Mercury 3 Explained: Inception's 1,068-Tokens-Per-Second Diffusion LLM
Inception launched Mercury 3 on October 19, 2026, a diffusion LLM that generates text in parallel at a claimed 1,068 tokens per second. This guide explains what it changes, what the benchmark claims really say, and how businesses can integrate it.
The Short Answer
Mercury 3 is the new flagship diffusion language model from Inception, launched on October 19, 2026, and it is making one of the boldest claims the AI industry has seen this year: a model that generates text at a company-reported 1,068 tokens per second while matching the quality of frontier speed-optimized models like OpenAI's GPT-6…
Mercury 3 is the new flagship diffusion language model from Inception, launched on October 19, 2026, and it is making one of the boldest claims the AI industry has seen this year: a model that generates text at a company-reported 1,068 tokens per second while matching the quality of frontier speed-optimized models like OpenAI's GPT-6 Luna and Anthropic's Claude Haiku 5.5. If you run anything that depends on fast AI responses (voice agents, browser automation, real-time search, or customer support copilots), this launch deserves your attention.
In this explainer, we break down what a diffusion LLM actually is, what Mercury 3 changes, what the benchmark claims say (and how to read them carefully), the business use cases where speed matters, and how developers can start integrating it today.
What Is a Diffusion LLM? Mercury 3 in Plain English
Almost every well-known language model, from GPT-6 to Claude to Gemini, writes text the same way: one token at a time, left to right. Each new word depends on everything that came before it. This is called autoregressive generation, and it is reliable but slow by design, because the model cannot write word six until it has finished word five. No matter how fast the hardware gets, that sequential step-by-step process is a built-in speed limit.
Mercury 3 takes a different path. It is a diffusion language model, a technique borrowed from image generation tools like Midjourney. Instead of writing left to right, a diffusion model starts with a rough, noisy draft of the entire response and then refines every part of it in parallel across a handful of passes, the same coarse-to-fine process that turns static into a photograph. Because the model works on many token positions at once rather than queuing them in sequence, it produces complete responses dramatically faster.
The trade-off, historically, is that diffusion was the fast but less capable sidekick. Early diffusion language models were cheaper and quicker than autoregressive rivals but could not match them on quality. Inception, founded by Stanford's Stefano Ermon alongside Volodymyr Kuleshov and Aditya Grover, has spent the last year building the commercial case that diffusion can scale. Mercury 1 (2025) proved diffusion LLMs could be commercially useful, Mercury 2.5 added tunable reasoning and a 260,000-token context window, and Mercury 3 is the company's argument that diffusion scaling laws match those of autoregressive models.
Mercury 3: What Actually Changed
The headline number is 1,068 tokens per second on NVIDIA GPUs, a company-reported figure that Inception says is four to eight times faster than the rival models it compared against (reportedly around 240 tokens/sec for Claude Haiku 5.5 and 130 for GPT-6 Luna). For perspective, a 500-word response that takes a conventional speed-optimized model several seconds could complete in well under a second with Mercury 3.
Here are the key specs, all from Inception's launch announcement:
- 1M-token context window: up from 260,000 tokens in Mercury 2.5, giving developers roughly four times the room for long documents, extended conversations, and agent workflows. The maximum output length is reported at 131,072 tokens.
- 40x larger than Mercury 1: Inception says Mercury 3 carries more than 40 times the total parameters of the original Mercury 1, at comparable speed. The exact parameter count was not disclosed.
- 3x intelligence improvement: the company claims roughly three times better aggregate intelligence on its internal benchmark suite versus Mercury 1 (mini).
- Tunable reasoning effort: four settings (instant, low, medium, high). Note that Mercury 3 defaults to low effort when the parameter is omitted, whereas Mercury 2.5 defaulted to medium, so pin the effort level explicitly when benchmarking or comparing.
- OpenAI-compatible API: the docs use the model string "mercury-3" with temperature adjustable between 0.5 and 1.0, and tool calling plus structured outputs are supported on the chat completions endpoint.
Access is API-only through Inception's own platform, OpenRouter, Baseten, and Vercel. There is no downloadable open-weights release. New API accounts get 100 million free tokens, and launch pricing is 50 percent off list: $0.10 per million input tokens and $0.35 per million output tokens (standard $0.20/$0.70), with cached input at roughly $0.01 per million.
