Inception Mercury 3: Release, 1,068 t/s, promotional price

Key Takeaways

1Mercury 3, a diffusion language model by Inception, is launched with a temporary discount and an OpenAI-compatible API

2The company claims 1,068 tokens/s and eight wins out of twelve benchmarks against GPT-6 Luna and Claude Haiku 5.5

3Early access partners report reduced latencies, but no independent validation is available

Why it matters : Mercury 3 targets real-time applications where generation speed is crucial, but its performance remains to be confirmed outside of internal testing.

Inception launches Mercury 3, a diffusion language model touted as significantly faster than its competitors. The price is temporarily halved, and availability spans multiple platforms, but performance gains are based on unverified internal measurements.

Launch Pricing and Multi-Platform Access

The standard price for Mercury 3 is set at $0.20 per million input tokens and $0.70 per million output tokens. Inception is applying a 50% discount at launch, without specifying an end date for this promotion. Once the promotion ends, the output cost would exceed that of GPT-6 Luna and Claude Haiku 5.5 for short inputs, with both models charging $0.50. For very long queries, these competitors become significantly more expensive, while Mercury 3 accepts up to one million tokens at a time. The model is accessible on Inception's platform as well as through OpenRouter, Baseten, and Vercel. It uses an API compatible with OpenAI's, and Inception claims that migration would require minimal effort from developers.

Claimed Speed and Comparisons Across 12 Benchmarks

Inception presents Mercury 3 as markedly faster than standard models and describes it as the most performant diffusion language model released. According to its internal measurements, Mercury 3 exceeds 1,000 tokens per second on Nvidia GPUs and outperforms GPT-6 Luna and Claude Haiku 5.5 in eight of the twelve tests conducted. The company claims a speed four to eight times greater, with a throughput measured at 1,068 tokens per second, compared to approximately 240 for Claude Haiku 5.5 and 130 for GPT-6 Luna. Results vary by task: the gap is widest in instruction-following, with Mercury 3 scoring 76.84 points on IFBench, 15 points higher than GPT-6 Luna, and a clear lead on SWE-Bench Atlas (51.08 versus 33.33 and 32.26). Conversely, Claude Haiku 5.5 excels in DeepSWE and for information retrieval in long texts, while GPT-6 Luna takes the lead in two Artificial Analysis tests. Inception summarizes the quality of Mercury 3 as equivalent or better. No independent tests have yet confirmed these figures.

Le brief IA que lisent les pros

Le brief IA que les pros lisent chaque soir

Les 7 actus IA du jour, décryptées en 5 min. Gratuit.

Inclus dès l'inscription : notre sélection des meilleurs guides & comparatifs IA.

Choisissez votre rythme

Gratuit · Pas de spam · Désabonnement en 1 clic

Real-Time Uses and Early Access Feedback

Mercury 3 is primarily aimed at real-time applications such as AI agents, search, customer support, and tutoring, where delays accumulate at each step. Early access partners report latency-focused improvements. Octen AI, which develops search functions for smart glasses, finds that a five-second response feels inadequate; the search takes only about 80 milliseconds, and the language model was the main bottleneck. With Mercury 3, the median response time would be 0.9 seconds and the complete process 2.2 seconds. General Learning indicates it provides hints and corrections in about one second. Lorikeet reports faster responses and better outcomes. At TinyFish AI, Shuhao Zhang mentions 50 navigation tasks on real sites, with 25 solved by Mercury 3, ahead of GPT-6.1-Sol, with no figures provided for the latter.

Technical Choices and the Mercury Line's Trajectory

Mercury 3 is based on a diffusion generation that starts with a broad response and then refines it through multiple passes, unlike models that generate word by word. Inception claims that Mercury 3 has more than 40 times the parameters of Mercury 1 while maintaining the same speed, and that performance has tripled compared to Mercury 1 mini. The company believes that diffusion models progress reliably as they increase in size, but does not disclose the exact number of parameters. In February, Mercury 2 was introduced as the first diffusion model capable of reasoning through intermediate steps. Prior to this launch, Inception raised $50 million from NVentures, M12, and Snowflake Ventures, the funds of Nvidia, Microsoft, and Snowflake.

A Field Where Google is Also Advancing with Diffusion

Google is also developing diffusion language models. After a demonstration of Gemini Diffusion, the company released DiffusionGemma in open source in June. Google claims that this model operates up to four times faster than standard Gemma 4, with slightly lower results.