Verdict: Inception Labs' Mercury 2.5 is the fastest LLM you can call from an API in September 2026 — 1,107 tokens per second on widely available NVIDIA GPUs, at a list price of $0.20/$0.75 per million tokens that undercuts every rival here on output. For latency-bound work (voice agents, search pipelines, coding subagents), Mercury wins. For cheap general-purpose chat with long context, GPT-5.6 Luna still wins. Claude Haiku 4.5 and Gemini 3.5 Flash-Lite are squeezed between the two.
TL;DR
- Mercury 2.5 (diffusion LLM): 1,107 tok/s vendor-reported, 260K context, $0.20/$0.75 per million (Inception Labs); OpenRouter observes 440 tok/s at P50 — still fastest of the four.
- Gemini 3.5 Flash-Lite: 382 tok/s (Artificial Analysis), $0.30/$2.50 per million — fast, but 3.3x Mercury's output price.
- GPT-5.6 Luna: 129 tok/s (Artificial Analysis), $0.20/$1.20 and the largest context window. Best default for general work.
- Claude Haiku 4.5: measured around 82 tok/s non-reasoning at $1.00/$5.00 per million — the slowest and most expensive of the group, bought for Anthropic instruction-following rather than throughput.
- The 80% launch discount that put Mercury at $0.04/$0.15 ended on 8 September 2026 at 07:00 UTC. Budget at list price.
- Last verified: 10 September 2026.
Which is the fastest LLM in 2026, and who should pick what?
Mercury 2.5 wins this comparison by both vendor figure and third-party measurement. Inception Labs reports 1,107 tokens per second and calls it the most capable diffusion LLM on the market, with a 40% intelligence gain over Mercury 2. OpenRouter's telemetry for the endpoint shows 440 tok/s P50 at 1.17 s latency. Both numbers sit above every measured rival.
The split:
- Latency-bound production: Mercury 2.5. Speed plus OpenAI-compatible tool calling, parallel tool calls and structured outputs make it a drop-in for existing agent code.
- General-purpose cheap chat and bulk reasoning: GPT-5.6 Luna. A far larger context window than Mercury's 260K and the maturity of the OpenAI ecosystem matter more than tokens per second when a human is reading the output.
- Google-native stacks: Gemini 3.5 Flash-Lite. Genuinely quick at 382 tok/s and easy to wire into Vertex AI, but you pay for it on output.
- Instruction fidelity over throughput: Claude Haiku 4.5. Choose it when you want Anthropic's behaviour in a small model and can absorb the price.
How do the four models compare on speed and price?
| Model | Measured output speed | Input $/M | Output $/M | Context |
|---|---|---|---|---|
| Mercury 2.5 | 440 tok/s (OpenRouter P50); 1,107 tok/s vendor-reported | $0.20 | $0.75 | 260,000 |
| Gemini 3.5 Flash-Lite | 382 tok/s (Artificial Analysis) | $0.30 | $2.50 | Long-context |
| GPT-5.6 Luna | 129 tok/s (Artificial Analysis, max) | $0.20 | $1.20 | Long-context |
| Claude Haiku 4.5 | ~82 tok/s (Artificial Analysis, non-reasoning) | $1.00 | $5.00 | Long-context |
Sources: Inception Labs, OpenRouter, OpenAI developer docs, Google DeepMind model cards, Anthropic pricing.
Read the table on the axis your workload cares about. On output price, Haiku 4.5 at $5.00 per million costs about 6.7x Mercury 2.5 (DeepMind model card; Anthropic pricing). On measured speed, Mercury's 440 tok/s OpenRouter P50 is roughly 5x Haiku's 82 tok/s and 3.4x Luna's 129 tok/s (OpenRouter; Artificial Analysis). Flash-Lite is the only rival close to Mercury on speed, and it charges 3.3x more per million output tokens (Google DeepMind model card).
Why is a diffusion LLM faster than a standard transformer?
Mercury is a diffusion language model, not an autoregressive one. Models like Luna and Haiku emit one token at a time, so latency scales with output length. A diffusion LLM refines a block of tokens in parallel across denoising steps, which decouples throughput from sequence length — and that is where the 1,107 tok/s figure comes from (Inception Labs).
