Mercury 3 is the fastest large language model most developers have ever heard of — and as of this week, it is also one of the most debated. Launched on 19 October 2026, this diffusion-based LLM claims a staggering 1,068 tokens per second on NVIDIA GPUs, a one-million-token context window, and benchmark wins over both GPT-6 Luna and Claude Haiku 5.5. It does all this at launch pricing of just $0.10 per million input tokens. In this 2026 guide, we break down what Mercury 3 actually is, how diffusion language models work, where the performance claims hold up, and where they deserve a sceptical look.
Table of Contents
- What Is Mercury 3?
- How Diffusion Language Models Work
- Mercury 3 Benchmarks: vs GPT-6 Luna and Claude Haiku 5.5
- Mercury 3 Pricing (USD)
- How to Access Mercury 3: Availability
- Mercury 3 vs GPT-6 Luna: Which Should You Choose?
- Mercury 3 in the Real World: Early Partners
- Honest Caveats About Mercury 3
- FAQ About Mercury 3
- Final Verdict on Mercury 3

What Is Inception Mercury 3?
Mercury 3 is a general-purpose diffusion large language model (dLLM) built by Inception Labs, a Redwood City startup founded by Stanford professor Stefano Ermon alongside Volodymyr Kuleshov and Aditya Grover. It is the third generation of the Mercury family — Mercury 1 launched in June 2025, Mercury 2 followed with reasoning capabilities, Mercury 2.5 added tunable reasoning and a 260,000-token context window — and it is by far the biggest leap of the series.
The headline numbers tell the story. Mercury 3 generates 1,068 tokens per second on NVIDIA GPUs, according to the company’s own measurements. That is four to eight times faster than comparable speed-optimised rivals. It accepts up to 1 million tokens of context — roughly four times the 260K of Mercury 2.5 — and it is more than 40 times larger in total parameters than the original Mercury 1. Inception characterises the intelligence jump from Mercury 1 (mini) to Mercury 3 as a threefold rise across standard benchmarks.
But raw speed is only half the pitch. The other half is quality. Inception claims that Mercury 3 matches or beats frontier speed-optimised models — specifically OpenAI’s GPT-6 Luna and Anthropic’s Claude Haiku 5.5 — across a range of standard benchmarks. If that combination of speed and intelligence holds up under independent testing, it would be a genuine milestone: the AI industry has long treated speed and smarts as a trade-off you cannot escape.
There is a catch, and we will spend an entire section on it below: every benchmark number comes from Inception itself. None has yet been independently verified. Treat this guide as the full picture — the promise and the fine print.
How Diffusion Language Models Work
To understand why Mercury 3 is different, you need to understand the difference between diffusion and autoregressive generation. Nearly every major model you know — GPT-6, Claude, Gemini — is autoregressive. It writes text the way you read it: left to right, one token at a time. To produce the sixth word, the model must first finish the fifth. That sequential bottleneck is the fundamental reason even the fastest conventional models feel sluggish when generating long responses.
Mercury 3 takes a completely different route, borrowed from image generation tools like Midjourney and Stable Diffusion. Instead of writing one token at a time, a diffusion language model starts with a noisy draft of the entire response and then refines all of the tokens in parallel across a handful of steps. It is a coarse-to-fine process: the first pass produces a rough approximation of the whole answer, and each subsequent pass sharpens it until the final text is clean and coherent.
The practical consequence is parallelism. Where an autoregressive model performs hundreds of sequential steps for a long answer, a diffusion model can get there in a few parallel refinement passes. That is the entire basis of Mercury 3’s 1,068 tokens-per-second figure: the architecture was designed from the ground up for parallel generation.
Why does this matter now? The most important use cases for AI speed are not chatbots — they are agent loops. AI agents wait on the model at every single step: navigate to a page, read it, decide the next click, call a tool, repeat. A two-second delay per step compounds into a minute-long task across thirty steps. Halve the per-step latency and you do not just save time — you give the agent room to try more approaches before answering. That is why Inception is pitching Mercury 3 at voice assistants, coding copilots, search agents, and tutoring tools rather than at slow, deliberate reasoning tasks where bigger models still dominate.
The bigger philosophical claim is about scaling. When Inception launched Mercury 1, the open question in the AI research community was whether diffusion language models could scale to rival autoregressive models on complex tasks. Mercury 3 is the company’s answer: 40x more parameters than Mercury 1, with reported speed holding steady above 1,000 tokens per second. As co-founder Stefano Ermon put it, the limits to scaling diffusion are compute and data, not fundamental algorithmic barriers.

