Mercury 3 vs GPT-6 Luna vs Claude Haiku 5.5: Which Fast Model Should You Use?
If you want the fastest possible output, Inception's Mercury 3 is the model to test: it claims 1,068 tokens per second and beats GPT-6 Luna and Claude Haiku 5.5 in 8 of 12 benchmarks Inception published. If you want a model you can use right now with a normal chat product around it, GPT-6 Luna and Claude Haiku 5.5 are both live on aiwith. Mercury 3 is not on aiwith or Router yet, so you would reach it through Inception's API, OpenRouter, Baseten or Vercel.
Those claims come from Inception's launch material, not from an independent evaluator. This is a practical comparison of what each model is for, what the vendor numbers do and do not prove, and where to try two of the three today.
Why Mercury 3 is different
GPT-6 Luna and Claude Haiku 5.5 are autoregressive models: they write one token at a time, and every token waits for the one before it. Mercury 3 is a diffusion language model. It sketches a rough version of the whole answer and refines it over several passes, so it can generate many tokens in parallel.
That architecture is the reason for the speed gap. Inception reports 1,068 output tokens per second on NVIDIA GPUs, against roughly 240 tokens/s for Claude Haiku 5.5 and 130 tokens/s for GPT-6 Luna in its own comparison. Inception describes the overall advantage as 4-8x faster. There are no independent speed tests yet, and throughput on your provider will depend on serving stack, batch size and prompt shape.
The practical case is not "faster chat replies". It is workloads where the model is called many times in a row and each call adds latency: browser agents, multi-step tool use, real-time search, voice support and tutoring.
Speed and context at a glance
| Mercury 3 | GPT-6 Luna | Claude Haiku 5.5 | |
|---|---|---|---|
| Architecture | Diffusion | Autoregressive | Autoregressive |
| Claimed output speed | 1,068 tokens/s | ~130 tokens/s* | ~240 tokens/s* |
| Context window | 1M tokens | Family-specific | Family-specific |
| API | OpenAI-compatible | OpenAI, on aiwith | Anthropic, on aiwith |
| On aiwith | No | Yes | Yes |
*Rival speeds are Inception's own figures from its launch comparison. They are vendor numbers, not a neutral benchmark.
The 1M-token context matters for long documents and agent transcripts. It does not automatically make Mercury better at reading them: long-context accuracy is a separate skill, and that is where Haiku 5.5 has an edge below.
Quality: where Mercury wins, and where it does not
Inception's benchmark table gives Mercury 3 8 wins out of 12 tests, with all three models run at medium settings. The two clearest wins:
- Instruction following: 76.84 on IFBench, about 15 points ahead of GPT-6 Luna.
- Autonomous coding: 51.08 on SWE-Bench Atlas, versus 33.33 and 32.26 for the other two.
But the same table is not a clean sweep:
- Claude Haiku 5.5 wins DeepSWE, the longer coding benchmark.
- Claude Haiku 5.5 is better at finding information in long texts.
- GPT-6 Luna wins two tests from Artificial Analysis.
So the honest summary is narrower than "Mercury 3 beats both". Mercury 3 looks strongest at fast instruction-following and agent-style coding. Haiku 5.5 looks stronger for long, careful coding and long-document retrieval. Luna is competitive enough that its price and availability become the deciding factors.
All of this is vendor-reported and has not been reproduced independently. If quality decides your switch, run your own eval set on your prompts and failure cases before moving production traffic.
Price: the launch discount changes the answer
Mercury 3 list pricing is $0.20 per 1M input tokens and $0.70 per 1M output tokens. Inception is running a 50% launch discount: $0.10 input and $0.35 output, with cached input at 90% off. The company has not said when the discount ends.
At launch pricing, Mercury 3 is cheaper than both rivals on output. At list price, the picture flips for short prompts: Inception itself notes that GPT-6 Luna and Claude Haiku 5.5 charge about $0.50 per 1M output tokens, so Mercury's $0.70 output is more expensive once the discount ends. For very long prompts, the rivals' higher long-context pricing can make Mercury cheaper again.
Budget from the standard rate, not the launch rate, or you will be surprised when the promotion ends.
Which one should you pick?
Pick Mercury 3 if your product is latency-bound and the model is in a loop: an agent that takes many steps, a real-time search or voice feature, or a high-volume support workflow where a second of delay per call compounds. Test it through Inception's API, which currently comes with 100 million free tokens, or through OpenRouter if you already route models there. Read our Mercury 3 pricing and access guide for setup details and the exact model IDs.
Pick GPT-6 Luna if you want a fast, cheap general model inside a finished chat product, with no API integration work. It is the cost-optimized member of OpenAI's GPT-6 family and is available on GPT-6 on aiwith.
Pick Claude Haiku 5.5 if your work is long-form coding, careful editing or pulling facts out of long documents, and you would rather trade some speed for reliability. It is available on Claude on aiwith.
If you need all three eventually, keep the autoregressive pair as your default and pilot Mercury 3 on one latency-sensitive path. That gives you the speed test without betting the whole product on a model that has no independent benchmarks yet.
Try the fast models you can use today
Mercury 3 is a promising release, but it is not the only fast option, and it is not the one with the most mature product around it. GPT-6 Luna and Claude Haiku 5.5 are both running on aiwith now, with no API key, no setup and no per-token billing to think about.
Compare fast models on aiwith.chat