Inception shipped the first commercial diffusion LLM, Mercury, in February 2025, then Mercury 2 in February 2026. The company was founded in 2024 by Stefano Ermon (Stanford), Aditya Grover (UCLA) and Volodymyr Kuleshov (Cornell), and raised a $50M seed led by Menlo Ventures with Microsoft's M12, NVIDIA's NVentures and angel cheques from Andrew Ng and Andrej Karpathy. A larger model is already in training.
If you would rather compare fast open-weight options, our GLM-5.3 Flash vs Qwen 3.8 Flash-Next vs DeepSeek V4 Flash comparison covers that tier, and Qwen 3.8 Flash-Next versus the larger Qwen variants sets out the tradeoff inside one family.
Does the speed advantage show up in real deployments?
Two vendor-published customer figures: Augment Code reports compaction latency fell from 150 seconds to 27 seconds, an 82% reduction with cost down 90%; voice-agent company OpenCall reports median response latency around 170 ms with P99 dropping from minutes to about one second. Both are vendor numbers, not independent benchmarks — treat them as directional.
OpenRouter's traffic pattern fits that story: the top consumers of the endpoint are agentic clients — Hermes Agent at 3.86B tokens, Claude Code at 2.3B — a workload shape where a coding tool fires many short internal calls and the user waits on the sum.
What are the honest limitations of Mercury 2.5?
Four caveats:
- Headline speed is not what you will measure. 1,107 tok/s is Inception's own hardware figure; OpenRouter's observed P50 is 440 tok/s. Benchmark on your prompts first.
- The discount has ended. The $0.04/$0.15 pricing was an 80% launch promotion that closed on 8 September 2026. Plan against $0.20/$0.75.
- Quality parity is a vendor claim. Inception positions Mercury 2.5 as comparable to Luna (Low), Flash-Lite and Haiku 4.5, measured on its own eval suite; practitioners note its speed comparisons reference older competitor releases.
- Context is the smallest here. 260K tokens is ample for agent loops but well below Luna, which matters for whole-repository passes.
For the frontier tiers, see our GPT-5 versus Claude 4 verdict, the Claude Fable 5.1 pricing comparison, and the best local LLM for coding.
How should you test this on your own workload?
A one-afternoon evaluation:
- Point your existing OpenAI client at the Inception endpoint — the API is OpenAI-compatible, so base URL and key are usually the only changes. Mercury 2.5 also runs via Baseten and OpenRouter, and Inception offers 100 million free API tokens to start.
- Replay 200 real production prompts through all four models, recording time-to-first-token, wall-clock time and output token count. Aggregate at P50 and P99, not the mean — tail latency is what users notice.
- Score outputs with your existing rubric. Mercury's claim is parity at the cheap tier; the question is whether parity holds on your prompts.
- Recompute cost at list prices using your own token mix — input-heavy RAG and output-heavy generation rank these models differently.
- Check tool-calling fidelity explicitly: tool calling, parallel calls, structured outputs and tunable reasoning are supported, with max output of 65,536 tokens per OpenRouter's model page, but verify schema adherence before routing agents through it.
FAQ
Q: What is the fastest LLM in 2026?
A: Mercury 2.5, reported at 1,107 tokens/sec (Inception), observed at 440 tok/s P50 (OpenRouter). Next best: Flash-Lite 382, Luna 129 (Artificial Analysis).
Q: Is Mercury 2.5 cheaper than GPT-5.6 Luna?
A: Input matches at $0.20 (Inception; OpenAI); Mercury output is $0.75 vs $1.20. Luna's bigger context window is the counterweight for long documents.
Q: Why is Claude Haiku 4.5 still worth using?
A: Not for throughput. At 82 tok/s and $1.00/$5.00 per million (Artificial Analysis) it is the slowest, priciest option here — worth it only when you want Anthropic's instruction-following in a small model.
Q: Does a diffusion LLM handle tool calling and agents?
A: Yes — tool calling, parallel tool calls, structured outputs and tunable reasoning through an OpenAI-compatible API, and agentic clients dominate its OpenRouter traffic.
Q: Is the $0.04 per million input price still available?
A: No. The 80% launch discount closed 8 September 2026. List price is $0.20/$0.75 per million (Inception Labs).
Q: What is the catch with the 1,107 tokens per second figure?
A: It is vendor-reported from Inception's own hardware. Third-party observation on OpenRouter is well below it, so benchmark your own prompts before planning capacity.
Corrections log
No corrections yet. Speed and price figures were last verified against primary vendor documentation on 10 September 2026 and will be rechecked when any of the four models ships a new revision.
Top comments (0)