Inception Mercury 3 Benchmarks: vs GPT-6 Luna and Claude Haiku 5.5
Now the numbers. Inception compared Mercury 3 against GPT-6 Luna and Claude Haiku 5.5 — the two models that define the speed-optimised frontier — across 12 benchmark suites. The company reports that its model leads in 8 of 12, with its largest margins on instruction-following and coding tasks. Here are the headline figures, exactly as Inception reported them at medium reasoning settings:
| Benchmark | Mercury 3 | GPT-6 Luna | Claude Haiku 5.5 |
|---|---|---|---|
| IFBench (instruction following) | 76.84 | 61.82 | 54.65 |
| GPQA Diamond (hard science QA) | 89.23 | 86.36 | 84.51 |
| SWE-Bench Atlas (software engineering) | 51.08 | 33.33 | 32.26 |
| BrowseComp+ (web research) | 55.18 | 50.50 | 45.78 |
| DeepSWE v1 (deep coding) | 46.31 | — | 51.62 |
| MRCR v2 (long context) | — | — | 78.40 |
Read the table carefully, because it tells two stories at once. The first story is genuinely impressive: a 15-point lead on IFBench, a 19-point lead on GPQA Diamond, and a striking advantage on SWE-Bench Atlas, where Mercury 3 scores 51.08 against Luna’s 33.33 and Haiku’s 32.26. On agentic coding benchmarks, Inception’s own numbers place its model in a different league from its direct speed-tier rivals.
The second story is the one Inception’s marketing glosses over. On DeepSWE v1, a harder coding evaluation, Claude Haiku 5.5 leads with 51.62 against Mercury’s 46.31. On MRCR v2, a long-context retrieval test, Haiku scores 78.40 — and Inception does not appear to have beaten it. GPT-6 Luna, meanwhile, edges one of the long-context evaluations. So the honest summary is narrower than the headline: Mercury 3 leads on instruction-following, science reasoning, and web-research benchmarks, but it loses to Haiku on deep coding and long-context recall.
The Inception Mercury 3 Honest Scorecard
- Instruction following: Mercury 3 wins big (IFBench 76.84 vs 61.82 vs 54.65)
- Science reasoning: Mercury 3 leads (GPQA Diamond 89.23 vs 86.36 vs 84.51)
- Software engineering tasks: Mercury 3 dominates SWE-Bench Atlas (51.08 vs 33.33 vs 32.26)
- Web research: Mercury 3 edges ahead (BrowseComp+ 55.18 vs 50.50 vs 45.78)
- Deep coding: Claude Haiku 5.5 wins (DeepSWE v1 51.62 vs Mercury’s 46.31)
- Long-context recall: Claude Haiku 5.5 wins (MRCR v2 78.40)
One more critical detail: all of these scores were reported at medium reasoning settings and come entirely from Inception’s own testing. As of 20 October 2026, no independent third party has replicated them. That does not mean they are wrong — vendor benchmarks are usually directionally accurate — but benchmark-tuning is a real phenomenon, and a day-one announcement is the single most optimistic data point a company will ever publish. Expect independent evaluations to paint a more nuanced picture.
Inception Mercury 3 Pricing (USD)
Speed is only half of Mercury 3’s value proposition. The other half is price. The list rates are already competitive for the speed-optimised tier — but the launch discount makes them genuinely aggressive.
| Token type | List price (per 1M) | Launch price (per 1M) |
|---|---|---|
| Input | $0.20 | $0.10 |
| Output | $0.70 | $0.35 |
| Cached input | $0.02 (90% below input) | $0.01 |
A few things to note about the pricing structure:
- The 50% launch discount has no published end date. Inception’s pages simply say Mercury 3 is 50% off at launch. Company co-founder Stefano Ermon has suggested the lower price could even become permanent, but that is a hope, not a commitment. If you are budgeting a production deployment, plan around the $0.20/$0.70 list rates and treat the launch price as a bonus.
- Cached input is 90% cheaper than standard input. For agent loops that repeatedly process the same system prompts and conversation history — exactly the workloads Mercury 3 targets — prompt caching can cut your bill dramatically. During the launch window, cached input costs just $0.01 per million tokens.
- New API keys come with 100 million free tokens. No payment details are required. That is enough to run a serious evaluation or even a small pilot before spending a dollar.
For context, $0.10 per million input tokens puts launch-priced Mercury 3 among the cheapest capable models available to US developers — comparable to DeepSeek V4.1 Flash territory — while claiming benchmark quality well above that class. The pricing clearly aims to pull developers off the OpenAI and Anthropic speed tier.

