Inception's Mercury 3 Diffusion LLM Runs 1,068 tok/s, Beats GPT-6 Luna in Its Own Tests

TL;DR

Inception released Mercury 3 on October 19, its fourth-generation diffusion language model. Inception says it generates 1,068 tokens per second on NVIDIA GPUs, has a 1M-token context window, and matches or beats GPT-6 Luna and Claude Haiku 5.5 on 8 of the 12 capability benchmarks in its launch table. List price is $0.20 per million input tokens and $0.70 per million output, so the output rate is above both rivals' $0.50. A 50% launch discount brings it down to $0.10 and $0.35, and that discounted price is the one the comparison table uses. Every benchmark and speed figure so far comes from Inception. Artificial Analysis hasn't listed the model yet, and the last time it measured a Mercury, the result came in at 60% of the advertised speed.

What shipped

Mercury 3 is live on Inception's own API, on OpenRouter, on Baseten, and on Vercel's AI Gateway. The API is OpenAI-compatible, so switching usually means changing a base URL and a model string. Inception's API platform comes with 100 million free tokens.

The spec sheet from the launch post and the Business Wire release:

• Speed: 1,068 output tokens per second on NVIDIA GPUs, by Inception's measurement, which it pitches as 4-8x faster than the two comparison models.
• Context: 1M tokens, up from 260K on Mercury 2.5. OpenRouter lists a 131,072-token cap on completions.
• Price: $0.20 in and $0.70 out per million tokens at list. During launch it's $0.10 and $0.35, and cached input is 90% off the input price.
• Features: OpenRouter's endpoint (inception/mercury-3-20261019) supports reasoning effort, tool calls, tool choice and structured outputs.
• Size: undisclosed. Inception says only that it has more than 40x the total parameters of Mercury 1 (mini), with "3x" the benchmark intelligence.

Weights and a technical report weren't published.

Why a diffusion LLM is fast

An autoregressive model like Luna or Haiku runs one forward pass per token, strictly left to right. Mercury starts from a masked block of text and refines many token positions in parallel over a series of denoising passes. Each pass costs more than one autoregressive step, but it can settle many tokens at once, so throughput depends on how many positions each pass commits. A regular LLM types like a typewriter. Mercury develops the page more like a Polaroid, with the whole paragraph coming into focus over a few passes.

The open question for diffusion LLMs has been whether quality keeps up as the models grow. Inception co-founder Stefano Ermon's answer, in the release: "We're showing that the limits to scaling diffusion models are compute and data, not fundamental algorithmic barriers." The launch post's scaling chart puts all five Mercury generations on one line and claims that the autoregressive scaling laws "hold for diffusion too."

The benchmark table, read closely

Inception ran all three models at "medium" reasoning and published 12 capability rows plus speed and price. Mercury 3 wins eight of them. The biggest margins are on SWE-Bench Atlas (51.08 vs Luna's 33.33 and Haiku's 32.26) and IFBench (76.84 vs 61.82 and 54.65). It also leads on GPQA Diamond (89.23 vs 86.36), τ³-bench Banking, MultiChallenge, BrowseComp+, DeepSearchQA, and GDPval, where it scores 1,237.2 Elo against Haiku's 1,225.0.

Two of the four losses are long-context benchmarks. Luna edges it on AA-Omniscience-Public accuracy (46.94 vs 46.44) and AA-LCR, a long-context reasoning test (78.30 vs 77.67). Haiku 5.5 wins MRCR v2 long-context retrieval (78.40 vs 75.50) and DeepSWE v1 (51.62 vs 46.31). Inception is now selling a 1M-token window, and long context is where its own table shows it trailing. If your workload is mostly "read this giant log and find the thing," test that case before you migrate.

The launch post also describes a web-agent run on 50 live Mind2Web navigation tasks. Mercury 3 completed 50% of them, "significantly more than GPT-6.1-Sol," with the same token volume and step count, according to the post. No Sol score was given, so all you can take from it is that Mercury 3 finished half the tasks.

