Agnes 2.5 Pro Beta: Singapore's $0.10/M Reasoning Model Scores 49 on AA Intelligence Index
Sapiens AI’s Agnes 2.5 Pro Beta, released August 26, is the price-performance story at the bottom of the frontier reasoning tier. The Singapore startup moved its model from an Artificial Analysis Intelligence Index of 39 to 49 in a single iteration — a 26% intelligence jump — while cutting input price from $0.45 to $0.10 per million tokens. Output pricing dropped from $0.90 to $0.30.
The result: a model that sits in the same bracket as DeepSeek V4 Flash (Index 52) and GPT-5.6 Luna (Index 52) on price-per-token, while coming in cheaper on input cost than either.
Benchmark Numbers
- AA Intelligence Index: 49 (v4.1.1)
- AA Agentic Index: 44
- GPQA Diamond: 91% (graduate-level science reasoning)
- Terminal-Bench v2.1: 70% (agentic coding and CLI tasks)
- AA-LCR: 78% (long-context reasoning)
- AA-Omniscience non-hallucination rate: 67%
- Output speed: 159.5 tokens per second
- Context window: 1M tokens; max output 65,536 tokens
The GPQA Diamond score of 91% is the headline. Most models in the $0.10–0.50/M input tier score in the 70s on that benchmark. Hitting 91% at $0.10/M places Agnes within striking distance of models costing 20–50x more on pure scientific reasoning.
Weaker areas are clearly documented: AA-Omniscience non-hallucination rate at 67%, and 16% on CritPt (physics reasoning). These gaps locate the model accurately — strong on graduate-level structured Q&A, weaker on long-tail factual recall and hard physics.
Intelligence Index vs. Price
The comparison that matters for anyone building in the sub-$0.50/M input tier:
| Model | AA Index | Input $/M | Output $/M |
|---|---|---|---|
| GPT-5.6 Luna (max) | 52 | $0.20 | $1.20 |
| DeepSeek V4 Flash | 52 | $0.22 | $0.66 |
| Agnes 2.5 Pro Beta | 49 | $0.10 | $0.30 |
| MiniMax-M3 (open) | 45 | $0.30 | $1.20 |
Agnes trades three index points off DeepSeek V4 Flash in exchange for a 55% lower input price and a 55% lower output price. For inference-heavy workloads where cost scales with call volume, that tradeoff is material.
Cache pricing: $0.01 per million input tokens on cache hits — a 90% discount that significantly reduces the effective cost for repeated-context workloads like RAG pipelines and long-session agents.
Version History
Agnes AI has shipped three versions in five weeks:
- Alpha (July 24): Open-weights, Apache 2.0, Index 39, $0.45/$0.90
- Stable (August 1): Proprietary, Index 39, same pricing as Alpha
- Beta (August 26): Proprietary, Index 49, $0.10/$0.30
The Alpha remains available for self-hosting. Beta is the commercial API only. The trajectory from 39 to 49 in one generation is steep; whether the next iteration holds the pricing while continuing to climb is the question.
The Company
Sapiens AI is headquartered in Singapore and operates under the public brand Agnes AI. CEO Bruce Yang appeared on Forbes Asia’s 30 Under 30. The company raised a $10M Series A in February 2026 led by LOOK FORWARD VCC, bringing total funding to $20M.
At $20M raised, Agnes 2.5 Pro Beta is punching significantly above its capitalization weight relative to the frontier labs it is benchmarking against. The model is priced and positioned as infrastructure for budget-sensitive enterprise inference — the tier where GPQA-class reasoning matters but cost-per-token determines deployment viability.