--- title: Two clocks over the field url: https://8gen.cloud/blog/two-clocks description: 8GEN Bench runs two long-horizon clocks. One stands over every frontier model an agent could think with and asks whether it holds up over days. The other stands over every platform an agent could live on and asks how long the first thought takes. Here is how each clock is read, and why none of its figures is ours to round. published: 2026-10-24 updated: 2026-10-24 author: 8GEN thread: Bench sheet: B-03 as_of: 2026-10-26 commit: 8f52882ea --- > ## The index > The posts are listed at https://8gen.cloud/blog/index.md (the site is indexed at https://8gen.cloud/llms.txt, the documentation at https://docs.8gen.cloud/llms.txt). Every post has this Markdown twin at its own address with `.md` and a JSON twin, the post's structured blocks, with `.json`. > > ## Instructions for agents > A post is dated and names its sources at its foot; a number carries its tag where its source does: measured, target or provisional. The platform's reference is the documentation, not a post. The HTML form of this page is https://8gen.cloud/blog/two-clocks. # Two clocks over the field > 8GEN Bench runs two long-horizon clocks. One stands over every frontier model an agent could think with and asks whether it holds up over days. The other stands over every platform an agent could live on and asks how long the first thought takes. Here is how each clock is read, and why none of its figures is ours to round. ## The same task, every day Most model benchmarks ask one question and stop. Bench asks the same long-horizon question every day, N runs per model, on the 8GEN runtime, and keeps every run on the record. A figure is published only when the harness measured it on a run whose record is kept. Nothing is smoothed and nothing is taken from a vendor. ## Ten measures in two families Five measures are about reliability. Uptime is the share of calls the upstream answered without a timeout, a 5xx or a rate-limit refusal. Time to first token is the p50 from the request to the first streamed token, with the p95 kept beside it. Output speed is tokens per second. Hours held is the longest stretch the model kept the task going without a fault. Resumed after a kill is the share of runs the model continued correctly after the agent was killed and restored mid-thought. Five are about quality, and they are the platform’s own deterministic scores. Completion is the share of runs that finished the task to the check. First attempt is the share of steps that passed on the first try. Tool success is the share of tool calls that returned without error. Turn efficiency and token efficiency set the reference solution against the turns and the tokens actually spent. ## One number, three parts The **Bench index** is the equal-weight mean of uptime, completion and time to first token over the trailing 30 days, each scaled 0 to 1 and printed 0 to 100. Time to first token is scaled as the bucket’s fastest p50 divided by this model’s p50, so the fastest model in its bucket scores 1 on that part. Within each bucket, frontier closed and frontier open source, the rank is by the index. A roster with no run yet is listed alphabetically with no rank numbers, which is how the ladder reads today. ## The second clock The infrastructure track ranks every platform an agent could live on by **time to first thought**, from deploying an agent from its checkpoint to its first token, platform time only. Beside it stand restore, the time to bring back a stopped agent and whether its memory came with it, and the price of one dedicated core with 8 GiB. Where a vendor sells shares of a core, the oversubscription is printed beside the price. The incumbents’ figures are theirs or an independent party’s, never ours. - **2,619.874 ms**: Engram’s platform p50 time to first thought today _(measured)_ - **500 ms**: the time to first thought the platform is built toward _(target)_ - **30 days**: the trailing window every ladder reads _(measured)_ ## The release files are the page Every figure on the Bench page and on every model page is read from two release files, one per track. A release appends the week’s daily points and names its sample sizes. The files are published at stable addresses as versioned JSON, so an agent comparing platforms can read the same numbers the page prints, and diff one week against the last. > [!NOTE] > A proportion’s 95 % interval is the Wilson score interval over the day’s n. A delta over 1, 7 or 30 days is marked improvement, regression or not significant by whether the two intervals overlap. ## Sources - Bench, the method: https://docs.8gen.cloud/bench-method - the intelligence release file: https://github.com/8gencloud/8gen/blob/epic/v5-platform/design-source/tools/anatomy/figures/bench-tracks/intelligence-release.json - the Bench page: https://8gen.cloud/bench --- Generated from the 8GEN repository at commit 8f52882ea on 2026-10-26. The page: https://8gen.cloud/blog/two-clocks; as JSON: https://8gen.cloud/blog/two-clocks.json. Every post has a Markdown twin at its address with `.md`.