Inception released Mercury 2.5 on September 8, 2026, positioning it as a reasoning-focused language model that prioritizes speed and cost efficiency over raw intelligence. The proprietary model scores below average on the Artificial Analysis Intelligence Index but delivers output at a rate nearly seven times faster than its peers in the same price tier.
Where Mercury 2.5 sits on intelligence
The Artificial Analysis Intelligence Index gives Mercury 2.5 a score of 12, compared to a median of 13 among comparable reasoning models. The index evaluates models across reasoning, knowledge, mathematics, and coding. By that measure, Mercury 2.5 is not the strongest performer available. Inception has not disclosed the model's parameter count, so the architecture behind that score remains opaque.
What the index also reveals is how Mercury 2.5 achieves its result. During evaluation, the model generated 35 million output tokens, well below the median of 85 million for its tier. That makes it notably concise — it reaches its answer with fewer tokens than competitors, which is part of what keeps its cost low.
The speed advantage
Mercury 2.5's standout characteristic is output speed. At 770.4 tokens per second through Inception's API, it is roughly seven times faster than the median of 108.6 tokens per second for comparable reasoning models. For developers building latency-sensitive applications — real-time chat, interactive tools, or high-throughput services — that speed matters.
The trade-off is that its time to first token is 2.95 seconds, somewhat slower than the median of 2.23 seconds. The extended thinking process that reasoning models use before producing an answer takes longer to initiate, even though the actual output streaming is fast.
Pricing and efficiency
Mercury 2.5 costs $0.25 per million input tokens and $0.75 per million output tokens. The input price matches the median for its tier exactly, while the output price sits below the median of $0.90. On a blended basis — using a 7:2:1 ratio of cache hits, input, and output — the effective rate is $0.14 per million tokens. Evaluating Mercury 2.5 on the Intelligence Index costs approximately $0.06 per task.
The combination of low token usage and modest pricing makes it one of the more economical reasoning models available. The Artificial Analysis analysis notes it is "well priced when comparing to other models of similar price," which is the clearest positioning statement available for a model that does not lead on intelligence.
What it can and cannot do
Mercury 2.5 is a text-only model. It accepts text input, produces text output, and supports a 260,000-token context window — roughly equivalent to 390 pages of A4 text at 12-point Arial. It does not support image input and is not multimodal. It is also not open weights; the model is proprietary and available through a single API provider.
As a reasoning model, it uses extended thinking or chain-of-thought processing to work through complex problems before generating a final answer. That architectural choice is what separates it from non-reasoning models that produce output more directly, and it explains both the slower first-token latency and the efficient token usage.
The practical picture
Mercury 2.5 is a deliberate trade. Inception has built a model that sacrifices some intelligence in exchange for speed and cost — a combination that appeals to teams running high-volume inference where per-token expense and response time are the dominant constraints. For developers who need a reasoning-capable model that can sustain a conversational pace without the computational overhead of larger systems, Mercury 2.5 fills a specific niche.
Whether the below-average intelligence score is a limitation or a calculated optimization depends on the use case. For straightforward reasoning tasks, the model may perform adequately while costing less and responding faster than alternatives. For tasks that push the boundaries of what a reasoning model can handle, the score gap is worth noting before committing to it in production.