How to Access Inception Mercury 3: Availability
Getting started with Mercury 3 is deliberately frictionless. The model is available today through four channels:
- The Inception API Platform. This is the direct route, with 100 million free tokens on every new API key. The docs identify the model string as
mercury-3, served from Inception’s chat completions endpoint, with parameters includingreasoning_effortand temperature. OfficialinceptionaiSDKs are available for Python and TypeScript. - OpenRouter. Mercury 3 is live on OpenRouter at launch pricing ($0.10/M input, $0.35/M output, $0.01/M cache read) with a 1M context window. You can check the live provider routing on OpenRouter’s Mercury 3 page.
- Baseten. Inception’s long-standing inference partner is serving the model with enterprise-grade throughput options.
- Vercel. The model works through the Vercel AI SDK, making it a drop-in option for Next.js developers building AI features.
The single most developer-friendly detail: the API is OpenAI-compatible. You can point your existing OpenAI client at Inception’s endpoint, swap the model string, and your code keeps working — no rewrites, no SDK migration, no refactoring of your tool-calling logic. The docs also show it working through AISuite, LiteLLM, LangChain, and the standard OpenAI client library.
For builders, Inception Mercury 3 supports the features that matter in production:
- Reasoning controls — tunable reasoning levels, so you can dial intelligence up or down against latency
- Tool calling — native function calling for agent workflows
- Structured outputs — schema-aligned JSON output for reliable data extraction
- Enterprise deployments — dedicated capacity, autoscaling, compliance controls, and configurable data retention
One administrative note for existing Inception customers: Mercury 2 and Mercury 2.5 are now closed to new users. Existing customers with usage before 19 October 2026 keep access until 15 November 2026, and enterprise customers remain supported under their existing agreements.

Inception Mercury 3 vs GPT-6 Luna: Which Should You Choose?
The most common question US developers will ask is the one Inception’s marketing is designed to provoke: should I pick Inception Mercury 3 or GPT-6 Luna? They occupy the same strategic slot — the “fast and cheap enough to call a hundred times” tier — so the comparison is fair. Here is how to think about it honestly.
Choose Inception Mercury 3 if:
- Latency is your bottleneck. If you are building agent loops, voice pipelines, or real-time search where every millisecond compounds, the 1,068 tokens/sec figure — 4 to 8x faster than Luna-class rivals — is the single strongest argument for switching.
- You want the cheapest capable option. At launch pricing, Inception Mercury 3 undercuts the speed tier significantly. The 100M free tokens make the evaluation essentially free.
- Instruction-following quality matters. On Inception’s own numbers, the IFBench lead over Luna (76.84 vs 61.82) is the largest margin in the comparison.
- Migration cost is a concern. The OpenAI-compatible API means switching is close to a config change.
Choose GPT-6 Luna if:
- You trust ecosystem over novelty. Luna runs on OpenAI’s battle-tested infrastructure with the deepest third-party tooling in the industry. For a production system handling customer data, that maturity has real value.
- You need long-context reliability. Luna edges at least one long-context evaluation where Inception Mercury 3 does not lead — and OpenAI’s context handling is independently well-documented.
- You want verified benchmarks. Luna’s capabilities have been probed by independent evaluators for far longer than Mercury 3’s one day on the market.
The short version: Inception Mercury 3 is the ambitious challenger with a speed advantage and aggressive pricing; GPT-6 Luna is the safer incumbent with a deeper ecosystem. For latency-sensitive new builds, Mercury 3 deserves a serious evaluation — which, at 100M free tokens, costs nothing to run.
Inception Mercury 3 in the Real World: Early Partners
Vendor benchmarks tell you what a model can do in a lab. Early customer statements tell you what it is doing in production. Inception launched Mercury 3 with statements from four partners — treat them as customer claims, not independent verification, but they paint a consistent picture of where the model fits.
- TinyFish AI ran Inception Mercury 3 inside its production web-agent harness on 50 live Mind2Web navigation tasks. Mercury 3 completed 50% of them — more than GPT-6.1-Sol managed — using the same token volume and agent steps at a fraction of the per-token cost.
- Octen AI tested the model for smart-glasses search, where anything over a few seconds feels broken. Mercury 3 delivered a median generation time of 0.9 seconds and a median end-to-end response time of 2.2 seconds — the best speed-quality balance among the fast models they tested.
- RevisionDojo uses Mercury 3 for interactive education. Hints, marked answers, and definitions come back in about a second, with quality as good as or better than the models it ran before — and it was a drop-in swap for their existing stack.
- Lorikeet reports that Mercury 3 beat other models on latency while also performing higher on evals for customer support agents at regulated businesses.
