Inception Mercury 2.5 Explained: Diffusion LLM Speed, Pricing, Use Cases and How to Evaluate It

Inception's Mercury 2.5 diffusion LLM generates 1,107 tokens per second with a 260K context window at $0.20/$0.75 per million tokens. How diffusion language models differ, where Mercury is used, what the company-reported numbers mean, and how to test it. Updated with Mercury 3.

PUBLISHED // UPDATED
DURATION // 06 MIN READ
Server room aisle with cool white lighting and a laptop on a cart showing a terminal, no readable text or logos
FIG.01 // VISUAL DOCUMENTATION Server room aisle with cool white lighting and a laptop on a cart showing a terminal, no readable text or logos

Inception released Mercury 2.5 on September 8, 2026. It is a diffusion large language model (dLLM): instead of writing one token at a time like most chatbots, it generates and refines blocks of text in parallel. Inception reports 1,107 tokens per second on widely available NVIDIA GPUs, a 260K-token context window, and pricing of $0.20 per million input tokens and $0.75 per million output tokens, discounted 80% at launch. The company positions it against cost-optimised models such as GPT-5.6 Luna at low reasoning, Gemini 3.5 Flash-Lite and Claude Haiku 4.5. This article explains how diffusion LLMs differ, summarises what Inception announced, and suggests how to evaluate the speed claim for your own workload.

Update (November 2026): Inception has since released Mercury 3, with a 1M-token context window, 1,068 tokens per second on NVIDIA GPUs and pricing of $0.20 input and $0.70 output per million tokens (50% off at launch, with cached input 90% off). Inception says it is more than 40 times larger than Mercury 1. The evaluation advice below applies to both models.

Mercury 2.5 at a glance

All figures below are from Inception’s announcement and are company-reported.

ItemMercury 2.5
ArchitectureDiffusion language model
Throughput1,107 tokens per second on widely available NVIDIA GPUs
Context window260K tokens
Standard price$0.20 per million input tokens, $0.75 per million output tokens
Launch price80% off: $0.04 input, $0.15 output per million tokens
Quality claim40% increase in intelligence over Mercury 2; comparable to GPT-5.6 Luna (Low), Gemini 3.5 Flash-Lite and Claude Haiku 4.5
CapabilitiesTunable reasoning, parallel tool calls, schema-aligned JSON
AvailabilityInception API and chat, Baseten, OpenRouter; 100 million free API tokens via Inception

Inception says Mercury 2.5 is, to its knowledge, the largest diffusion language model ever trained. It also previewed Mercury Voice, a dLLM for voice agents with time to first token under 170 milliseconds, and Mercury Router, which analyses prompts and routes them to open or closed models based on quality, speed and cost. TestingCatalog’s report on the launch matches these figures.

How diffusion language models differ

Most production LLMs are autoregressive. They generate one token, feed it back in, generate the next, and repeat. That is effective but sequential: token 50 cannot be produced before token 49.

A diffusion language model starts from a masked or noisy version of the whole response and refines it over several steps, updating many positions in parallel. The idea comes from image diffusion models, which gradually denoise a picture. As Shattered’s explainer notes, the parallelism is what allows diffusion models to report throughput several times higher than comparable autoregressive models. The historical trade-off has been reasoning depth and coherence on long outputs, which is why Inception emphasises that Mercury 2.5 supports tunable reasoning rather than only speed.

Practical implications for developers:

  • Throughput versus latency. High tokens per second matters most for long outputs. For short replies, time to first token and network overhead may dominate.
  • Streaming behaviour. Because diffusion refines blocks rather than appending tokens, streaming can feel different from a typical chat model. Check how your interface handles it.
  • Model behaviour. Prompting techniques tuned for autoregressive models usually transfer, but test structured outputs and tool calls carefully.

Where Inception says Mercury is used

Inception describes three main production uses, with customer figures it attributes to those customers:

Search agents and retrieval pipelines. One search request can trigger dozens of model calls for planning, query rewriting, reranking, fact structuring, summarising and answer checking. Faster calls help keep the whole chain within a single user interaction.

Voice agents. OpenCall, which builds AI phone agents, reported median model response latency close to 170 milliseconds on its workload. Its CEO said P99 response time dropped from several minutes to one second and P50 from 0.4 seconds to under 0.2 seconds, including reasoning.

Coding sub-agents. Augment Code uses Mercury for context compaction, model routing and MCP tool search. Inception says moving compaction to Mercury cut latency by 82%, from about 150 seconds to 27 seconds, and reduced cost by 90% while maintaining quality.

These are customer testimonials published by the vendor, not independent benchmarks. They are useful as indications of where speed matters most: high-volume supporting calls rather than the single most demanding reasoning step.

Cost comparison in practice

At standard prices, a workload of 10 million input tokens and 2 million output tokens a day would cost about $2.00 for input and $1.50 for output, or $3.50 a day, on Mercury 2.5. During the launch discount, the same workload would be about $0.70. These are list-price calculations; real costs depend on prompt length, retries and reasoning settings. Compare against the actual prices of the alternatives you are considering, since prices for small models change frequently.

How to evaluate Mercury for your workload

Vendor speed figures are measured under the vendor’s conditions. A practical evaluation:

  1. Pick representative tasks. Use real prompts from the workflow you want to speed up, such as query rewriting, summarisation or classification.
  2. Measure end to end. Record time to first token, total response time and P50/P95/P99 latency from your own region, not just tokens per second.
  3. Compare quality on your data. Score outputs against your current model with the same rubric. A faster model that needs more retries may not save time.
  4. Test structured output and tools. Check JSON schema adherence and parallel tool calls under load.
  5. Check long context. If you plan to use large windows, test retrieval accuracy at different positions; see our overview of how context windows evolved.
  6. Price the full workflow. Include retries and reasoning tokens.
  7. Review deployment options. Inception says enterprise deployments support dedicated capacity, autoscaling, compliance controls and configurable data retention.

Safety and robustness also deserve testing. Small, fast models can behave differently under adversarial prompts; our article on perturbation probing and LLM safety describes one approach. For a contrasting frontier release from the same period, see our coverage of GPT-6 Astra’s safety and hidden reasoning.

Our analysis

Mercury 2.5 is a meaningful data point for diffusion language models: a commercial model reporting quality comparable to established small models while generating text much faster. The strongest case is for workflows with many short, supporting model calls, such as search pipelines, voice agents and coding sub-agents, where latency compounds. The numbers are company-reported, so the right next step is a side-by-side test on real tasks measuring end-to-end latency, quality and cost. With Mercury 3 now available, the gap between diffusion and autoregressive small models appears to be narrowing, which makes such tests worth repeating as new versions arrive.