Dekrov InsightsCharts / AI Benchmarks
GitHub ↗
AI Models · Research / Oct 21, 2026

Mercury 3:
Speed vs Evidence

Inception's new diffusion language model promises unusually fast output at low launch pricing. Here's what is published, what is measured, and what remains uncertain.

1,068Tokens/sec · vendor claim
1MContext tokens
100MFree API tokens · new users

01 / Generation speed

Inception-reported throughput on NVIDIA GPUs. This is not a head-to-head independent benchmark.

Mercury 3 · output tokens per second

Mercury 31,068
No directly comparable, independently verified speed numbers for GPT-6 Luna or Claude Haiku 5.5 are included here. The vendor says Mercury 3 runs 4–8× faster, but benchmark settings and methodology matter.

02 / API pricing

USD per million tokens. Compare the same token category; lower is cheaper. Launch prices are temporary and have no published end date.

Input price · launch vs standard

Launch$0.10
Standard$0.20

Output price · launch vs standard

Launch$0.35
Standard$0.70
CategoryLaunchStandard
Input / 1M$0.10$0.20
Cached input / 1M$0.01$0.02
Output / 1M$0.35$0.70

03 / Published quality results

Medium-reasoning scores attributed to Inception by BenchLM. Different benchmarks measure different abilities: do not treat the bar lengths as one unified ranking.

GPQA Diamond · knowledge/reasoning

Score89.23

IFBench · instruction following

Score76.84

DeepSWE v1 · software engineering

Score46.31

BrowseComp+ · web research

Score55.18

Additional reported results: MultiChallenge 60.35, MRCR v2 75.50, tau3-bench Banking 30.24, DeepSearchQA 71.33. Exact evaluation configurations and independent replication remain limited.

Why diffusion matters

Traditional autoregressive language models produce tokens sequentially. Diffusion language models iteratively refine multiple tokens in parallel. This can reduce generation latency, but speed alone does not guarantee correct code, robust tool use, or faster end-to-end agents.

What is actually available?

Mercury 3 is offered through Inception's API and announced via OpenRouter, Baseten and Vercel. The API is OpenAI-compatible. The model is proprietary: public weights are not available. The 100M-token offer applies to new accounts and is subject to provider terms.

What I'd test next

Run the same real coding-agent tasks through Mercury 3, DeepSeek V4.1 Flash and GLM-5.3 Flash. Record task success, first-token latency, end-to-end completion time, tool-call reliability and total billed tokens. Don't infer any of those results from vendor throughput alone.

Sources

Evidence policy: no invented cross-model bars. Performance, price, and quality charts use different scales and are labeled separately. Last researched Oct 21, 2026.