Two of the rows carry Artificial Analysis benchmark names (AA-Omniscience-Public and AA-LCR). They're Inception's own runs of those evals, and AA hasn't scored the model on its leaderboard.

The price is a launch price

The speed and cost section of Inception's table puts Mercury 3 at $0.10 / $0.35 next to $0.10 / $0.50 for both Luna and Haiku 5.5, which matches what OpenRouter lists for those two today. That Mercury figure is the 50%-off launch rate. At the $0.20 / $0.70 list price, Mercury 3 charges twice as much for input and 40% more for output than either model it says it beats.

Inception's last launch discount stuck around. Mercury 2.5 launched on September 8 at 80% off ($0.04 / $0.15 against a $0.20 / $0.75 list), and OpenRouter still shows that discounted price six weeks later, which makes for a long launch. Inception hasn't said when the Mercury 3 discount ends.

Speed matters most where latency compounds. A model that finishes a long agent loop in a fraction of the wall time lets you run more attempts in the same latency budget. Inception's partner examples are aimed there: Octen AI reported a median generation time of 0.9 seconds and 2.2 seconds end to end for smart-glasses search, and RevisionDojo co-founder Michael Tong said hints and marked answers "now come back in about a second."

Discount the speed claim, too

Inception's table gives Luna ~130 tokens per second and Haiku 5.5 ~240, which is where the "4-8x faster" line comes from. Those rival numbers are also Inception's.

The only independent data point on a recent Mercury is Artificial Analysis's page for Mercury 2.5. Inception launched that model claiming 1,107 tokens per second. AA measures it at 668 tokens per second through its API, about 60% of the claim, with an Intelligence Index score of 12. Vendor numbers often come from a quiet, well-tuned box, and production endpoints serve real traffic, so the gap isn't a scandal. You should still expect it.

Even at the 2.5 ratio, Mercury 3 would land around 640 tokens per second, still well ahead of the rival figures Inception quotes. The speed advantage is probably real. The exact multiple is marketing until someone outside Inception measures it.

What to do with it

Agent loops and voice are where Mercury's pitch is strongest, because per-step latency adds up there. Swap it into one non-critical loop through OpenRouter and compare wall time and task success against your current small model.

For coding helpers, the SWE-Bench Atlas margin is the most interesting number in the table. Run your own repo tasks before you believe it, since DeepSWE in the same table points the other way. For long-document work, test retrieval at the lengths you actually use, because the long-context rows are the model's weakest.

Budget at $0.20 / $0.70. If the discount lasts as long as Mercury 2.5's has, enjoy the savings.

Key Takeaways

• Inception released Mercury 3 on October 19: a diffusion LLM with a 1M-token context window, tool calling and structured outputs, on Inception's API, OpenRouter, Baseten and Vercel.
• In Inception's own medium-reasoning tests it beats GPT-6 Luna and Claude Haiku 5.5 on 8 of 12 capability benchmarks, with its biggest leads on SWE-Bench Atlas and IFBench.
• It trails on two long-context benchmarks (AA-LCR, MRCR v2), on AA-Omniscience knowledge accuracy, and on DeepSWE.
• List price is $0.20 / $0.70 per million tokens, above the rivals' $0.10 / $0.50. The table compares the 50%-off launch price.
• The 1,068 tokens per second figure is vendor-reported. Artificial Analysis measured Mercury 2.5 at about 60% of its claimed speed and hasn't scored Mercury 3 yet.

Sources: Inception: Diffusion at Scale, Introducing Mercury 3, Inception launch release (Business Wire via Yahoo Finance), OpenRouter: Mercury 3, OpenRouter: GPT-6 Luna, OpenRouter: Claude Haiku 5.5, Inception: Introducing Mercury 2.5, OpenRouter: Mercury 2.5, Artificial Analysis: Inception models, Inception API docs