Reflex Instinct 0.6B (Atlas AI)

Composite (secondary)

0.1 · Not ranked on this board

Cost and measurement conditions

$0.0047 per 1,000 decisions (estimated)

ESTIMATE (*est.), size-class reference; self-hosted open weights, no public bookable tariff for these weights. The UNMODIFIED frozen pricing_v15 path is asserted below: the declared base Qwen/Qwen3-0.6B is in UNLISTED_BASE_MODELS, so reference() returns no market reference rather than raising, and Florian's standing 24 Sep rule gives a clearly labelled size-class estimate. The sub-1B line of the PUBLISHED v1.6 field is used unchanged - USD 0.01 per 1M input and 0.05 per 1M output, the line liquid-d1-omni-600m (600M), jiwo-0.8b and vegaml-0.8b (0.8B) carry there. No row of THIS draw carries that line; it comes from the earlier draws of the same benchmark - so no new release-lane pricing line is created. x this row's own measured tokens: the author's serve.py returns usage.input_tokens as the length of the one sequence it encodes and usage.output_tokens as a literal 0 (one forward pass, a pointer head, nothing generated). Estimated, not charged. COMMON COST BASIS (v1.6.30): cost per 1,000 decisions is computed on the common 1,485 items of this draw, leaving out the 15 very long items (at least 76,358 input tokens as counted by the rows that answered them). They are the 15 items the Jev reference refused with max_tokens_exceeded, which the other rows that did not answer them refused or failed on; their tokens count for no row, whether the row refused or answered them, and a refused item still counts as a wrong answer for Intelligence. This row: USD 0.004712 per 1,000 decisions (0.073x the official cost cap USD 0.064595), against USD 0.006301 over all 1,500 items (v1.6.29).

x2 + 0.15 s (assumption, not measured)

Sandy, the evaluator's own shared CPU host (AMD Ryzen 5 3600, 6 cores / 12 threads, Zen 2, AVX2), NOT a dedicated machine: other jobs ran beside it and the 1-minute load average was logged every 60 s into logs/CONDITIONS-<key>.jsonl. The author's own serve.py, shipped in their weights repo, run with their own documented line `python serve.py --port 11472 --device cpu --repo <the pinned local snapshot directory>` inside a loopback-only network namespace with an outbound probe that failed as required, no credentials in the environment, offline (HF_HUB_OFFLINE=1). float32 on CPU, which is their own code's choice for CPU and the right one on Zen 2 (AVX2 only, where bf16 is about 5x slower). Their serve.py sets no torch thread count, so their own default is every core of the host; this run bounded it to 6 of Sandy's 12 (OMP_NUM_THREADS=6) to stay polite to the other jobs on the box - OUR cap, not theirs, and an unbounded host would use all 12. Peak RSS about 7.6 GB, so the row ran in a 14 GiB systemd scope rather than this job's inherited 4 GiB. Speed is therefore a loaded-shared-CPU figure and a lower bound for this model on a dedicated core or a GPU; the standard self-hosted x2 + 0.15 s adjustment is applied on top and the raw p50/p95 are kept. Re-measured for this release on the frozen v1.6-regular-20261012 self-hosted input (sha256 dfd36720077ef7c41ace367524896507c7f44ce71306fb3e659e8b85069d995b) under CoS deputy decision #18209, at USD 0 on Sandy’s own CPU, 6 of 12 threads, starting 2026-10-13 06:14:53Z; the 1-minute load average is logged for the whole window and reported with the row. DISCLOSED: the author’s own encode() truncates whole state lines above its own max_len=16384 rather than refusing them, so this row answers the draw’s 15 items of 106,804–122,936 characters that four other rows of this draw recorded as context refusals (attest-31b-v1.0, julia-1, standard_one_3b_sh and uniphore-4b-decider; krino-2b, both MeTask rows and torchcast-decision-4b-v1.1 answered them) — the author’s design choice, not ours, and those items are the slowest in the row.