Mercury 3 claims 1,068 tokens/sec via diffusion, but real-world OpenRouter tests show ~160 tok/s and it loses to Claude Haiku 5.5 on DeepSWE coding (46 vs 52).
Vamsi Tallapudi
Manager, Architect Technology at Cognizant
Mercury 3 is a diffusion-based large language model from Inception Labs that generates text at a claimed 1,068 tokens per second by refining an entire response in parallel instead of writing it one token at a time. Inception says it beats GPT-6 Luna and Claude Haiku 5.5 on several benchmarks while running 4 to 8 times faster, but it loses to Haiku 5.5 on real coding benchmarks like DeepSWE, and early testers on OpenRouter report real-world speeds closer to 160 tokens per second.
That gap between the headline number and what people actually see matters more than the headline itself. Let's get into what Mercury 3 actually is, where the claims hold up, and where diffusion models are genuinely useful for coding versus where they're still a worse bet than a normal autoregressive model.
Mercury 3, released October 19, 2026, is Inception Labs' flagship diffusion language model (dLLM). Every other mainstream chat model you've used (Claude, GPT, Gemini) is autoregressive: it predicts the next token, then the next, then the next, each one conditioned on everything before it. That's why streaming responses trickle out word by word.
Diffusion models don't work that way. Mercury starts with a noisy draft of the whole response and refines it over a handful of passes, denoising the entire sequence at once, closer to how image diffusion models like Stable Diffusion build a picture than how a chatbot normally writes a sentence. Fewer sequential steps means fewer round trips through the model, which is the entire source of the speed claim.
Inception isn't new at this. Mercury 1 shipped in June 2025 as, by their own description, the first commercial-scale diffusion LLM. Mercury 2.5 added adjustable reasoning effort and a 260K context window. Mercury 3 is the fourth release in that line, and Inception says it's more than 40 times larger than Mercury 1 in total parameters, with roughly 3x the aggregate intelligence on standard benchmarks while keeping similar throughput.
Here's the number everyone's repeating: 1,068 tokens per second on Nvidia GPUs, which Inception frames as 4 to 8x faster than GPT-6 Luna and Claude Haiku 5.5. That number comes straight from Inception's own announcement. Artificial Analysis hadn't independently verified it as of launch, and nobody outside the company has published the hardware config or batch size behind it.
And once people started actually running it, the picture got messier. On the Mercury 3 Hacker News thread, one commenter reported building a SwiftUI app in 45 minutes, processing 26 million tokens (at 160 tokens per second through OpenRouter, not 1,068). Another flagged bad latency on the same route. That's roughly a 6-7x gap between the marketing number and what a third party measured on a public API.
This isn't automatically dishonest. Diffusion models are known to have batch-size scaling quirks (Google's own diffusion LLM research ran into similar issues), so a number measured on dedicated hardware at a specific batch size can look very different from what you get hitting a shared inference endpoint. But if you're planning capacity around "4-8x faster," test it on your own traffic pattern first. Don't take the slide at face value.
| Mercury 3 | Claude Haiku 5.5 | |
|---|---|---|
| Architecture | Diffusion (parallel denoising) | Autoregressive |
| Claimed speed | 1,068 tok/sec (company number) | Anthropic's fastest standard-speed model |
| Context window | 1M tokens | Standard Claude context tiers |
| Input price | $0.10/M (launch promo, $0.20 list) | $0.10/M (up to 100K tokens) |
| Output price | $0.35/M (launch promo, $0.70 list) | $0.50/M (up to 100K), jumps to $2.50/M past 100K |
| DeepSWE v1 (coding) | 46.31 | 51.62 |
| Cached input | $0.01/M (promo) | $0.01/M |
A few things jump out. Input pricing is basically a tie at launch promo rates. Output is cheaper on Mercury 3 at the promotional price, though once the promo ends the list price ($0.70/M) is actually higher than Haiku 5.5's sub-100K rate. And Haiku 5.5 has its own pricing trap worth knowing about: it's a flat rate under 100K tokens, then a 5x cliff on every token in the request once you cross that line. I covered that cliff in detail in the Claude Haiku 5.5 breakdown if you're choosing between the two for a cost-sensitive pipeline.
On quality, Inception claims wins over Haiku 5.5 on IFBench, GPQA Diamond, SWE-Bench Atlas, and BrowseComp+. But on DeepSWE v1, a benchmark that actually measures software-engineering task completion, Haiku 5.5 wins outright: 51.62 to Mercury 3's 46.31. So "Mercury 3 beats Haiku 5.5" is true on some benchmarks and false on the one that looks most like real coding work. Read the fine print before you believe the headline claim either way.
One more wrinkle: Mercury 3 defaults to "low" reasoning effort, down from "medium" on the prior model. If you're benchmarking it yourself and comparing against someone else's numbers, make sure you're both running the same reasoning setting, or you're not comparing the same model.
Honestly, I think diffusion LLMs are a real tool for a specific job, not a general replacement for autoregressive models in coding work. Here's where I'd reach for one and where I wouldn't.
Good fit:
Bad fit:
The honest summary: Mercury 3 is a legitimate step forward for diffusion LLMs at scale, and the architecture bet (parallel denoising instead of token-by-token generation) is interesting enough that I expect more labs to try it. But for actual software engineering agent work today, Claude Haiku 5.5 still wins on the one benchmark that looks most like the job. Pick Mercury 3 for the use cases where raw latency is the bottleneck and a human is in the loop, not as a drop-in swap for your coding agent's backbone model.
If you're tracking the broader fast/cheap-model race, I've also written up MiniMax M3.1 Flash and DeepSeek V4.1 Flash, both chasing the same low-latency niche from the autoregressive side. And if you want the full state of play across every major model release this year, the 2026 AI tools roundup has the rest.
Mercury 3 is live on the Inception API Platform and through OpenRouter, Baseten, and Vercel, and it's OpenAI-API compatible, so swapping it into an existing integration is mostly a base-URL and model-name change. Launch pricing ($0.10/M input, $0.35/M output, $0.01/M cached input) runs at roughly half the list price, so lock in your testing now if you want to benchmark it against your own workload before the promo ends.
Direct answers below for the most common questions about Mercury 3.
New AI tools, automation workflows, and course drops — straight to your inbox. Join 2,400+ builders.
Mistral's Timestral forecasts time series on a CPU in 10-140ms, no GPU. How to install and run it in Python (5M-209M models), plus the license's $20M catch.
MiniMax M3.1 Flash costs $0.15/M input tokens and turns out to be the stealth "Space Bunny" model. Here's the pricing, benchmarks, and how it compares to Claude Haiku 5.5.
HybridAgent lets LLM agents think in latent vectors and act in text, cutting tokens up to 54.7% and cost up to 3.7x. Here's what it means for agent builders.
Comments