What is a diffusion LLM and how does it differ from a standard autoregressive model?

A diffusion LLM (dLLM) generates multiple tokens simultaneously by iteratively denoising a block of tokens rather than producing each token sequentially from left to right. Autoregressive models like GPT-4 or Claude generate one token at a time, which creates a hard ceiling on throughput. Because a diffusion model can refine an entire output block in parallel, it can produce thousands of tokens per second on the same GPU hardware that an autoregressive model uses to produce a few hundred.

What benchmarks support Mercury 2.5's quality claims?

Inception positions Mercury 2.5 as comparable in quality to cost-optimized frontier models including GPT-5.6 Luna (Low), Gemini 3.5 Flash-Lite, and Claude Haiku 4.5. However, the company notes that Mercury 2.5's specific 40% intelligence gain over Mercury 2 uses internal evaluations rather than a fully disclosed independent benchmark. Production deployments, such as Augment Code's context compaction pipeline, provide a more concrete signal: a reported 82% reduction in latency and 90% reduction in cost while maintaining output quality.

Is Mercury 2.5 available in enterprise environments with compliance requirements?

Yes. Mercury 2.5 is available via Inception's API, Baseten, and OpenRouter. Enterprise deployments support dedicated capacity, autoscaling, configurable data retention, and compliance controls. Inception also distributes through AWS Bedrock and Azure Foundry, which bring each cloud's native security and governance controls to the model.

What is Mercury Voice and when does it ship?

Mercury Voice is a diffusion model variant optimized for voice agents, targeting under 170 milliseconds time to first token. Inception announced it alongside Mercury 2.5 as a separate preview with no stated general availability date. Mercury Router, also in preview, selects across open and closed models based on quality, speed, and cost.

How does Mercury 2.5 pricing compare to other fast inference models?

At its 80% launch discount, Mercury 2.5 costs $0.04 per million input tokens and $0.15 per million output tokens. At standard pricing it is $0.20 in and $0.75 out. For comparison, Claude Haiku 4.5 and Gemini 3.5 Flash-Lite sit in a similar intelligence band; Mercury 2.5 claims substantially higher throughput (1,107 tokens per second versus the 100-400 range typical for autoregressive speed models) at a comparable or lower cost.

Mercury 2.5: Inception's Diffusion LLM Built for Enterprise Agents

On September 8, 2026, Inception launched Mercury 2.5, the latest release in its diffusion large language model (dLLM) line. The model produces 1,107 tokens per second in production, extends its context window to 260,000 tokens, and ships at $0.04 per million input tokens at launch discount. For enterprise teams running AI agents that make dozens or hundreds of model calls per task, those numbers rewrite the cost and latency calculation.

The launch confirms a shift that was theoretical only 18 months ago: diffusion LLMs have moved from research paper to production infrastructure.

What Diffusion LLMs Mean for Enterprise AI Inference

Every major language model in production today, including GPT-6 Astra, Claude Fable 5.1, and Gemini 3.8 Flash, generates text one token at a time. The autoregressive architecture is well understood and highly optimized, but it creates a throughput ceiling. A model that generates 200 tokens per second cannot generate 2,000 tokens per second on the same hardware simply by better engineering: the sequential dependency is a structural constraint.

Diffusion LLMs take a different path. Instead of predicting the next token conditioned on all previous tokens, a diffusion model starts with a block of masked or noisy tokens and iteratively refines them in parallel. The result is a model that can produce many tokens simultaneously. Inception’s Mercury 2.5 achieves 1,107 tokens per second on widely available NVIDIA GPUs at production scale, a throughput level that puts it in a different category from any autoregressive competitor at the same intelligence tier.

For agentic workloads, this matters more than it does for simple chat. An agent running a research loop, compacting context between tool calls, routing sub-tasks, or parsing returned documents at each step may invoke a language model 20 to 100 times per user task. At 150 tokens per second, those calls accumulate. At 1,107 tokens per second, the same pipeline finishes in seconds rather than minutes, and the cost per completed task drops accordingly.

This is the specific gap Mercury was built for. As the enterprise AI agent adoption gap has widened, inference cost and latency have appeared consistently as production blockers. Mercury 2.5 addresses both simultaneously.

Mercury 2.5 Capabilities and Specifications

Mercury 2.5 ships with several features that make it usable as a drop-in component in production agent pipelines:

Specification	Mercury 2.5	Mercury 2
Context window	260,000 tokens	128,000 tokens
Throughput	1,107 tokens/sec	~600-800 tokens/sec
Input price (standard)	$0.20/M	$0.25/M
Output price (standard)	$0.75/M	$0.75/M
Launch discount input	$0.04/M	n/a
Launch discount output	$0.15/M	n/a
Tunable reasoning	Yes	No
Parallel tool calls	Yes	No
Native JSON/schema output	Yes	Limited

The doubled context window matters for enterprise document processing, where a single contract, report, or codebase section may exceed 100,000 tokens. Tunable reasoning lets developers dial the model’s internal computation budget based on task complexity, trading tokens for quality on harder problems. Parallel tool calls reduce round-trip latency when an agent needs to invoke multiple tools before proceeding. Schema-aligned JSON output removes the post-processing step that causes failures in structured pipelines.

