Inception releases Mercury 3, a diffusion LLM with a 1M-token context that it says beats GPT-6 Luna and Claude Haiku 5.5 at 1,000+ tokens/s
Partly confirmed
The takeaway
On Oct 19, 2026 Inception released Mercury 3, its flagship diffusion language model.
Status
- Claim
Partly confirmed
- Our reporting
- Medium confidence
- Importance
- 3 of 5
- Last verified
- 20 October 2026
Your AI and this story
- GPT-6 Astra172 days after its cutoff
- Claude Opus 5.5111 days after its cutoff
- Gemini 4 Argon202 days after its cutoff
- Grok 4.7141 days after its cutoff
None of these four assistants can know about it. The closest, Claude Opus 5.5, stops 111 days before it.
Key facts
- API id: mercury-3 (Inception API, OpenAI-compatible; also OpenRouter as inception/mercury-3, Baseten and Vercel)
- Context 1M tokens (Mercury 2.5: 260K); max output 131,072 tokens; reasoning effort instant/low/medium/high (Inception docs)
- Speed: 1,068 tokens/s on Nvidia GPUs, per Inception; claimed 4–8x faster than GPT-6 Luna, Claude Haiku 5.5 and GLM-5.3 Flash
- Price: $0.20 input / $0.70 output per 1M tokens, 50% off at launch ($0.10 / $0.35); cached input 90% off. The Deep View: launch input price equals Haiku 5.5 and GPT-6 Luna, output 35% cheaper
- Claims without published scores: beats GPT-6 Luna and Claude Haiku 5.5 ‘across a range of benchmarks’; completed 50% of 50 live Mind2Web browser tasks, ‘significantly more than GPT-6.1-Sol’ (Sol’s score not given)
- Scale: over 40x the total parameters of Mercury 1 (Feb 2025); count not disclosed
- The Deep View: Inception says swarms of Mercury 3 agents were used for recursive self-improvement that raised speed a further 16%; Inception’s own blog post does not repeat this claim
- CEO Stefano Ermon (The Deep View): ‘Mercury 3 is the most capable diffusion-based language model on the market’; he said Inception is working on making the launch price permanent
What happened
Inception, the Stanford-founded startup behind the first commercial diffusion LLMs, released Mercury 3 on Oct 19, 2026. Instead of writing one token at a time, the model drafts and refines many tokens in parallel, which Inception says gives it over 1,000 tokens per second. The release adds a 1M-token context window and four reasoning-effort levels, and is sold at half price at launch.
Inception’s comparisons are claims: the blog post gives no benchmark table, and the only task figure (50% of 50 live Mind2Web tasks) has no matching number for the GPT-6.1 Sol model it is compared with. No independent measurement was available on Oct 20.
Why it matters
Diffusion LLMs are the main alternative to autoregressive decoding. If Mercury 3’s claims hold, a small-model tier of speed and price now comes with long context and agentic ability close to the cheap tiers of OpenAI and Anthropic.
Sources
3 sources from 3 sites. Numbers match the chips in the text.
3 sources: 2 primary, 1 press
Primary
- Inception: Diffusion at Scale: Introducing Mercury 3inceptionlabs.ai, official
- Inception docs: Models and pricing (mercury-3)docs.inceptionlabs.ai, docs
Press
- The Deep View: Why Inception’s Mercury 3 upends AI economicsthedeepview.com, press
Changes
- Filed (Techmeme / The Deep View; Inception blog and docs read)