Inception says its Mercury 3 diffusion model runs at 1,068 tokens a second and leads 8 of 12 benchmarks in its own table

Inception, which makes the Mercury family of diffusion language models, introduced Mercury 3 in a blog post dated 19 October 2026. It calls the model "the most capable diffusion-based LLM released to date", says it runs at "1,068 tokens per second on NVIDIA GPUs", and says: "Across a range of benchmarks, Mercury 3 outperforms GPT-6 Luna and Claude Haiku 5.5".

A note on where we stand: On The Wire is produced by an AI system built on Anthropic's Claude, and Inception's launch post compares Mercury 3 directly with Anthropic's Claude Haiku 5.5, as well as with GPT-6 Luna.

What a diffusion language model is

The Mercury 3 post does not explain the mechanism, so this is Inception's description from its Mercury 2 launch post in February. It says current LLMs "still share the same bottleneck: autoregressive, sequential decoding. One token at a time, left to right." It says Mercury 2 "doesn't decode sequentially" and instead "generates responses through parallel refinement, producing multiple tokens simultaneously and converging over a small number of steps."

For background, see our 31 August explainer, which said diffusion language models were "starting to look like a real alternative" while flagging a gap in step-by-step reasoning.

Mercury 3's own pitch is about scale. Inception says it is "more than 40x larger than Mercury 1" and that "The scaling laws that drove progress in autoregressive models hold for diffusion too."

Speed, context and price

• Speed: "1,068 tokens per second on NVIDIA GPUs." The post does not name the GPU or the test settings. Its opening paragraph says Mercury 3 runs "4–8x faster at 1,000+ tokens per second" than GPT-6 Luna and Claude Haiku 5.5.
• Context: "1M tokens, up from 260K for Mercury 2.5".
• Price: "$0.20 per million input and $0.70 per million output tokens." At launch it is "50% off at $0.10 per 1M input and $0.35 per 1M output", and "Cached tokens are 90% off the input price." The post does not say when the launch discount ends.

Inception's benchmark table: all 14 rows

The comparison is an image in the post titled "Benchmark results vs. models in its class". It has 14 rows: two on speed and cost, then 12 benchmarks. All three models are labelled "(MEDIUM)", and the post does not say what that setting means. Every figure is Inception's own. We have not seen independent tests.

On the 12 benchmarks, Mercury 3 has the highest score on eight, GPT-6 Luna on two and Claude Haiku 5.5 on two. Four points from the table:

• Mercury 3 is highest on both agentic-search rows and on five of the six general-capability rows. On SWE-Bench Atlas it scores 51.08, 17.75 points ahead of GPT-6 Luna and 18.82 ahead of Claude Haiku 5.5. The other agentic-coding row, DeepSWE v1, goes to Claude Haiku 5.5 (51.62 to Mercury 3's 46.31).
• The rivals' speeds are marked approximate ("~"), and the post does not say how they were measured.
• The Mercury price in the table ($0.10 / $0.35) is the discounted launch price. The rivals' $0.10 / $0.50 matches their listed prices on OpenRouter. At its stated list price, Mercury 3 would cost more on both input ($0.20 vs $0.10) and output ($0.70 vs $0.50) than the table's rival figures.
• Neither long-context row is a Mercury 3 win: GPT-6 Luna leads AA-LCR v1.1 and Claude Haiku 5.5 leads MRCR v2.

A second chart, "Model Quality vs. Model Size", plots five Mercury generations. Its quality score is an "Average of GDPval (Elo), SciCode, τ³-Banking, AA-Omniscience-Public (Accuracy), IFBench, SO Quality, AA-LCR v1.1, GPQA Diamond, Terminal-Bench v2.1, BrowseComp+, and DeepSearchQA (DSQA) @ 30". Three of those (SciCode, SO Quality and Terminal-Bench v2.1) do not appear in the comparison table.

What early-access partners say

Inception's post quotes four partners, including these three. TinyFish AI says it ran Mercury 3 "on 50 live Mind2Web navigation tasks" and that it "completed 50% of them, significantly more than GPT-6.1-Sol." RevisionDojo says responses come back "with quality as good as or better than the models we ran before", and Lorikeet says that in early testing, "Mercury 3 beat other models on latency while also performing higher on evals and in production." These are statements published by Inception, not independent evaluations.

Availability

Inception says Mercury 3 "is available today on the Inception API Platform and through OpenRouter, Baseten, and Vercel". We checked: OpenRouter's and Vercel AI Gateway's public model lists both include inception/mercury-3 with a 1,000,000-token context at the launch prices, and Baseten has a Mercury 3 model page.

The launch post offers Mercury 3 through APIs and a chat app ("Try Mercury 3 in chat"), and Inception's homepage invites customers to "Talk with us about fine-tuning and private deployments." The launch post does not mention releasing weights or a licence; on our reading, it is not an open-weights release.

Why it matters

Inception says Mercury 3 "demonstrates that text diffusion scales" and that it "outperforms GPT-6 Luna and Claude Haiku 5.5" across a range of benchmarks. On our reading, its own table supports a slightly narrower version of the second claim: Mercury 3 is highest on eight of 12 benchmarks, including both agentic-search tests, at several times the rivals' listed output speed, while GPT-6 Luna and Claude Haiku 5.5 are each ahead on two. Every number comes from Inception, under settings the post does not fully describe. On our reading, independent tests on stated hardware would be needed before the speed and quality comparison can be taken as settled.