run against the new checkpoint before the switch, rather than an event that happens to you on a Tuesday.
The takeaway
The snapshot and the alias are the same model at the same price; the difference is whether your Tuesday looks like your Monday. Most traffic takes the moving target, which is correct for ad-hoc calls and wrong for anything measured.
For the workloads that do pin, the evidence is decent: forty-eight billion tokens in a week, 199.2 tokens per second, a 1,453 ms median first token, and an error rate of nine requests in a hundred thousand.
Sourcing note: pricing, context, modalities and the snapshot-versus-alias distinction are from the OrcaRouter model page, checked 2026-10-27 [OURS]. Latency, throughput, error rate and token volumes are OrcaRouter production telemetry for the seven days ending 2026-10-26 [OURS]. The global weekly volume figure comes from an independent public API-traffic ranking covering 2026-10-19 to 2026-10-25 and measures a different population [INDEPENDENT].
See More:misalignment.net