Inception just dropped Mercury 3, a model 4-8x faster than Luna / Haiku for the same price and benchmark performance.
In real world use, OpenRouter shows that it’s effectively 2-4x faster and 2-4x cheaper effective price.
The time for diffusion language models is finally here!
In real world use, OpenRouter shows that it’s effectively 2-4x faster and 2-4x cheaper effective price.
The time for diffusion language models is finally here!
Oct 20, 2026 · 7:33 AM UTC
68
24
336
24,382
Source:
Mercury 3 is live: the most capable diffusion LLM released to date.
It outperforms GPT-6 Luna and Claude Haiku 5.5 across a range of benchmarks while running at over 1,000 tokens per second.
More than 40x the size of Mercury 1, 3x more intelligent across standard benchmarks, a 1M-token context window.
Available today on the Inception API, @baseten, and @OpenRouter .
1
2
13
4,178
@deedydas Still not fast enough for the performance compromise!
2
0
1
530
@deedydas Diffusion LMs having a real moment is the actual story here, benchmarks aside. If parallel decoding holds quality at 4x the speed, autoregressive's monopoly on "good enough" might finally be over.
0
0
1
82
@deedydas this is the subagent model. file search and test loops don't need opus.
anyone running mercury 3 as the cheap worker inside codex or claude code yet?
anyone running mercury 3 as the cheap worker inside codex or claude code yet?
0
0
1
257
@deedydas mercury 2 and 2.5 are already closed to new users, and existing api customers lose access nov 15
0
0
1
200
@deedydas The part that makes this more than a speed demo: Mercury 3 is 40x the size of Mercury 1 and just as fast. If that holds, diffusion LLMs get smarter without paying for it in latency.
0
0
0
94
@deedydas Sub-$0.40 output at 300+ tok/s while matching Haiku/Luna benchmarks... the inference speed war is getting crazy.
0
0
0
8
@deedydas Mercury 3 is cheaper and faster. Other AI companies must be furious someone tried both.
1
0
0
163
@deedydas diffusion models being 4x faster means the loading spinner finally gets to retire
0
0
0
5
Trump has made one of Xi’s most dangerous potential miscalculations significantly more likely. foreignpolicy.com/2026/09/28/xi-…
451
901
3,679
1,869,030
@deedydas 2-4x cheaper at the same quality is the part that actually changes buying decisions.
0
0
0
29
@deedydas Mercury 3 sounds like it skipped the line at the token counter.
0
0
0
127
@deedydas Most of the 2-4x is input: $0.022 vs $0.043 for Luna and $0.086 for Haiku in your OpenRouter chart. Output is $0.36 vs $0.45 and $0.57, only 1.3-1.6x. Output-heavy agent runs get the smaller number.
0
0
0
103
@deedydas Agent loops chaining dozens of tool calls will notice this speedup way more than anyone just chatting with it.
0
0
0
251
@deedydas 2-4x cheaper in real traffic is what'll get people switching.
0
0
0
200
@deedydas Man I wish it was actually good... Idk if it's just launch issues or what but my benchmarks of it are extremely swingy on latency. Taking as much as 5x longer for the same prompt within seconds randomly and regularly
0
0
0
104
@deedydas 2-4x faster in real use at Haiku level is already huge. Speed irons out so many of the small issues, can't wait to see diffusion models go further
0
0
0
146
@deedydas Diffusion LMs write the whole answer at once, like a polaroid developing. Token streaming suddenly looks like a typewriter 🙂
0
0
0
136
@deedydas 2-4x faster and cheaper on OpenRouter is actually the part that matters most for real work right now. Performance benchmarks are nice but speed/cost wins in deployment.
0
0
1
80
@deedydas in my agent loops one extra wrong turn would eat that whole speedup pretty fast
0
0
0
185
@deedydas the whole answer just appearing at once is gonna feel so weird in a chat ui ngl
0
0
0
108
@deedydas 2-4x on openrouter next to 4-8x on the lab chart, i'll take the openrouter one
0
0
0
68
@deedydas The OpenRouter numbers matter more than the benchmark. 2-4x cheaper in real use can actually change model routing.
0
0
0
104
@deedydas same price at 4x the speed makes it a scout model candidate
0
0
0
54
@deedydas speed only matters to me if it returns the same structure every time, a fast model I have to retry twice isnt actually faster
0
0
0
65
@deedydas 4 to 8x faster at the same price. Would you switch your default model for speed alone?
0
0
0
23
@deedydas 拡散LLMが速いのは分かるけど、私は節約でモデルを安く替えたのに、検証し直す手間で結局コストが増えて転んだから、本番投入前の品質確認が一番怖い。
0
0
0
40