The pattern across all four is unmistakable: nobody is choosing Inception Mercury 3 for deep philosophical reasoning. They are choosing it for fast feedback loops — education, search, support, web navigation — where a one-second answer beats a five-second brilliant one.

Honest Caveats About Inception Mercury 3
A responsible guide has to say the quiet parts out loud. Here is everything that should temper your enthusiasm about Inception Mercury 3:
- Every benchmark number is Inception’s own. As of 20 October 2026, no independent third party has replicated the benchmark results. Vendor-run benchmarks are usually directionally right, but they are also the most flattering numbers the company will ever publish.
- One early independent voice is cautious. A day-one analysis (Daily_Daily, 20 October) argues that Mercury 3 trails Claude Haiku 5.5 on coding — a direct counterpoint to Inception’s SWE-Bench Atlas headline. The truth will emerge as more evaluators run the model.
- We do not know which NVIDIA GPU produced 1,068 tokens/sec. Inception’s announcement does not disclose the specific GPU. Throughput numbers are hardware-dependent, so the figure may not transfer directly to your serving setup.
- The marketing is inconsistent on the speed multiplier. Inception’s own pages say “5x” on the models page and “4 to 8x” in the blog post. Both cannot be the precise figure; the range is honest, the single number is marketing.
- 50 tasks is 25 completions. TinyFish’s 50% completion rate on 50 Mind2Web tasks means 25 successful navigations. It is a real result, but it is a small sample — do not extrapolate it to every web task.
- It loses two headline benchmarks. DeepSWE v1 (Haiku 51.62 vs Mercury 46.31) and MRCR v2 (Haiku 78.40) go to Anthropic. Anyone telling you Mercury 3 beats everything is not reading the full table.
- The launch discount has no end date — in either direction. There is no published schedule for when $0.10/$0.35 reverts to $0.20/$0.70, and the co-founder’s hope that the discount becomes permanent is not a commitment.
None of this makes Inception Mercury 3 a bad model. It makes it a day-old model. The honest position is curiosity with a spreadsheet: run the free tokens, measure your own workload, and decide on your numbers — not theirs.
FAQ About Inception Mercury 3
What is Inception Mercury 3?
Inception Mercury 3 is a diffusion-based large language model launched by Inception Labs on 19 October 2026. It generates up to 1,068 tokens per second on NVIDIA GPUs, supports a 1-million-token context window, and targets latency-sensitive applications like agents, voice assistants, and real-time search.
How fast is Inception Mercury 3 really?
Inception reports 1,068 tokens per second on NVIDIA GPUs — 4 to 8 times faster than comparable speed-optimised models. The specific GPU model is undisclosed, and the figure has not been independently verified as of 20 October 2026.
How much does Inception Mercury 3 cost?
List pricing is $0.20 per million input tokens and $0.70 per million output tokens. A 50% launch discount (no end date announced) brings that to $0.10 and $0.35. Cached input costs 90% less than standard input, and new API keys include 100 million free tokens.
Is Inception Mercury 3 better than GPT-6 Luna or Claude Haiku 5.5?
On Inception’s own benchmarks, it leads both on instruction following, science reasoning, and web research — but trails Claude Haiku 5.5 on deep coding (DeepSWE v1) and long-context recall (MRCR v2). These are vendor-reported numbers, not independent results.
How do I try Inception Mercury 3?
Sign up on the Inception API Platform for 100 million free tokens, or use it through OpenRouter, Baseten, or Vercel. The API is OpenAI-compatible, so existing code needs minimal changes. Read the details in Inception’s official announcement.
What happened to Mercury 2 and Mercury 2.5?
Both are closed to new users. Existing customers with usage before 19 October 2026 keep access until 15 November 2026.
Final Verdict on Inception Mercury 3
Inception Mercury 3 is the most interesting model launch of October 2026 — and the one most worth evaluating with your own hands. The diffusion architecture is a genuine alternative to the autoregressive consensus, the 1,068 tokens-per-second speed claim is remarkable if it holds, and the launch pricing is aggressive enough to force a response from OpenAI and Anthropic.
But it is one day old, its benchmarks are all self-reported, and it loses two of its own comparison table’s headline categories to Claude Haiku 5.5. The smart move is not to believe the launch post or dismiss it — it is to spend the 100 million free tokens measuring Inception Mercury 3 on your workload. For agent loops and real-time applications, it might be the fastest upgrade you make this year. For everything else, wait for the independent numbers.
Published 20 October 2026. Benchmark figures are Inception’s self-reported results at medium reasoning settings and had not been independently verified at the time of writing. Pricing was accurate at publication; the launch discount has no announced end date.
