← Back to feed

Inception Labs released Mercury 2.5 on September 8, completing the most consequential upgrade to its diffusion LLM platform since the architecture went public. The model runs at 1,107 tokens per second on standard NVIDIA GPUs — unchanged from the preview — but the intelligence jump is the story: Inception claims a 40% gain over Mercury 2, bringing quality up to cost-optimized frontier models including GPT-5.6 Luna (Low), Gemini 3.5 Flash-Lite, and Claude Haiku 4.5.

Numbers

• Speed: 1,107 tokens/sec on widely available NVIDIA GPUs
• Context: 260K tokens
• Price: $0.20/M input, $0.75/M output (launch promotion: $0.04/$0.15 — 80% off)
• Intelligence: Self-reported 40% gain over Mercury 2
• Capabilities: Tunable reasoning, parallel tool calls, schema-aligned JSON output

The 80% launch discount puts Mercury 2.5 among the cheapest capable models in production. At $0.04/M input, a developer could run 250 million input tokens for $10 — the economics designed for search agents that need dozens of calls per query.

Production signals over benchmarks

Inception built Mercury 2.5 training from production failure data rather than benchmark optimization. Thousands of developers and dozens of enterprises deployed Mercury 2, generating a feedback loop that identified where the model broke down at latency boundaries.

Two case studies illustrate the target workloads:

Voice: OpenCall builds AI phone agents. On Mercury, its P99 response time dropped from several minutes to one second. P50 fell from 400ms to under 200ms. Mercury Voice, announced in preview alongside Mercury 2.5, targets under 170ms time-to-first-token — a threshold where latency becomes audible as hesitation on a live call.

Coding subagents: Augment Code runs Mercury for context compaction, model routing, and MCP tool search within a larger coding agent stack. Compaction latency dropped 82% (from 150 seconds to 27 seconds) and cost fell 90% while quality held. Tool-search summaries now return in under a second.

Mercury Voice and Mercury Router

Both arrived in preview alongside the main model.

Mercury Voice is a separate serving configuration optimized for audio pipelines. The TTFT target is sub-170ms — tight enough for real-time call agents where standard transformer models create noticeable pauses. The underlying architecture is the same diffusion backbone; the optimization is in how tokens are streamed to speech synthesis.

Mercury Router addresses orchestration: routing requests across the Mercury family (and potentially other models) based on task type, latency requirements, and cost. The announcement did not detail routing logic, but the pairing with Mercury Voice suggests a system where latency-sensitive tasks get routed to Voice while higher-quality tasks go to Mercury 2.5.

Architecture position

Inception describes Mercury 2.5 as the largest diffusion language model ever trained. The claim is consistent with what Artificial Analysis tracked on Mercury 2 — it led output speed rankings at launch and held the position until Celeris-1 entered at 1,664 tokens/sec in August. Mercury 2.5 does not reclaim the outright speed record but pairs 1,107 tokens/sec with a meaningful intelligence step up.

The competitive set Inception names — Luna Low, Gemini 3.5 Flash-Lite, Haiku 4.5 — positions Mercury 2.5 as a latency play, not a frontier intelligence play. The bet is that a large segment of production workloads do not need GPT-5.6 Sol or Claude Fable 5.1. They need calls that complete inside a user interaction window, at costs that scale to billions of tokens per day.