Inception Mercury 3: 1,068 Tok/s, 1M Context, 8 of 12 Benchmarks

Inception has released Mercury 3, the fourth release in its diffusion LLM family, claiming 1,068 tokens per second on NVIDIA GPUs and a 1M-token context window while leading 8 of 12 self-reported benchmarks against GPT-6 Luna and Claude Haiku 5.5. Unlike autoregressive models that emit one token per forward pass, a diffusion LLM starts from a rough draft of the full response and refines every position across a small number of parallel passes. For latency-sensitive voice, search, and agent workloads, that architectural difference is the core of the speed claim.

Inception reports that Mercury 3 carries more than 40x the parameters of Mercury 1 (mini), and that average quality across standard benchmarks rose about 3x between those two models while speed stayed above 1,000 tokens per second. The quality-versus-size curve spans five Mercury generations, but it is vendor-reported evidence, not independent verification.

Benchmark profile

All figures below come from Inception's benchmark table, with all models run at medium settings. These are self-reported results.

Mercury 3's largest margins appear in agentic coding and instruction following. SWE-Bench Atlas reaches 51.08 against 33.33 for GPT-6 Luna and 32.26 for Claude Haiku 5.5. IFBench reaches 76.84, about 15 points above Luna's 61.82. GPQA Diamond lands at 89.23, ahead of Luna's 86.36 and Haiku 5.5's 84.51.

The losses are instructive. Haiku 5.5 wins DeepSWE v1, 51.62 to 46.31, and MRCR v2, 78.40 to 75.50. Luna edges Mercury 3 on AA-Omniscience-Public (46.94 vs 46.44) and AA-LCR v1.1 (78.30 vs 77.67). Longer coding tasks and long-context retrieval remain relative weak spots.

Pricing and availability

Mercury 3 is live on the Inception API Platform, OpenRouter, Baseten, and Vercel via an OpenAI-compatible API using the mercury-3 model string and a reasoning_effort parameter. Inception lists $0.20 per 1M input tokens and $0.70 per 1M output tokens, halved at launch to $0.10 input and $0.35 output; cached input carries a 90% discount, and new accounts receive 100M free tokens. MarkTechPost's comparison table puts GPT-6 Luna (short context) and Claude Haiku 5.5 (prompts up to 100K) at $0.10 input and $0.50 output, so the launch output price undercuts both rivals, though the $0.70 list price changes that. AI Mastery's pricing data shows Gemini 3 Flash at $3.00 output and Mistral Small 4 at $0.60 output. Inception has not said when the launch discount ends. The release also includes Mercury Voice, aimed at time-to-first-answer-token in voice agents, and Mercury Decide, a classification model.

AI Mastery analysis

Three implementation details need attention before production evaluation. Parameter count is not disclosed, maximum output length is not disclosed, and input modalities are not disclosed—no image support is confirmed. For multimodal pipelines this is a hard constraint rather than a tradeoff. Second, MarkTechPost notes that OpenRouter showed a P50 throughput of 383 tokens per second on launch day, well below the 1,068 headline. Dedicated enterprise capacity may recover closer to Inception's figure, but teams on shared infrastructure should benchmark against the lower number, a pricing and capacity problem as much as a capability one. Third, the 1M context window is a specification, not a retrieval guarantee: MRCR v2 at 75.50 trails Haiku 5.5's 78.40.

The win pattern—strong on short agentic tasks, weaker on long sequential ones—is consistent with how parallel refinement works. Early-access partners TinyFish, Octen AI, RevisionDojo, and Lorikeet describe short-horizon deployments: web navigation, smart-glasses search, single-turn hints, and support routing. That aligns with bounded-output workloads replacing open-ended prose generation in production. Mercury 3 looks strongest where pipeline design, not raw model intelligence, sets the throughput economics. The broader signal is that architectural specificity may be starting to outperform raw GPU scaling at the inference layer—still an open experiment, but one worth benchmarking for teams building agent loops or voice interfaces.

Sources

• Reported by: MarkTechPost: Inception Releases Mercury 3 Diffusion LLM at 1,068 Tokens per Second
• Competitor prices: AI Mastery token calculator, figures as of 2026-09-14