The cheapest route depends on the workload.

On RouterPlex the cheapest model is DeepSeek V4 Flash at $0.09 per 1M input tokens and $0.18 per 1M output tokens — about $0.011 for a 100K-token-in, 10K-token-out agent turn. Prices are vendor list rates, so the same model costs the same whether you call it here or direct.

Ranking on input price alone favours models that charge little to read and a lot to write, which is the wrong way round for coding and agent work. This table ranks on a fixed 100K-in / 10K-out turn so input and output rates both count, and the same profile is applied to every model.

Across this catalog the spread is about 139× on the same agent turn: DeepSeek V4 Flash costs roughly $0.011 where GPT-6 Astra costs about $1.50. Whether that trade is worth it depends on task difficulty — a cheap model that needs three attempts is not cheap.

Not meaningfully. RouterPlex bills vendor list prices with no markup and no top-up fee, so per-token cost matches going direct to each vendor. What you save is the overhead of separate accounts, keys and minimum balances per provider.

RouterPlex applies no additional RPM or TPM cap beyond the upstream vendor's. The real caveat is capability: budget models are generally weaker at long agentic tool loops, so compare quality on the benchmarks pages before moving production traffic onto one.