What the release page reports
Inception presents Mercury 3 as a diffusion-based language model. Its announcement reports 1,068 tokens per second on NVIDIA GPUs and a one-million-token context window. Those figures describe the vendor's reported conditions; FyreLinkz has not reproduced them or established throughput on a consumer graphics card.
Separate ordinary pricing from a launch offer
The announcement lists $0.20 per million input tokens and $0.70 per million output tokens. It also advertises a 50% launch discount. Treat the discount as an offer to verify, not a permanent rate. At the listed non-discounted rates, one million input plus one million output tokens totals $0.90 before any other charges.
Throughput does not equal response time
Output tokens per second is only one component of an interaction. A long prompt can take time to process, and an agent can spend most of a task waiting for tools. Streaming behavior also affects how soon a person sees something useful. Record the first useful response, completion time and full task duration separately.
| Measure | What it answers | What it misses |
|---|---|---|
| Output throughput | How quickly output is emitted under stated conditions | Input delay and correctness |
| Time to first useful response | When the user can start acting | Whether the full answer finishes well |
| End-to-end task time | How long the actual workflow takes | Why the run was slow |
| Accepted outcome cost | What usable work costs | Whether another workload behaves similarly |
A fair evaluation for your application
Select a fixed set of representative inputs and a written answer rubric. Keep tool access, output limits and retries equal across candidates. Include a short request, a longer context task and a tool-dependent workflow rather than extrapolating from one favorable case. Save quality failures even when they arrive quickly. Report hardware or service endpoint, date and concurrent workload alongside timing so others can interpret the result.
Context size is a ceiling to investigate
A large context allowance does not prove reliable retrieval from every part of a document collection. Put a known answer near the beginning, middle and end of your test material. Add a contradictory document and require dated source citations. Measure whether the system notices uncertainty and preserves the relevant version. This is a proposed protocol; it does not establish Mercury 3's accuracy on those tests.
Choose by workload
A low-latency assistant and a batch report generator may need different trade-offs. Compare the cost of reviewed outcomes rather than choosing from a speed headline alone. Our news index contains related agent and local-model coverage. Verify access and current terms with Inception before building an integration; an API launch does not establish downloadable model weights.
Frequently asked questions
Is Mercury 3's reported 1,068 tokens per second an independent FyreLinkz test?
No. It is a figure reported by Inception under its stated conditions.
What are the listed non-discounted token rates?
The announcement lists $0.20 per million input and $0.70 per million output tokens, with a separate launch discount.
Does a 1M context window guarantee perfect document recall?
No. Test retrieval, conflicting evidence and citation accuracy on representative documents.
Sources & further reading
Primary sources checked Oct 25, 2026. Vendor statements are attributed; editorial advice is our own.
- 1Introducing Mercury 3 ↗Inception
Help us keep this useful. Report a factual correction →
