Speed benchmarks are easy to game. Production data is not. Inception Labs built Mercury 2.5 by feeding real customer failures and usage patterns back into training, and the result is a model that delivers a 40% intelligence jump over Mercury 2 while holding the same low-latency, low-cost profile that made the original useful in the first place. That’s a harder thing to pull off than it sounds.
The company announced Mercury 2.5 as its most capable production model to date. According to Inception, it’s the largest diffusion language model ever trained, and it now sits in the same quality tier as GPT-4.1 Mini, Gemini 2.5 Flash-Lite, and Claude Haiku 4.5. Those are the models most developers reach for when they need something fast and cheap. Mercury 2.5 is now competing directly in that bracket.
What the model actually does
Mercury 2.5 runs at 1,107 tokens per second on standard NVIDIA GPUs. List price is $0.20 per million input tokens and $0.75 per million output tokens. At launch, it’s 80% off: $0.04 input and $0.15 output. That pricing is aggressive, especially for a model Inception is positioning against established frontier alternatives.
Key capabilities include:
- Tunable reasoning depth depending on task complexity
- Parallel tool calls for agentic workflows
- Schema-aligned JSON output for structured data use cases
- Sub-170ms time-to-first-token in voice-optimized configurations
The model is available through the Inception API, Baseten, and OpenRouter. Enterprise deployments get dedicated capacity, autoscaling, compliance controls, and configurable data retention. YC companies can claim $500,000 in deployment benefits, and Baseten customers can deploy through their existing setup without changes.
Where it’s already running in production
The clearest signal that Mercury 2.5 is more than a benchmark story is where it’s actually deployed. Search infrastructure is one of the biggest use cases. A single search request can trigger a dozen or more model calls: query rewriting, reranking, fact structuring, summarization, answer checking. Mercury handles those calls fast enough to keep everything inside a single user interaction. Several search companies are running it in production.
Voice is the other critical test. Latency in voice isn’t an abstract metric. It’s the pause a caller hears. OpenCall, which builds AI phone agents for live customer calls, reported that after switching to Mercury, median model response latency dropped to around 170 milliseconds. P99 response time fell from several minutes to one second, and P50 dropped from 0.4 seconds to under 0.2. Oliver Silverstein, co-founder and CEO of OpenCall, called it significantly faster than any other provider they tested, including reasoning models.
Coding agents are a third area. Augment Code uses Mercury for context compaction, model routing, and tool search. Moving compaction to Mercury cut latency by 82% (from around 150 seconds to 27 seconds) and reduced cost by 90%, with no reported quality drop. Tool-search summaries now return in under a second.
Two new products launching alongside it
Mercury Voice is a diffusion LLM built specifically for voice agents, targeting sub-170ms time-to-first-token. Mercury Router uses a diffusion LLM to analyze incoming prompts and route them to whichever model, open or closed, offers the best balance of quality, speed, and cost for that specific task. Both are in preview.
The router is worth watching. If it works reliably, it turns Mercury into infrastructure rather than just a model. Developers stop choosing between providers and let the router decide. That’s a different kind of product entirely.
Why this matters for the diffusion LLM space
Most production AI workloads are not single large inference calls. They’re chains of smaller calls, many of them happening in parallel, where latency and cost stack up fast. That’s the problem Mercury is actually solving. GPT-4o and Claude Sonnet are excellent, but they’re expensive and slower when you’re making 20 calls per user request.
Inception says it has already started training its next model, which it describes as its largest yet, with a release targeted in the coming months. If Mercury 2.5 is the first output of a real production feedback loop, the next version will be worth watching closely.
