MEASURED RESULTS · PRODUCT VIDEOS
Evidence behind the videos.
The product videos combine recorded benchmark timings with schematic animation, and the standard-tool race on the landing page is the first experiment below. Each result belongs to its stated experiment.
GLM serving benchmark
69.642 s → 9.273 s p99
GLM-5.2, 2026-08-31, 400 attempted requests per arm. Run glm-5.2-2026-08-31. The percentile includes 386 completed, nonempty solo responses and 400 airmet responses, using linear interpolation.
13 solo requests timed out near the 300-second cap. One further record has no output and no recorded error. Neither class is treated as a completed response. There are no such records in the protected arm of this experiment. Zero observed failures is not an uptime guarantee.
The earlier 69.542-second solo figure included the extra zero-duration record. Removing it gives 69.642 seconds, rounded to 69.6 seconds in V5. The CDF uses these same retained samples; the video shows the 0–80-second portion.
The median is about tied: end-to-end p50 7.8 s solo vs 6.8 s with airmet (1.14×), on the same completed, nonempty samples. One model, one pair, one window.
Solo was faster in 98 of 386 completed pairs (25.4%). Output lengths vary. End-to-end timing is not a claim of equal task accuracy, identical answers, or a universal improvement in token throughput.
Recorded timing, reconstructed interface
Manifest declares identical prompts; interleaved rounds use seeds 42 and 99. Indices follow the recorded request arrays. Prompt hashes are not retained.
Recorded stream-arrival timings; neutral blocks reconstruct the interface. Playback speed is labelled in the film. Arrivals include reasoning deltas. No prompt or response text is exported.
One stream event appears at first-token time, followed by one per recorded inter-arrival interval. The interface is reconstructed, not a live customer session.
V6 labels the ordinary replay at 3× speed and the stalled replay at 6× speed. Both sides use the same acceleration, then hold after both requests finish. V5.1 used real-time clocks with a labelled freeze; its stalled solo request remained unfinished when that shot ended.
Separate cost and resilience experiment
149 → 0 observed failures / 3,000 turns per arm
Run g4-stall-highN-20260911, 2026-09-11. Total spend was 1.09× solo (+9%); cost per served turn was 1.04×. This was a stall95 rescue configuration. It is a different experiment from the GLM p99 comparison.
Scope: one model served by two fronts, the second leg fired when a turn stalled past the 95th percentile, 3,000 turns per arm. Solo, 4.97% of turns failed; with the second leg, 0 did.
These results depend on workload, provider behavior and policy. The 2× graphic is an illustrative reference for equal-priced duplication of every call. Zero observed failures does not establish an availability guarantee.
Product and use-case scope
The agent runs in the customer’s infrastructure. Signed factory signals seed local decisions; local observations take precedence as the measurement window fills. Prompt and response traffic goes directly to the configured inference providers, bypassing airmet’s cloud.
Each provider maintains its own cache. Reuse is best effort; no cross-provider KV copy is claimed. The API example assumes a configured local agent and supported OpenAI-compatible Chat Completions features.
Enterprise, chat, coding and batch/RL workflows are schematic use cases. The historical coding pilot’s filtered speed and accuracy claim is omitted from these videos. No coding-task or RL-training multiplier is asserted.