Mercury 2 ranks #170 of 228 on the public leaderboard with a score of 33.61/100. It does not yet have enough sourced coverage for a verified position.
Mercury 2 is a proprietary model with a 128K context window. It uses an explicit reasoning mode, which can improve complex problem solving while adding latency and token use.
Mercury 2 remains supported for existing customers. Inception’s September 8, 2026 models page directs customers to their representative or documentation for access and migration to Mercury 2.5.
Inception released Mercury 2 on February 24, 2026 as a proprietary diffusion reasoning model. Its launch materials report AIME 2025, GPQA, IFBench, LiveCodeBench, SciCode, PinchBench, DeepSearchQA, and a fresh-facts WideSearch run. Those source-backed values replace their earlier entries; all remaining historical rows are retained pending primary-source verification. Independently run display rows remain separate and do not influence ranking.
9 of 766 tracked benchmark slots currently have displayable evidence. Missing categories stay blank.
Its strongest eligible category is Agentic at #122, while its lowest eligible position is Coding at #134. particularly useful for coding agents, browser research, and computer-use workflows.
Model research
Mercury 2.5 is now available
Inception’s September 8 launch introduces Mercury 2.5 with 260K context, tunable reasoning, parallel tool calls, and schema-aligned JSON. Its published standard rates are $0.20 per million input tokens and $0.75 per million output tokens; the launch offer reduces those to $0.04 and $0.15. Cached input is $0.004 during the offer.
The models page says Mercury 2 remains supported for existing customers. Ask Inception for access or migration guidance. We keep the two generations separate so Mercury 2.5 results cannot be mistaken for a new measurement of Mercury 2.
What Mercury 2 is built for
Inception released Mercury 2 on February 24, 2026 as a proprietary diffusion reasoning model. It accepts text, returns text, exposes a 128K context window, and supports tunable reasoning, tool use, schema-aligned JSON, and an OpenAI-compatible chat-completions interface.
The intended fit is latency-sensitive text work that can benefit from high sustained output: agent steps, search synthesis, coding, structured extraction, and the reasoning layer inside a cascaded voice system. Mercury 2 is not a native audio model; voice applications still need speech recognition and speech generation around it.
How to read the speed evidence
The current hosted runtime row measures Mercury 2 at 925 output tokens per second and 4.17 seconds to the first answer. Inception separately reports 1,009 tokens per second on NVIDIA Blackwell. The workloads, hardware, and timing methods differ, so the provider number remains separate from the cross-provider row.
Output rate and first-answer delay measure different parts of a request. Mercury can produce a long response quickly after generation begins while still making a user wait longer for the initial chunk than another model. Voice teams should measure end-of-speech to first audible output across the complete speech stack.
Quality evidence is sourced, but mostly provider-run
Official API pricing is $0.25 per million input tokens, $0.025 per million cached input tokens, and $0.75 per million output tokens. Inception documents the 128K window, reasoning controls, tool use, and structured output on the same first-party release page.
Mercury 2 now has source-attached results for GPQA, LiveCodeBench, SciCode, AIME 2025, IFBench, PinchBench, DeepSearchQA, and WideSearch. Most are provider-run, so the profile links each published row to its exact source and keeps unsupported historical rows hidden. The public score uses only evidence admitted by the ranking pipeline.