What the Mercury 2.5 770 tokens per second benchmark buys in an agent loop

Mercury 2.5 is a proprietary model released in September 2026, with text input and text output only. The page in question shows the reasoning version, and Artificial Analysis notes that a non-reasoning variant may also exist. Because it is proprietary, it is compared against both proprietary and open weights models in the $0.15 to $1 per 1M band, using a blended 3:1 input/output price ratio.

Artificial Analysis defines output speed as tokens per second received while the model is generating, measured after the first chunk has been received from the API for models that support streaming. Figures represent the model's first-party API, or the median across providers where a first-party API is not available. So 770.4 is a steady-state decode rate, not an end-to-end response time.

The companion metric on the same page, Time per Intelligence Index Task, is described as weighted average decode time in minutes per task and states that it excludes TTFT and overhead time. It is computed by dividing output tokens per task by output speed. That exclusion is the caveat that matters most for anyone sizing a serving loop from this number.

The arithmetic follows directly from the rate. A 10,000-token response decodes in roughly 13 seconds at 770.4 tok/s, and a 20-call agent loop emitting 500 output tokens per call spends about the same 13 seconds in decode across the whole loop. That loop also pays a time-to-first-token charge on each of the 20 calls, and the benchmark measures none of it.

Mercury 2.5 generated 35M output tokens to run the Intelligence Index, against a class median of 85M, ranking 14th of 174 for verbosity. That is roughly 41% of the median model's token count for the same set of tasks. For a fixed per-turn latency budget, fewer tokens and faster tokens multiply rather than add.

The page shows the reasoning version, so the 35M includes reasoning tokens. Artificial Analysis splits Output Tokens per Intelligence Index Task into Answer and Reasoning components in its charts, but the model summary does not give that split for Mercury 2.5. Anyone budgeting tokens for a loop where reasoning is billed at the output rate should treat 35M as an aggregate, not an answer-token count.

Artificial Analysis tracks Time To First Answer Token, Latency by Prompt Type, Latency Variance and Latency Over Time, and defines the first metric as seconds to first answer token received, accounting for reasoning model thinking time. On a reasoning model the first answer token arrives only after the thinking tokens are spent, so a 770.4 tok/s decode rate says nothing about when the first useful byte lands. The summary for Mercury 2.5 carries no value for it.

That gap is where practitioners are aiming. One commenter on Hacker News said they have a product that cannot use an LLM with a p99 of more than 700ms for TTFT. For that class of workload, a speed rank of 2 of 174 is not the deciding number, and the sources do not supply the one that is.

At 770.4 tokens per second sustained, a single stream emits about 2.77M output tokens per hour, which at $0.75 per 1M is roughly $2.08 per hour of continuous generation. Input is $0.25 per 1M, and the 90% cache discount puts cache hits near $0.025 per 1M; Artificial Analysis notes that cache write and cache storage are billed separately and vary by provider. Running the full Intelligence Index costs $0.06 per task on average, which is rank 21 of 174 for cost.

The 260k context window plus the 90% cache discount is the combination that matters for an agent loop resending a growing transcript each turn. Whether it pays off depends on whether the model finishes the task in fewer iterations than a slower, more capable model would, and the sources report no per-task iteration counts.

The Hacker News reaction was largely hostile and mostly aimed at the intelligence column rather than the speed column. The recurring objections:

None of these comparisons appear in the sources as measured results. Artificial Analysis publishes no head-to-head against gpt-oss-120b or deepseek v4 flash, and the model summary gives no per-evaluation breakdown across the ten evaluations in Intelligence Index v4.3.2: AA-Briefcase v1.1, GDPval-AA v2.1, AutomationBench-AA, Terminal-Bench 4.0, SciCode, Humanity's Last Exam, GDP.pdf, CritPt, AA-Omniscience and AA-LCR v1.1. What the sources do support is the ranking: 90th of 174 on intelligence, 2nd on speed, 21st on cost.

Commenters also argued about whether diffusion language models are a dead end, with Mercury as the example. The Artificial Analysis page does not state the model's architecture. One argument in the thread is that serving diffusion at larger batch sizes loses the speed gains, which is why a single-user or on-device setting might be the natural fit; others countered that DiffusionGemma shipped alongside the gemma-4 models and that K2-Horizon-7B has diffusion and non-diffusion variants with the same claimed intelligence. These are practitioner claims, not documented facts in the sources.

A cost of $0.06 per Intelligence Index task is a point on a curve that keeps moving. An Epoch AI report from Emberson and Roodman, summarized by Marginal Revolution on the same day the Mercury 2.5 page appeared, puts the rate of decline at about 47% per quarter over the past three years.

Their worked example: OpenAI o3 cost about $0.30 per question to attain 75% on GPQA Diamond in January 2025, while GPT-5.6 Luna attained roughly the same score for $0.0004 in mid-2026, a roughly 725-fold decline in under 18 months. Marginal Revolution draws the conclusion that the open-model threat is smaller than it appears, because frontier models are becoming cheaper to run at any given level of performance rather than only smarter. Read against that, a speed-and-price position has to be re-earned each quarter.

First, the Time To First Answer Token value and its variance, since the reasoning version pays thinking time before the first answer token. Second, prefill throughput across the 260k window, which the output speed metric does not cover by definition. Third, the per-evaluation breakdown behind the score of 12, because a model that is weak on Humanity's Last Exam but adequate on AutomationBench-AA is a different proposition for agent work than one that is uniformly weak; the page also marks CritPt as under review.

For tracking movement, bestmodelforyourbudget.terrydjony.com pulls the Artificial Analysis free data API on a daily GitHub Actions cron and tabulates Intelligence, Coding, Math, blended $/1M, input $/1M, output $/1M, Tokens/s and TTFT s per model, plus a diff of new, removed and re-scored or re-priced models between consecutive fetches. Its own caveats carry over: the blended price is the 3:1 input/output figure and excludes cached-input discounts, batch pricing and fast modes, and Artificial Analysis re-bases the Index between versions, so a score of 12 under v4.3.2 is not comparable to an older snapshot.