AXON◉LABS VOL.26
--:--:-- UTC

RESEARCH BULLETIN · SEP 10 2026

Measuring the rise of machine cognition.

Axon Labs is an independent institute measuring, scaling and interrogating frontier AI systems. This bulletin collects our live training telemetry, benchmark curves and interpretability findings — updated as the runs progress.

Explore the findings ↓ Latest paper ↗
0Frontier parameters
0×1026Training FLOP
0Cheaper inference vs 2022
0GPQA-Diamond · AXON-1
SCROLL
01 / BENCHMARKS

The reasoning climb

Frontier systems are closing the gap with expert humans on the hardest closed-form evaluations. We re-score every checkpoint under identical decoding settings.

FIG. 1
  • MMLU
  • GPQA-Diamond
  • HumanEval
2018 — 2026 · ACC %

FIG. 1 — ACCURACY ON THREE CANONICAL EVALUATIONS. HOVER LEGEND TO ISOLATE · MOVE CURSOR FOR VALUES.

02 / SCALING LAWS

Bigger compute, smaller loss

L(C) ∝ C−0.076

Loss keeps falling as compute grows — a clean power law with no visible bend through 10²⁶ FLOP. The second curve is the one that matters to everyone else: the price of intelligence.

  • Slope α ≈ 0.076 for loss vs. training compute
  • No saturation observed up to 10²⁶ FLOP
  • Inference price fell ≈150× since 2022
  • Compute-optimal frontier keeps shifting right
FIG. 2LOG–LOG · VALIDATION LOSS

FIG. 2 — VALIDATION LOSS VS TRAINING COMPUTE. MARKERS: PUBLIC CHECKPOINTS + AXON-1.

FIG. 3$ / 1M OUTPUT TOKENS

FIG. 3 — COST PER 1M FRONTIER-QUALITY OUTPUT TOKENS, NORMALIZED BY CAPABILITY.

03 / INTERPRETABILITY

Opening the black box

We train sparse autoencoders on internal activations to recover the features models actually use, then trace the circuits that bind them.

Fig. 4 renders live attention flow from layer 24 of AXON-1 while it solves a planning task — each pulse is a head reallocating weight toward a token it considers evidence. Below: the current experiment log.

  • EXP-128Sparse autoencoders at 34M featuresΔ +12.4%RUNNING
  • EXP-121Circuit tracing: induction heads87% mappedRUNNING
  • EXP-132Activation steering: honesty vectorσ 0.91REVIEW
  • EXP-117Feature splitting in residual streamp < .001REPLICATED
  • EXP-109Entropy floors & calibrationr = .84REPLICATED
FIG. 4LAYER 24 · HEAD 7 · LIVE RENDER

FIG. 4 — ATTENTION FLOW, 16×16 TOKEN WINDOW. BRIGHT CELLS = HIGH WEIGHT. RETARGETS EVERY ~2.4S.

04 / COMPUTE & ENERGY

The physics of scale

Frontier runs are now municipal-scale consumers. We publish power, water and heat figures with every model card.

FIG. 5HALL-B POWER DRAW
0MW
OF 120 MW CAPACITY

FIG. 5 — CURRENT CLUSTER DRAW DURING RUN-4213.

FIG. 6PARAMETERS · LOG SCALE
1.5BGPT-2 ’19
175BGPT-3 ’20
≈1.8TFrontier ’23
2.6TAXON-0 ’25
3.4TAXON-1 ’26

FIG. 6 — FRONTIER PARAMETER COUNT, 2019–2026. BAR HEIGHT IS LOG₁₀-SCALED.

0Power usage effectiveness
0%Waste heat reused by district grid
0Litres of water / kWh
05 / PUBLICATIONS

Selected papers

Peer-reviewed and preprint output from the last two research cycles.

  1. arXiv:2604.01882Emergent Planning in Autoregressive Transformers at Depth 96Kovács A., Lindqvist M., Okafor T., et al.Reasoning04 · 2026→
  2. arXiv:2602.11407Sparse Autoencoders Reveal Composable World ModelsIbarra S., Chen W., Novák P.Interpretability02 · 2026→
  3. arXiv:2511.20931Latent Chain-of-Thought Compresses Reasoning 11×Okafor T., Almeida R.Efficiency11 · 2025→
  4. arXiv:2508.04415Compute-Optimal Distillation under Power ConstraintsLindqvist M., Sato H., Kovács A.Scaling08 · 2025→
  5. arXiv:2410.33268Attention Entropy as a Predictor of Downstream CalibrationNovák P., Ibarra S.Evaluation10 · 2024→
  6. arXiv:2403.09912Thermodynamic Bounds on Inference EfficiencyChen W., Almeida R., Sato H.Theory03 · 2024→
Liquid-cooled GPU training cluster in Hall B
HALL B — LIQUID-COOLED TRAINING CLUSTER · 9,216 ACCELERATORS · ZÜRICH-NORD

“We treat interpretability as an empirical science — every capability gets a circuit diagram.”

Dr. Ada Kovács — Research Director, Axon Labs

9,216Accelerators
41 PBCurated corpus
24/7Telemetry feed