Inception also ships two sibling models alongside Mercury 3: Mercury Voice, optimized for time-to-first-answer in voice agents, and Mercury Decide, a classification-style model for structured decisions. Existing Mercury 2 and 2.5 users should note that those models are now closed to new users, with existing non-enterprise users given until November 15, 2026 to migrate.
Mercury 3 Benchmark Claims: Reading Them Carefully
Inception's central quality claim is that Mercury 3 is "on par or better" than GPT-6 Luna and Claude Haiku 5.5 across a range of benchmarks. These are company-reported, vendor-run benchmarks, not independent evaluations, and they should be read with that in mind.
According to coverage of Inception's own benchmark table, Mercury 3 reportedly leads 8 of 12 tests against GPT-6 Luna and Claude Haiku 5.5, with the biggest reported wins on SWE-Bench Atlas (51.08) and IFBench (76.84), plus GPQA Diamond. Where it reportedly trails: DeepSWE v1 at 46.31, about 5.31 points behind Haiku 5.5's 51.62, and MRCR v2, where Haiku 5.5 also still wins.
Early customer results shared by Inception hint at where the speed matters most: TinyFish AI says Mercury 3 completed half of 50 live Mind2Web browser-navigation tasks in its production setup, and Octen AI reported a median generation time of 0.9 seconds and a median end-to-end response of 2.2 seconds. Because these results were shared by Inception, they are not independent comparisons, but they point at agentic loops, where a model is called again and again to complete one task and every millisecond of latency multiplies.
One caution flagged by analysts: Inception's comparison claims do not state which reasoning-effort setting produced them. A model answering at "low" effort by default will look faster and may score lower than the same model at "high". If you evaluate Mercury 3 yourself, pin the effort level explicitly, or your numbers will not be comparable to anyone else's.
Why Mercury 3 Matters for Business Use Cases
Speed is not just a nicer user experience. For many AI products, latency decides whether the product is usable at all. Inception positions Mercury 3 squarely at five areas, and each maps to a real business scenario:
- Browser-use agents: autonomous agents that click through websites need to make dozens of model calls per task. At 1,000+ tokens per second, multi-step web automation stops feeling broken. This is one of the strongest fits for diffusion LLMs, and it connects to the broader wave of always-on AI agents for business.
- Voice pipelines: in phone-based support or sales, a five-second pause feels broken. Sub-second generation makes real-time conversation possible.
- Real-time search agents: tools that search, read, and summarize on the fly, like Octen AI's smart-glasses search with its 0.9-second median generation time.
- Customer support copilots: support agents that draft replies, look up policies, and suggest next actions while the customer is still on the line.
- Coding subagents: the SWE-Bench Atlas result (reported 51.08) suggests Mercury 3 is credible for agentic coding loops where a coding agent iterates rapidly through draft, test, and fix cycles.
Companies exploring this space should also watch how Mercury 3 compares to other speed-optimized models in the segment. Our Claude Haiku 5.5 breakdown covers the other leading low-latency contender, which still leads Mercury 3 on DeepSWE v1 and MRCR v2 according to Inception's own numbers.
How Businesses Can Start Integrating Mercury 3
The good news for engineering teams is that integration is straightforward. Because the API is OpenAI-compatible, teams already using that standard can typically point their existing client libraries at the Inception endpoint (or at OpenRouter, Baseten, or Vercel) and change the model string to "mercury-3", with no rewrite of their application code. Structured outputs and tool calling work through the standard chat completions shape, so agentic workflows built on function calling patterns carry over.
A sensible rollout path looks like this:
- Start with the 100M free tokens to benchmark Mercury 3 on your own workloads, pinning the reasoning_effort level explicitly so results are reproducible.
- Pilot one latency-sensitive path first, such as voice responses or browser-agent subtasks, measuring end-to-end time to completion, not just tokens per second. Throughput is not the same as the time a user waits: network delay, prompt processing, and tool calls all add up.
- Compare against your current speed-tier model on accuracy, instruction following, and cost per completed task before migrating production traffic.
- Plan for the Mercury 2/2.5 migration deadline (November 15, 2026 for existing non-enterprise users) if you are already on Inception's platform.
If your team needs help designing that evaluation or wiring a new model into an existing product, our API development and integration services cover exactly this kind of work, from connector builds to agentic workflow design. For a broader grounding in integration strategy, see our complete 2026 API integration guide.
