Inception has released Mercury 2.5, calling it the most capable diffusion LLM on the market and, to its knowledge, the largest ever trained. The company reports a 40% intelligence gain over Mercury 2 while retaining the same low-latency, low-cost serving profile. It says quality is comparable to cost-optimized frontier models including GPT-5.6 Luna at Low reasoning, Gemini 3.5 Flash-Lite, and Claude Haiku 4.5.
Mercury 2.5 generates 1,107 tokens per second on widely available NVIDIA GPUs and supports a 260K-token context window. Its capabilities include tunable reasoning, parallel tool calls, and schema-aligned JSON. Standard pricing is $0.20 per million input tokens and $0.75 per million output tokens, with an 80% launch discount cutting those rates to $0.04 and $0.15.
Today we're excited to announce Mercury 2.5
— Stefano Ermon (@StefanoErmon) September 8, 2026
It’s the most capable diffusion LLM on the market. It is a 40% jump in intelligence over Mercury 2 and runs at over 1,100 tokens/sec on widely available @NVIDIAAI GPUs.https://t.co/xFnPBuJ577 pic.twitter.com/5h3FLhjQqY
The release follows rapid adoption of Mercury 2. Inception says thousands of developers have built with it, dozens of enterprises have deployed it in production, and usage has risen by more than an order of magnitude. Customer feedback and production failures shaped new evaluations and training priorities, making Mercury 2.5 the first result of that feedback loop.
The model targets latency-sensitive search, voice, and coding systems. Search-agent requests can require dozens of calls for planning, query rewriting, reranking, fact structuring, source summaries, and answer checks. OpenCall uses Mercury for live customer calls, where median response latency is close to 170 milliseconds. Its P99 response time fell from several minutes to one second, while P50 dropped from 0.4 seconds to under 0.2 seconds, including reasoning.
Augment Code uses Mercury for context compaction, model routing, and MCP tool search. Moving compaction to Mercury cut latency by 82%, from about 150 seconds to 27 seconds, and reduced cost by 90% while maintaining quality. Tool-search summaries arrive in under a second.
Inception is also previewing Mercury Voice and Mercury Router. Mercury Voice targets voice agents with time to first token below 170 milliseconds, while Mercury Router analyzes prompts and directs them to open or closed models based on the required balance of quality, speed, and cost.
Mercury 2.5 is available through Inception’s chat product and API, Baseten, and OpenRouter, with 100 million free API tokens offered through Inception. Enterprise deployments support dedicated capacity, autoscaling, compliance controls, and configurable data retention. Training is already underway on Inception’s next and largest model, which is targeted for release in the coming months without giving up the speed and token efficiency of diffusion models.
