Mercury 3 Claims 1,068 Tokens per Second—Independent Tests Find Trade-offs

On October 19, 2026, Inception introduced Mercury 3 with a claim of 1,068 output tokens per second on NVIDIA GPUs, a one-million-token context window and launch prices of $0.10 per million input tokens and $0.35 per million output tokens; the announcement also quoted TinyFish AI’s account of its browser agent completing 50% of 50 live Mind2Web navigation tasks in the partner’s production harness. Mercury 3 is available through Inception’s API platform, OpenRouter, Baseten and Vercel. The browser result measures an agent using the model, while the speed figure is the provider’s output-throughput claim.

The first independent task results give developers a more mixed picture. Spring Prompt’s October 20 review ranks Mercury 3 18th of 24 models on slide decks and ninth of 13 on survey analysis; in its simulated advertising task, Mercury 3 scored 26.1 against GPT-6 Astra’s 56.6. Those tests assess finished work under their stated settings. They do not directly measure the output-token rate in Inception’s launch claim.

What Mercury 3 offers through the API

The advertised price is a launch discount, with published list rates twice as high. No end date for the promotion has been given. Cached input has separate pricing, so the cost of an application depends on how much input it can reuse as well as how many new input and output tokens it sends.

The large context window describes how much material a request can supply; it is not a measure of how reliably the model will use every part of that material. The service is compatible with OpenAI-style API calls. Early-access examples in the announcement span browser navigation, tutoring, customer support and real-time search, including search for smart glasses. Each application adds its own prompts, tools and serving work around the model, which can affect both the time to a useful answer and the quality of that answer.

What each Mercury 3 result measures

The launch chart, partner account, API specifications and independent tasks answer different questions. Their configurations matter: Inception’s comparative chart uses medium reasoning, the business-task evaluations use the provider default, and the timed chess runs change reasoning settings. Reading them as a single league table would erase those differences.

• Vendor benchmarks: Inception’s chart compares Mercury 3 with GPT-6 Luna and Claude Haiku 5.5 across selected instruction-following, coding, knowledge and agent tasks at medium reasoning. It presents the company’s chosen benchmark results under that setting. The chart does not establish the quality or latency of an answer in a partner application.
• Partner account: TinyFish AI’s Mind2Web navigation trial ran inside its production web-agent harness. A completed task reflects the model together with the agent’s instructions, tools and navigation steps. The account is evidence about that partner setup, with limited visibility into how the result would transfer to a different agent or task set.
• API specifications: The context allowance, token prices, supported routes and claimed generation rate describe the offered service. Tokens generated per second are only one part of request latency: time before generation starts and any tool calls or other application steps also affect when a usable response arrives.
• Independent task measurements: DeckBench assesses whether an analysis becomes a usable slide deck; SurveyBench evaluates survey plans and interpretations; ROASBench scores decisions in a simulated ad-budget exercise. The business runs used the provider default, and recorded API costs exclude judging. Their scores use different scales, so raw points cannot be combined into an overall ability measure.

Reasoning changes Mercury 3’s response time

The clearest measured latency contrast comes from timed chess. The published BulletBench results give Mercury 3 a median move time of about 0.4 seconds with reasoning off and 4.2 seconds at low reasoning in one-minute games; the low-reasoning configuration lost 87.5% of those games on time, while the no-reasoning configuration had no losses on time. The same model therefore produced sharply different outcomes when the reasoning setting changed.

Chess is a narrow test, but its clock makes delayed decisions visible. A fast stream of output tokens cannot guarantee a fast completed decision if the model spends longer reasoning before that output appears. The independent model listing still has no output-throughput measurement comparable to Inception’s advertised rate, leaving the provider’s headline speed claim without a direct outside replication in these results.

Finished work shows a separate trade-off

The business tasks found weaknesses that generation speed alone cannot resolve. In the deck evaluation, recommendations were often late or wrong and the completed decks were not rated ready to present. The survey evaluation found a similar divide between producing a plausible plan and interpreting responses accurately: correct figures did not consistently lead to a sound recommendation.

The advertising result comes from a deterministic simulation of an invented skincare brand, so it compares choices within that exercise rather than real advertising returns. The deck and survey tasks also use constructed cases and limited runs. They show where this initial evaluation struggled, while leaving broader performance across other workloads open.

For developers choosing a model for time-sensitive work, the immediate consequence is concrete: the configuration that made chess moves quickly differs from the settings used for the launch benchmark chart and the independent business tasks. Further independent throughput testing and more results on agent and long-context work will determine how far Mercury 3’s advertised speed carries into completed applications.

Also read:

• AWS Continuum Claims 89% on CyberGym—but the Result Is Vendor-Run
• OpenAI API vs Gemini API: Cheap Tokens Do Not Buy the Same Capability
• Mercury vs Brex: Three Premium Users Can Reverse the Price Advantage