Inception Labs is making an unusual pitch with Mercury 2.5: not a smarter model, but one fast and cheap enough to be called dozens of times inside a single user interaction. The startup says it is the largest diffusion language model ever trained, clocks 1,107 tokens per second on widely available NVIDIA GPUs, and lists at $0.20 and $0.75 per million input and output tokens β discounted 80%, to $0.04 and $0.15, at launch.
Key takeaways
- Mercury 2.5 delivers a claimed 40% intelligence gain over Mercury 2 while keeping the same low-latency, low-cost serving profile.
- Inception positions its quality against cost-optimised frontier tiers β GPT-5.6 Luna (Low), Gemini 3.5 Flash-Lite and Claude Haiku 4.5 β rather than against flagship reasoning models.
- Augment Code reports that moving context compaction to Mercury cut latency 82%, from roughly 150 seconds to 27 seconds, and reduced cost by 90%.
Why a diffusion LLM competes on latency
Conventional large language models emit one token after another, so wall-clock time scales with how much they say. Diffusion models refine a whole block of text across parallel passes instead, and that architectural difference is why Inception Labs can quote four-digit throughput on commodity accelerators rather than on specialised inference silicon.
The company frames Mercury 2.5 as the first output of a feedback loop rather than a benchmark push: usage grew more than an order of magnitude after Mercury 2 shipped, across thousands of developers and dozens of enterprise deployments, and Inception says it rebuilt its evaluations around the production failure cases that growth exposed before starting training.
Where customers are running it
Search shows why per-call latency compounds. One query can fan out into dozens of model invocations β planning, query rewriting, reranking, fact structuring, summarisation, answer checking β and the whole chain has to finish before the user notices a pause, which is why Inception says several leading search-infrastructure companies now run Mercury in production.
Voice agents are the harder constraint, because latency there is audible as dead air on a phone call. OpenCall, which builds AI phone agents for live customer conversations, measures median model response latency near 170 milliseconds on its production workload, and co-founder and CEO Oliver Silverstein says the switch cut the company's P99 from several minutes to one second and its P50 from 0.4 seconds to under 0.2:
significantly faster than any other provider we've seen, and that's including reasoning.
Coding agents present the third pattern, and the numbers there are the most concrete in the announcement. Augment Code moved context compaction onto Mercury and watched it drop from roughly 150 seconds to 27 β an 82% cut β while costing 90% less at what the company says is unchanged quality, with MCP tool-search summaries returning in under a second.
Two previews aimed at the latency floor
Mercury Voice, previewed alongside the main release, is a dLLM tuned for the tightest budgets in spoken interaction, with time-to-first-token under 170 milliseconds. A second preview, Mercury Router, uses a diffusion model to read an incoming prompt and pick a downstream model, open or closed, on quality-speed-cost grounds.
Shruti Koparkar, a senior product manager in NVIDIA's Accelerated Computing Group, read the release as a marker of how quickly a new architecture can harden into production-ready systems on existing AI infrastructure. That framing is the real claim here: diffusion for text has spent years as a research curiosity, and speed numbers alone have not moved buyers before β OpenAI's own ultrafast tier leaned on Cerebras silicon to get a comparable jump, where Inception is claiming the gain comes from the architecture rather than the hardware.
Availability and what comes next
Mercury 2.5 serves through the Inception API, Baseten and OpenRouter, with 100 million free tokens for new API users and enterprise options covering dedicated capacity, autoscaling, compliance controls and configurable data retention. Tunable reasoning, parallel tool calls and schema-aligned JSON round out the feature set, the last mattering for agent pipelines that depend on structured output.
A successor is already in training. Inception describes it as its largest model yet, targeted for release within months, and says the goal is a capability leap that does not surrender diffusion's speed or token efficiency.
FAQ
How much does Mercury 2.5 cost?
List pricing is $0.20 per million input tokens and $0.75 per million output tokens. At launch Inception is discounting that by 80%, to $0.04 per million input and $0.15 per million output. New API users also get 100 million free tokens.
Is Mercury 2.5 open source?
No. Mercury 2.5 is a proprietary hosted model served through the Inception API, Baseten and OpenRouter. Enterprise customers can get dedicated capacity and configurable data retention, but the weights are not published.
What is a diffusion LLM?
A diffusion language model generates text by iteratively refining a block of tokens in parallel rather than predicting them strictly left to right. The approach trades architectural familiarity for throughput, which is why it is aimed at latency-sensitive work such as voice agents, search pipelines and agent scaffolding.
