Agent Mag
SUBSCRIBE
Mercury 2
All models
InceptionInceptionReleased 2026-03-04

Mercury 2

128K context$0.250/M input$0.750/M output

Mercury 2 is a reasoning diffusion large language model (dLLM) designed for high-speed token generation and refinement in parallel, achieving over 1,000 tokens per second on standard GPUs. It supports tunable reasoning levels, 128K context length, native tool use, and schema-aligned JSON outputs, making it suitable for coding workflows, real-time voice/search applications, and agent loops. Mercury 2 is notable for being 5x faster than leading speed-optimized models like Claude 4.5 Haiku and GPT 5 Mini, while maintaining cost efficiency.

What is Mercury 2?

Mercury 2 is an AI model from Inception that Agent Mag tracks for pricing, context window, modalities, benchmarks, and API compatibility. Builders can use this page to compare Mercury 2 against other models for agent workflows and production deployments.

Model ID

Mercury 2 is a reasoning diffusion large language model (dLLM) designed for high-speed token generation and refinement in parallel, achieving over 1,000 tokens per second on standard GPUs. It supports tunable reasoning levels, 128K context length, native tool use, and schema-aligned JSON outputs, making it suitable for coding workflows, real-time voice/search applications, and agent loops. Mercury 2 is notable for being 5x faster than leading speed-optimized models like Claude 4.5 Haiku and GPT 5 Mini, while maintaining cost efficiency.

Architecture & Specifications
Architecture
Reasoning Diffusion LLM (dLLM)
Tokenizer
Other
License
Proprietary
Released
2026-03-04
Modalities
Input
text
Output
text
Supported Parameters
include_reasoningmax_tokensreasoningresponse_formatstopstructured_outputstemperaturetool_choicetools
Strengths
  • Extremely fast token generation (>1,000 tokens/sec)
  • Supports 128K context length
  • Tunable reasoning levels
  • Native tool use and schema-aligned JSON outputs
  • Cost-efficient compared to leading models
Limitations
  • Lower performance in research-level physics reasoning (CritPt: 0.8%)
  • Moderate accuracy in knowledge-based tasks (AA-Omniscience Accuracy: 20.5%)
  • Higher structured output error rate (2.44%) compared to some models
  • Tool call error rate of 4.82%
  • Limited economic task performance (GDPval-AA: 23.0%)
Recommended Use Cases
Coding workflows with low latency requirements
Real-time voice and search applications
Agent loops for autonomous systems
Schema-aligned JSON output generation
High-speed reasoning tasks

Related content

Browse the AI models directory

Compare pricing, local installs, context windows, and modality filters across the full model catalog.

Explore AI agent tools

Find frameworks, SDKs, and infrastructure tools that pair with this model in production workflows.

Read Agent Mag model and agent articles

See Agent Mag coverage of model benchmarks, agent frameworks, and deployment patterns.

Data enriched Apr 24, 2026. Pricing from OpenRouter API.