Diffusion-based reasoning LLM that refines tokens in parallel, built for latency-bound agent loops, voice and extraction pipelines.
Output Speed *
Intelligence Index *
Context Window *
Input price
Output price
Performance
Parallel Token Refinement Moves the Latency Bottleneck
Benchmarks
Artificial Analysis: Fastest in Class, Competitive but Not Leading in Intelligence
Output Speed
Intelligence Index
MMLU *
GPQA *
HLE *
LiveCodeBench *
Technical Specifications
What the model supports
Comparison
Mercury 2 or GPT-5 Mini for Speed-Optimized Reasoning?
Inception positions Mercury 2 directly against speed-optimized reasoning models such as GPT-5 mini and Claude Haiku 4.5, claiming comparable quality at 5-10x the throughput. Both are low-cost reasoning models for agents, extraction and high-volume workloads. Mercury 2 also has lower output pricing than GPT-5 mini. The decision is whether the architecture's speed advantage outweighs the broader capabilities and ecosystem of a mainstream autoregressive model.
Choose Mercury 2 when end-to-end latency compounds across many sequential calls, such as agent loops, voice, search and bulk extraction. Its text-only 128K window is enough for those jobs. Choose GPT-5 mini when you need image input, a larger context window, or the maturity of the OpenAI ecosystem. Also choose it when stronger general intelligence matters more than speed. Mercury 2's quality parity is largely Inception's claim, and Artificial Analysis does not rate it as a leader in intelligence.
Limitations & Trade-offs
Best-Fit Workloads
Where this model earns its place
Multi-step agent loops and subagents
Agent workflows chain many model calls, so latency accumulates. Mercury 2 pairs native tool use and tunable reasoning with over 1,000 tokens per second reported by Inception. Artificial Analysis highlights agentic coding and terminal use as strengths. In a joint case study, Augment Code reported an 82% latency drop on a context-compaction subagent versus Claude Opus 4.7. Use it for fast subagents; keep a stronger planner model for the hardest decisions.
Voice and real-time conversational agents
Voice agents need reasoning to finish within a conversational turn. Inception markets Mercury 2 for this use and cites time to first token under 300 ms on standard NVIDIA GPUs. Treat that as a provider claim and verify it on your own endpoint and reasoning setting. It fits turn-based assistants, call-handling logic and intent routing where delays break the experience. It is text-only, so speech-to-text and text-to-speech are separate components.
Bulk structured extraction and classification
Schema-aligned JSON output, a 128K window and lower output costs suit high-volume extraction, tagging and classification. Artificial Analysis measured 70% on IFBench, which supports its use where outputs must follow exact format rules. Fast throughput shortens batch jobs. Because factual recall is weak, use it to transform supplied text rather than to answer from memory.
Retrieval pipeline reasoning steps
Query rewriting, reranking, relevance judgments and answer synthesis from retrieved passages all add latency before a user sees a response. Mercury 2 can handle these steps quickly within its 128K window. Inception positions it for retrieval pipelines. Ground answers in retrieved context, since its standalone knowledge recall is weak.