The model is OpenAI API-compatible, using the identifier mercury-2.5, which means most existing integrations can adopt it by changing a model name and endpoint.

Who Is Building Mercury and Why

Inception was founded by Stefano Ermon (Stanford), Aditya Grover (UCLA), and Volodymyr Kuleshov (Cornell), the researchers behind the first published diffusion LLM. The founding team’s academic lineage matters because Mercury is not a fine-tune of a larger model: it is a ground-up implementation of a different generation architecture. That distinction determines what is improvable with standard scaling and what requires fresh research.

The company is backed by Menlo Ventures, Mayfield, Innovation Endeavors, M12 (Microsoft’s venture fund), Snowflake Ventures, and Databricks Ventures, with individual investors including Andrew Ng, Andrej Karpathy, and Eric Schmidt. That investor roster is notable because Snowflake Ventures and Databricks Ventures are strategic: both companies run enterprise data and AI platforms that would benefit from faster, cheaper inference inside their own products.

Inception says enterprise Mercury usage has grown by more than an order of magnitude since Mercury 2 launched, and that thousands of developers and dozens of enterprises are running production workloads on the model.

Production Evidence: Augment Code and OpenCall

Inception provided two named production deployments at launch.

Augment Code moved its context compaction, model routing, and MCP tool search to Mercury, and reports an 82% reduction in context-compaction latency (from roughly 150 seconds to 27 seconds) and a 90% reduction in cost while maintaining output quality. Tool search summaries returned in under one second. These are Augment’s figures rather than independently audited results, but the specificity is useful: 150 seconds to 27 seconds is a concrete claim, not a percentage improvement without context.

OpenCall runs AI phone agents on Mercury. Voice agents have the hardest latency constraints of any text-generating workload: a response that arrives after 400 milliseconds breaks conversational naturalness. Mercury’s parallel generation architecture, which produces a first token faster than autoregressive models at equivalent quality, is a structural advantage for this use case.

These deployments illustrate where diffusion inference changes the economics most: not in one-off completion tasks, but in pipelines where latency and cost compound.

What This Means for Enterprise AI Procurement

The ongoing repricing of enterprise AI inference has accelerated in 2026, but most of the price movement has come from autoregressive models competing on the same architectural footing. Mercury 2.5 introduces a different variable: architectural throughput as a competitive advantage.

For enterprise buyers, the practical evaluation question is not whether Mercury 2.5 benchmarks at GPT-6 Astra quality (it does not: it benchmarks at the Haiku/Flash-Lite tier). The question is whether the workloads they need to run are quality-limited or speed-and-cost-limited.

Search retrieval, context compaction, tool-call summarization, document parsing, agent routing, and classification tasks often do not require frontier reasoning depth. Running them at 1,107 tokens per second for $0.04 per million tokens instead of frontier models at 200 tokens per second for $10 per million tokens is not a quality trade-off for those tasks. It is the correct model selection.

Enterprise teams architecting multi-agent systems should route by task type. Heavy reasoning tasks (contract analysis, complex code generation, multi-step planning) go to frontier models. High-volume, latency-sensitive utility tasks (summarization, routing, extraction, compaction) are candidates for a fast, cost-optimized model like Mercury 2.5. That routing architecture is increasingly the difference between agentic AI that is cost-effective in production and agentic AI that is cost-effective only in demos.

Inception also previewed two additional products at launch. Mercury Voice targets under 170 milliseconds time to first token for voice agent workloads. Mercury Router would handle cross-model routing across open and closed model catalogs. Both are in preview without general availability dates.

The company also confirmed it has started training its next model, described as larger than Mercury 2.5, targeting release in the coming months.

Evaluating Mercury 2.5 for Your Stack

Mercury 2.5 is available now via Inception’s API, Baseten, and OpenRouter. Enterprise deployments support dedicated capacity, autoscaling, compliance controls, and configurable data retention. Inception is offering 100 million free tokens to new developers through its API portal.

The 80% launch discount makes September the lowest-cost window to evaluate the model under production traffic. The key evaluation metric is not tokens per second in isolation but completed-task latency and cost per successful outcome across the specific pipeline tasks being considered.

For enterprise teams building agent infrastructure who want to understand how model selection, routing, and deployment architecture interact with their GTM and operations workflows, Enera works with AI-native enterprises to design and deploy those systems.

Sources: Inception Labs blog, September 8 2026 | BusinessWire press release | RuntimeWire analysis | ReasonCore coverage