When do you need a reasoning model?
As of September 8, 2026, the cheapest live-measured option within 90% of top quality on multi-step reasoning is ling-3.0-flash at $0.0162 per 1,000 requests — 1.000 measured quality vs 1.000 at the top, with a 69× price spread across the measured frontier (frontier v6).
re-measured weekly · how these numbers are made
Multi-step reasoning is the workload frontier pricing was built on — and the one where the "do I need it" question has a measurable answer. The suite scores multi-step problems with deterministic ends, so a model’s reasoning is graded on reaching the right result, not on showing appealing work.
The measured frontier
Every row is a live measurement on Potion's held-out multi-step reasoning suite — same items, same scoring, per option. Names that are part of the product are withheld; their numbers are not.
| option | vendor | measured quality | $ / 1K requests | p95 latency |
|---|---|---|---|---|
| ling-3.0-flash | inclusionai | 1.000 | $0.0162 | 7067 ms |
| nemotron-3.5-lightning | nvidia | 0.980 | $0.0733 | 3343 ms |
| inkling-small | thinkingmachines | 1.000 | $0.3450 | 5181 ms |
| gemini-2.5-flash | 0.900 | $0.9105 | 3232 ms | |
| inkling | thinkingmachines | 1.000 | $0.9589 | 3973 ms |
| gemini-3.7-flash | 1.000 | $1.13 | 3887 ms |
frontier v6 · measured 2026-09-08 · live provider calls only — simulated evidence never appears on this page
Head to head
ling-3.0-flash vs nemotron-3.5-lightninginkling-small vs ling-3.0-flashgemini-2.5-flash vs ling-3.0-flashinkling vs ling-3.0-flashgemini-3.7-flash vs ling-3.0-flashinkling-small vs nemotron-3.5-lightninggemini-2.5-flash vs nemotron-3.5-lightninginkling vs nemotron-3.5-lightninggemini-3.7-flash vs nemotron-3.5-lightninggemini-2.5-flash vs inkling-smallinkling vs inkling-smallgemini-3.7-flash vs inkling-smallgemini-2.5-flash vs inklinggemini-2.5-flash vs gemini-3.7-flashgemini-3.7-flash vs inkling
Questions
When is a reasoning model worth it?
When the measured gap on your problem class exceeds the interval — which happens on genuinely multi-step, error-compounding work, and largely does not on retrieval, extraction, or classification dressed up as reasoning.
How is reasoning scored without a judge?
By final-answer correctness on problems with known results, plus journey-grain suites where each step feeds the next and only the end artifact is scored.
Do reasoning tokens change the cost math?
Substantially — reasoning models spend hidden tokens, so their cost per request is measured from real usage, not list price per token. The frontier below prices what requests actually cost.
Potion routes each request to the cheapest option measured at your quality bar — with a receipt on every answer and this page's evidence behind every pick. Get an API key or read the docs.