Model profile & tracker
LLM Speed Comparison: Latency and Tokens/sec on 9 Models
LLM speed comparison: tokens per second, first-answer latency, and time per task on nine models, with effort levels and fast modes. Checked 23 November 2026.
Speed is not one number. A model that writes 260 tokens per second can still keep you waiting four minutes before the first word appears, and a model that streams instantly can take half an hour to finish a hard task because it reasons too much. This page separates the three numbers people call "speed" and shows where each of the nine models we track actually sits, with effort levels and fast modes included.
Everything below comes from independent measurement — mainly Artificial Analysis's performance benchmarking, plus OpenRouter provider telemetry and local-hardware numbers where the model runs on your own machine. Nothing is vendor marketing. Checked 23 November 2026.
The three numbers that make up "speed"
- Time to first answer token (latency). How long from sending the request to seeing the first word of the actual answer. For reasoning models this includes thinking time, which is why it ranges from 3 seconds to over 4 minutes depending on effort level. Chat interfaces live or die on this number.
- Output speed (tokens per second). How fast the model writes once it has started. This drives total time for long generations — documents, code files, batch jobs.
- Time per task. Artificial Analysis's headline metric for the agentic era: how many minutes a model needs to finish a full Intelligence Index task, end to end. A fast-per-token but verbose model loses to a slower model that writes less. This is the number that matters for agents.
LLM speed comparison: the nine models
Median output speed and first-answer latency at each model's typical effort setting, plus time per Intelligence Index task (lower is better). Sources: Artificial Analysis, OpenRouter, vendor API telemetry.
| Model | Output speed | First answer after | Time per task | Notes |
|---|---|---|---|---|
| Gemini 4 Hydrogen | 166 tok/s | 16.8 s (default) | 8.4 min | Fastest of the nine at quality ≥70; $1.87 per task at default |
| Mistral Large 4 | 116 tok/s | ~1–3 s | ~10 min (est.) | 116 tok/s is "notably fast" per AA; preview pricing $0.70/$2.10 |
| Gemini 4 Neon | ~110 tok/s | 22.3 s (medium) | ~21 min | TTFA 16.8–60.1 s across low–xhigh; $6.44 per task |
| Claude Opus 5.5 | ~98–111 tok/s | ~26 s (max, fallback) | ~26 min (max) | Medium effort: ~6.8 min and $2.45 per task |
| Claude Fable 5.2 | 70 tok/s (Anthropic) | 240.7 s (max) | 42.0 min (max) | Fast mode: 263 tok/s at 2× price; thinking time unchanged |
| GPT-6.2 Astra | ~60 tok/s | 2.96 s (medium) | 10.4 min (max) | TTFA 178 s at max, ~3 s at medium; effort is the lever |
| Qwen4-Max | 54 tok/s | 148.6 s | 32.1 min | Writes 2× the median tokens per task — verbosity, not speed, is the cost |
| Inkling 2 Mini | 210 tok/s (local) | — | 9.0 min | 210 tok/s decode on a 96GB M5 Ultra; Apache 2.0, runs on laptops |
| DaisyRose 27B | not independently measured | — | — | Qwen3.8-27B fine-tune; expect base-model class speed (see below) |
Two rows need a caveat. DaisyRose 27B is a reinforcement-learning post-train of Qwen3.8-27B with no independent API throughput measurement published yet — its speed should track the base model's class, not a frontier MoE. Mistral Large 4's time-per-task estimate is derived from its measured output speed and token counts, not a published AA figure.
The fastest model depends on the job
Chat and voice assistants: first-answer latency is everything. GPT-6.2 Astra at medium effort answers in about 3 seconds. Gemini 4 Hydrogen's default lands at 16.8 seconds. The max-effort settings of Fable 5.2 (240.7 s) and Qwen4-Max (148.6 s) are unusable for live conversation — but both are one effort-level change away from seconds.
Batch and long generations: output speed wins. Claude Fable 5.2's fast mode (263 tok/s) is the quickest writer of the nine at standard API rates, at double the price ($16/$80). Inkling 2 Mini writes 210 tok/s locally on an M5 Ultra with no API bill at all. Hydrogen at 166 tok/s is the cheapest fast option through an API.
Agents: time per task is the only number that matters. Hydrogen finishes a full Intelligence Index task in 8.4 minutes at default for $1.87. Inkling 2 Mini does it in 9.0 minutes for $0.59, locally. Opus 5.5 at max effort takes ~26 minutes; at medium it drops to 6.8 minutes and $2.45 — the same model, one setting apart, 3.8× faster.
Effort level is a bigger lever than model choice
The same model can be 60× faster or slower depending on how hard you let it think. GPT-6.2 Astra waits 178 seconds before its first answer token at max effort and 2.96 seconds at medium. Claude Opus 5.5 falls from ~26 minutes per task at max to 6.8 minutes at medium for about a quarter of the cost. If a model feels slow, drop its effort level before you switch models — the intelligence cost is usually 5–11 points, and the speed gain is a multiple.
The reverse trap exists too: Qwen4-Max writes at an ordinary 54 tok/s but finishes a task in 32 minutes because it emits roughly twice the median output tokens. Speed per token tells you nothing about how many tokens a model needs.
Local speed: the laptop tier
Inkling 2 Mini decodes at 210 tok/s on a 96GB M5 Ultra Mac Studio (125 tok/s on a 48GB M5 Max) with 3,620 tok/s prefill. DaisyRose 27B, built on the same 27B base, has no published local benchmark yet; the base Qwen3.8-27B class runs comfortably on 32GB+ consumer hardware. Both are open-weight (Apache 2.0), so your throughput depends on your hardware, not a queue.
Speed vs. price
Fast and cheap are no longer opposites. Hydrogen (166 tok/s, $1.50/$7.50) and Inkling 2 Mini (210 tok/s local, $0.15/$1.875 API) sit at the fast end of this table and the cheap end of our price comparison. Fable 5.2's fast mode is the one true speed premium: 2.5× output for 2× price. To price your own workload across all nine, use the LLM cost calculator.
FAQ
Which LLM is the fastest? Gemini 4 Hydrogen at 166 tok/s is the fastest of the nine through a standard API at frontier quality. Claude Fable 5.2 in fast mode writes faster (263 tok/s) at double the price. Inkling 2 Mini is fastest locally at 210 tok/s on an M5 Ultra with no API cost.
What is a good time to first token? Under 2 seconds feels instant in chat. Most of the nine models hit this at low or medium effort; the multi-minute figures in the table are max-effort thinking time, which you control per request.
Is tokens per second or latency more important? For chat and voice, latency (time to first answer token). For batch jobs and long generations, tokens per second. For agents, neither — use time per task, because a verbose model at high tok/s can still be the slowest to finish.
Do reasoning models have to be slow? No, but their thinking time is what you're waiting on. The same reasoning model typically offers effort levels that trade a few intelligence points for seconds instead of minutes of first-answer latency.