Evaluate Inception Mercury 3 for Faster, Cost-Effective Agentic Workloads
Inception AI’s new diffusion model claims superior speed and coding scores against major rivals, offering system admins a potential upgrade path for latency-sensitive applications.
Key points
- Inception AI released Mercury 3, claiming it outperforms GPT-6 Luna and Claude Haiku 5.5 on eight of twelve agentic benchmarks.
- The model reportedly achieves inference speeds four to eight times faster than competitors, reaching 1,068 tokens per second in internal tests.
- Pricing is set at 20 cents per million input tokens, with a temporary launch promotion halving those rates for early adopters.
Inception AI has launched Mercury 3, a new large language model that the company says delivers significantly faster response times than current industry leaders. System administrators evaluating generative AI tools should note that this release targets agentic workloads, where automated agents perform complex tasks rather than simple chat interactions. The vendor claims performance advantages in both speed and specific coding benchmarks compared to established models from OpenAI and Anthropic.
In plain English
Most large language models generate text sequentially, predicting one word at a time before moving to the next. This autoregressive approach can become a bottleneck for long or complex responses. Inception uses a different method called diffusion. Instead of writing linearly, the model drafts the entire response structure simultaneously. It then refines every part of the answer in parallel passes. This parallel processing allows Mercury 3 to complete tasks much quicker than models that wait for each token to finish before starting the next.
The background
According to SiliconANGLE, Inception AI Inc. released Mercury 3 as its new flagship model. The company conducted its own benchmarks comparing Mercury 3 against OpenAI Group PBC’s GPT-6 Luna and Anthropic PBC’s Claude Haiku 5.5. In these tests, Mercury 3 won eight out of twelve evaluations. The speed difference was substantial, with Mercury 3 operating at four to eight times the rate of its competitors. Co-founder Stefano Ermon stated that the primary challenge for diffusion models is scaling compute and data, not fundamental algorithmic limits.
Performance metrics highlighted in the release show Mercury 3 achieving 1,068 tokens per second on Nvidia Corp. graphics processing units. By comparison, Haiku 5.5 reached about 240 tokens per second, and GPT-6 Luna hit approximately 130. Inception did not disclose the parameter count for Mercury 3, but noted it is more than 40 times larger than Mercury 1. The context window also expanded to 1 million tokens, up from 260,000 in the previous version.
Benchmark results showed mixed outcomes across different tasks. Mercury 3 scored 51.1 on the SWE-Bench Atlas coding test, outperforming GPT-6 Luna’s 33.3 and Haiku 5.5’s 32.3. However, Haiku 5.5 led on the DeepSWE coding benchmark with 51.6 to Mercury 3’s 46.3. GPT-6 Luna won two additional tests by narrow margins. No independent third-party group has published verification of these results yet, according to Techmeme.
What changes now
Early customers are already integrating Mercury 3 into products where latency is critical. Kuan Zou, CEO of Octen AI, reported that median generation time for real-time search on smart glasses dropped to 0.9 seconds. Michael Tong, CTO of RevisionDojo, noted that tutoring hints and marked answers now return in about one second. These improvements address user frustration with delays, which Zou described as feeling "broken" when answers take five seconds.
Pricing for Mercury 3 is listed at 20 cents per million input tokens and 70 cents per million output tokens. Cached inputs are 90% cheaper. A launch promotion currently halves both rates. This pricing compares to GPT-6 Luna and Haiku 5.5, which list at 10 cents and 50 cents respectively for typical prompts. Inception has also confirmed that training for Mercury 3’s successor is already underway.
Key terms explained
When evaluating new AI infrastructure, teams usually focus on inference latency and throughput. Inference latency is the delay between sending a prompt and receiving the first character. Throughput measures how many tokens the model processes per second. Diffusion models attempt to improve both by generating text in parallel rather than sequentially. Administrators should check whether their current applications suffer from sequential bottlenecks. If users experience noticeable lag during complex queries, a parallel-processing model might reduce wait times. Teams also need to verify that the new model supports their required context window size.
What to do and how to stay safe: Inception AI
- Review current application latency metrics to determine if slower inference times are impacting user experience or operational efficiency.
- Test the new model with your specific agentic workloads to verify that claimed speed improvements translate to your unique data patterns.
- Compare total cost of ownership, including input, output, and caching fees, against your existing provider contracts to assess financial impact.
- Monitor for independent benchmark results from third-party auditors before committing to large-scale production deployments of unverified performance claims.
Step-by-step guide: Stop AI Chatbots Leaking Sensitive Business Data
General security guidance from the Vector Update newsroom. It is not confirmed advice from the organisations named in this story.
Frequently asked questions
How does Mercury 3 compare to GPT-6 Luna and Claude Haiku 5.5?
According to Inception’s internal benchmarks, Mercury 3 beat both models on eight of twelve tests and ran four to eight times faster.
What is the pricing for Mercury 3?
The model lists at 20 cents per million input tokens and 70 cents per million output tokens, with a temporary launch promotion halving these rates.
Has anyone independently verified Inception's performance claims?
No, the source material states that no outside group has published its own results on the model yet.
