Workload requirements analysis (budget-friendly ≠ one metric)
For budget-friendly workloads, you should optimize the combined cost of quality per dollar and effective throughput (tokens/s). Quality is the gating factor for correctness; inference latency determines how many completed tasks you can run per unit time; and price/1M determines hard cost.
Primary metrics that matter
Quality (task success / correctness): prioritize for anything that’s not purely autocomplete (e.g., code generation with tests, extraction, structured outputs).
Price/1M tokens: directly impacts unit economics; compare apples-to-apples on $/1M.
Speed (tokens/s): impacts concurrency and wall-clock SLA; higher tokens/s reduces time-to-completion under the same token budget.
Open source (only if required): among these candidates, open source matters operationally (self-hosting, governance) but does not automatically win on price/quality.
Derived decision rule (practical)
Use a two-step rule:
Filter by high quality at non-zero, stated price (exclude entries with $0.00 unless you have a separate internal pricing policy).
Among the survivors, rank by the best quality-cost balance, then break ties with speed.
Tier 1 picks (top 2–3 with data)
These are your budget-forward default choices for most production-ish workloads.
Quality:49.6 (highest in the provided priced candidates set)
Price/1M:$0.53 (non-zero, stated)
Speed:43 tok/s Why it wins: It provides the strongest quality signal without the “free but unusable” pricing entries ($0.00). If your workload is quality-sensitive (extraction, summarization with constraints, code synthesis that must be correct), this is the safest budget choice.
2) GPT-5.4 nano (xhigh) — best “quality + throughput” among OpenAI budget options
Quality:44.4
Price/1M:$0.46
Speed:212 tok/s (massively faster than most peers)
It’s lower quality than , but the makes it ideal when you need high concurrency or short inference SLAs.
Fique por dentro
Análise semanal de LLMs direto no seu email. Sem spam.
Speed:34 tok/s (slowest among the top budget set)
Why it makes Tier 1: If your workload benefits from reasoning-style behavior (multi-step extraction, tool-less planning, complex classification), the low $0.32/1M can dominate total cost even with lower tokens/s—provided you can tolerate throughput.
Tier 2 / budget alternatives (when your constraints differ)
Use these when Tier 1 doesn’t match your specific constraint.
Qwen3.5 27B — open-source budget with decent speed
Quality:42.1
Price/1M:$0.82
Speed:89 tok/sPick it if: you explicitly need open source or prefer Qwen-family deployment options. It costs more than DeepSeek and MiniMax, and it doesn’t beat MiniMax/GPT-5.4 on quality or speed.
M2.1: Quality 39.4, $0.53, 38 tok/s Pick it if: you’re locked to the MiniMax endpoint variants and need a fallback, but don’t choose them ahead of M2.7.
Speed:201 tok/s Pick it if: your workload is throughput-heavy and you accept lower quality (e.g., drafting code scaffolds where a secondary verification pass exists).
Default budget pick (most workloads): MiniMax-M2.7. It has the highest provided quality (49.6) at a reasonable $0.53/1M, making it the most reliable way to reduce expensive retries.
If you have strict latency / high concurrency constraints: GPT-5.4 nano (xhigh). The 212 tok/s throughput dominates throughput-limited pipelines while still staying relatively low cost ($0.46/1M).
If you are cost-minimizing and can accept slower inference or add parallelism: DeepSeek V3.2 (Reasoning). It’s the cheapest priced option ($0.32/1M) with solid reasoning quality (41.7) and open-source availability.
To finalize your shortlist, run your own token-budgeted evals (same prompt templates, same max tokens, same scoring rubric). Then lock one primary model and one fallback. Use the Explore or LLM Selector to apply your workload constraints (latency, budget ceiling, open-source requirement) to this candidate set.