Mercury 2.5
Mercury 2.5 is a chat model from Inception, measured by Artificial Analysis at roughly 857 output tokens per second, placing throughput at the center of its profile.
API Pricing
| Provider | Input / 1M | Output / 1M | Cached / 1M |
|---|---|---|---|
| $0.040 | $0.150 | $0.0040 |
Prices updated daily. Last check: Sep 24, 2026
Mercury 2.5 pricing by provider
Input, output, and batch rates, plus alternatives, for one provider at a time.
Performance & Benchmarks
Source: Artificial Analysis →Reasoning & Knowledge
- Humanity's Last Exam11.8%
Coding
- SciCode38.5%
Instruction & Long Context
- Long-Context Reasoning71.7%
Benchmarks measured Sep 2026. Scores are independent evaluations, not vendor-reported.
Model Details
General
- Creator
- Inception
- Modalities
- Text
Capabilities
- Tool Calling
- No
- Open Source
- No
Strengths & Limitations
Strengths
- Measured output throughput of approximately 856.6 tokens per second by Artificial Analysis, far above typical autoregressive chat model rates
- Diffusion-based generation approach in the Mercury family gives it a different speed/latency profile from conventional token-by-token decoders
- Sustained generation speed shortens wall-clock time on long completions such as full files of code or multi-paragraph drafts
- Well suited to agent loops and batch jobs where total tokens emitted, not just request count, drives elapsed time
- Served through inference APIs, so throughput is available without self-hosting infrastructure
- Independent third-party performance measurement (Artificial Analysis) rather than vendor-reported speed figures
Limitations
- Time to first token measured at roughly 3,464 ms, which is slow for interactive chat where users expect immediate streaming to begin
- Speed advantage is diluted on short responses, where the initial latency dominates total time
- We do not track published benchmark scores for reasoning, math, or coding accuracy, so quality must be evaluated directly
- Context window size is not recorded in our database, which complicates planning for long-document workloads
- Fewer serving providers and less ecosystem tooling than the most widely deployed chat model families
Key Features
About Mercury 2.5
Common Use Cases
Mercury 2.5 fits workloads dominated by the volume of text generated rather than by per-request responsiveness. Code generation, whole-file refactoring, test scaffolding, and translation of long passages all benefit from the ~857 tokens/second output rate, since these tasks routinely produce hundreds or thousands of tokens per call. Multi-step agent pipelines that generate plans, tool arguments, and summaries at each hop also compress meaningfully when generation is fast. Batch offline jobs — synthetic data creation, bulk document drafting, large-scale content rewriting — are another natural fit, because the ~3.5 second time to first token is irrelevant when nothing is waiting on a stream. Conversely, latency-sensitive interactive surfaces such as autocomplete, voice assistants, or chat UIs where users judge quality by how fast the first words appear are a weaker match, and teams should benchmark Mercury 2.5 against their own accuracy requirements since we do not track published quality benchmarks for it.
Frequently Asked Questions
How much does Mercury 2.5 cost to use?
Pricing depends on which provider serves the model and on the pricing type — per-input-token, per-output-token, and batch or cached rates often differ. Rates also change over time. Check the pricing table on this page for current per-provider figures rather than relying on any fixed number.
What is Mercury 2.5 best used for?
It is best suited to workloads that emit a lot of tokens per request: code generation and refactoring, long-form drafting, bulk translation or rewriting, synthetic data generation, and agent loops with verbose intermediate steps. Its measured ~857 tokens/second output rate compresses wall-clock time most on long completions.
Why is Mercury 2.5 so much faster than other chat models?
Inception's Mercury family uses a diffusion-based approach to text generation rather than purely sequential token-by-token decoding, which allows tokens to be produced in parallel rather than strictly one after another. Artificial Analysis measures the result at roughly 856.6 output tokens per second.
Is Mercury 2.5 a good choice for real-time chat interfaces?
It depends on the interaction pattern. Time to first token is measured at about 3,464 ms, so users see a noticeable pause before text begins appearing. For long answers the fast generation rate can still make total completion time competitive, but for short, snappy replies a model with lower first-token latency may feel more responsive.
What benchmark data do you have for Mercury 2.5?
Our database currently records Artificial Analysis performance measurements — approximately 856.6 output tokens per second and 3,464 ms time to first token. We do not track published accuracy benchmarks such as reasoning or coding scores for this model, so evaluate output quality on your own workload before committing.