Decision Model Pricing and Cost Benchmarks
An overview of token rates, benchmark task expenditures, and empirical support routing costs for leading decision models.
👁 115 reads
Evaluating the economics of decision models requires examining raw input token pricing, standardized composite task costs, and domain-specific throughput. Unlike standard text-generation models, decision model providers bill exclusively for input tokens while returning decision probabilities without output token fees. Pricing data verified on October 13, 2026, shows significant variance across both native cloud endpoints and aggregated API brokers.
Provider Token Rates
Entry-level pricing for decision models begins at $0.02 per million input tokens, with specialized or dynamic-reasoning architectures commanding higher fees. OpenRouter currently provides unified access to nine of the eleven models tracked on the primary benchmarks.
| Model | Producer Price / 1M Inputs | OpenRouter Price |
|---|---|---|
| Decider v1.1 27B | $0.02 | $0.02 |
| Mercury Decide | $0.04 (promo $0.02) | $0.02 (+ free variant) |
| d1 | $0.04 | $0.04 |
| Microsoft-Decision-1 | $0.042 | $0.042 |
| Jev 1.13 | $0.042 | $0.042 |
| Clef-Flash | $0.038 | from $0.021 |
| GPT-6 Luna Decisions | $0.10 (+10% EU/EEA) | $0.10 |
| GLiDE | $0.15 (all passes billed) | Not available |
| Clef | $0.24 | from $0.042 |
Artificial Analysis Task Costs
In the Decision Index v1.0 published on October 13, 2026, Artificial Analysis calculated the cumulative cost to execute one million benchmark tasks across varied reasoning, knowledge, and routing inputs. Decider v1.1 27B logged the lowest composite cost at $16.5, followed by Mercury Decide at $30.4, d1 at $30.3, Microsoft-Decision-1 at $30.7, and Jev 1.13 at $41.5. GLiDE reached $119.0, driven by internal reasoning passes that bill across all evaluation phases, while Cloudflare Clef stood at $199.4.
When isolating moderation and classification via Ai2’s WildGuardTest, task costs drop markedly: Decider v1.1 27B required $12.1 per million tasks, Microsoft-Decision-1 cost $12.5, Mercury Decide cost $12.6, Celeris-1 Decision cost $16.0, and Jev 1.13 required $22.8. In comparison, running Claude Haiku 5.5 low on WildGuardTest cost $91.5, while Claude Opus 5.5 low reached $3,774 per million requests.
Empirical Support Routing Costs
A separate test conducted on 50 customer-support tickets written in a non-English language (Polish) - a small, indicative evaluation - demonstrated how tokenization alters practical running costs. Mercury Decide recorded $4.4 per million tasks under promotional pricing, Decider v1.1 27B registered $7.8, and Microsoft-Decision-1 logged $11.5.
Jev 1.13 recorded a higher practical cost of $26.8 per million tasks due to tokenization expansion on non-English text, consuming approximately 640 input tokens per task compared to approximately 275 tokens for Microsoft-Decision-1. Even with this overhead, all decision architectures operated at a fraction of Claude Haiku 5.5, which recorded $96.7 per million tasks across the same dataset.
More: AI decision models - Decision Index ranking, prices and tests
More on this from QUANTSLANT
guides·Decision Model vs LLM: Architecture and PerformanceCompare decision models and traditional LLMs across routing tasks, latency profiles, calibration metrics, and cascading costs.