Mercury 3 vs Speed-Optimized Alternatives
Mercury 3 does not exist in a vacuum. The two models Inception chose to compare against are the other leaders of the speed tier:
| Model | Reported speed (tokens/sec) | Context window | Notable strength (vendor claims) |
|---|---|---|---|
| Mercury 3 | 1,068 (Inception claim, NVIDIA GPUs) | 1M tokens | Leads 8 of 12 vendor-run benchmarks; strongest on SWE-Bench Atlas |
| GPT-6 Luna | ~130 (Inception comparison) | Not disclosed in launch coverage | Frontier OpenAI quality bar; broad ecosystem |
| Claude Haiku 5.5 | ~240 (Inception comparison) | Not disclosed in launch coverage | Still leads on DeepSWE v1 and MRCR v2 per Inception's own table |
The honest reading: Mercury 3's advantage is throughput per dollar in latency-bound workloads. It is not claiming to be the smartest model in the world; it is claiming to be fast enough that agent loops, voice pipelines, and real-time search become practical, at quality comparable to rival speed-tier models. Businesses choosing between them should weigh ecosystem maturity against Mercury 3's raw speed edge.
Mercury 3 Limitations and Risks to Know
No launch explainer is complete without the caveats:
- Benchmarks are vendor-run. Every quality comparison cited here comes from Inception's own testing. Independent evaluation has not yet happened, and the effort-level setting behind the comparison claims was not disclosed.
- API-only, no open weights. You cannot self-host Mercury 3 or run it on your own infrastructure. That means vendor lock-in, data flowing through Inception's servers (check data-retention terms for your compliance needs), and dependence on their uptime.
- Text in, text out. The model handles text only. Multimodal products will need companion models for vision or audio.
- Launch pricing is promotional. The 50% discount is a current offer, not a permanent promise. Budget projections should use the $0.20/$0.70 standard rates for planning beyond the launch window.
- Early independent reviews note uneven constraint following. One hands-on reviewer found strong business math and solid explanations but sloppy adherence to creative constraints (word counts, forbidden words), a reminder to test instruction-following on your own tasks.
- Diffusion is still a young paradigm. The tooling, evals, and prompting best practices around diffusion LLMs are less mature than the autoregressive ecosystem.
Mercury 3 FAQ
What is Mercury 3?
Mercury 3 is a diffusion large language model from Inception, launched on October 19, 2026. Instead of generating text one token at a time, it refines a draft of the entire response in parallel, which Inception says produces output at 1,068 tokens per second on NVIDIA GPUs with a 1-million-token context window.
How fast is Mercury 3 compared to other AI models?
Inception reports 1,068 tokens per second, four to eight times faster than the rivals in its comparisons. These are vendor claims, not yet independently verified.
Does Mercury 3 beat GPT-6 Luna and Claude Haiku 5.5?
According to Inception's own benchmark table, Mercury 3 leads on 8 of 12 tests, with its biggest reported wins on SWE-Bench Atlas (51.08) and IFBench (76.84). However, Haiku 5.5 still wins DeepSWE v1 and MRCR v2 in that same table. All of these are company-run results, so treat them as directional rather than definitive.
How can developers access Mercury 3?
Mercury 3 is available through Inception's own API platform, plus OpenRouter, Baseten, and Vercel. The API is OpenAI-compatible, so existing OpenAI client code can switch by changing the model string to "mercury-3". New API accounts receive 100 million free tokens, and there is no downloadable open-weights version.
What is Mercury 3's context window?
One million tokens, up from 260,000 on Mercury 2.5, with a reported maximum output of 131,072 tokens. That gives developers roughly four times the room for long documents, extended conversations, and agent workflows.
How much does Mercury 3 cost?
Standard rates are $0.20 per million input tokens and $0.70 per million output tokens. A 50% launch discount currently brings those to $0.10 and $0.35, with cached input around $0.01 per million. The launch discount is promotional, so plan long-term budgets against the standard rates.
Mercury 3 is the clearest signal yet that diffusion language models are moving from research curiosity to commercial infrastructure. Whether the speed-to-quality claim holds up will be decided by independent testing and real production workloads over the coming months, and businesses with latency-bound AI products have good reason to run their own pilots now. If you want help evaluating Mercury 3, integrating it into your stack, or designing the agentic workflows it enables, Contact our team: our AI and machine learning specialists build production AI systems that turn new model capabilities into working products.
Related Articles
Keep reading — more guides you may find useful.
Related Services
Services connected to this guide.
