Vol. I · No. 1Edition: Mid-morningEst. 2026AboutSearchRSS

The Sand Paper

Finding grains of sand with grains of sand.
A Journal of Intelligence Transfer
ArchiveOctober 2026
Lead Story · Industry

Claude Adds Monthly API Credits to Max and Team Plans

Max plans get $100 to $200 in monthly API credits; Team plans get up to $500, pooled.

Anthropic is rolling out monthly Claude Platform API credits for Max and Team plans, @ClaudeDevs announced in a post @robinebers flagged. Max 5x plans get $100 a month and Max 20x plans get $200. Team plans get up to $500, pooled. The credits work on any model, including Haiku 5.5, in your own code or third-party harnesses.

The credits reset monthly with no rollover, @toni_medic warns: spend them or lose them. @EricWiener5 adds that new Max subscribers wait seven days before claiming. @robinebers’s move is to paste the key into Vercel AI Gateway, which he uses across his apps, so his products burn the subscription credit first and reach for paid models only after.

From @robinebers

Models & Hardware

TensorFold 1.0 Drops Python for a Native Zig Engine

@ashxhart’s TensorFold 1.0 runs inference without a Python engine and gains up to 1.10x decode speed.

@ashxhart shipped TensorFold 1.0, a native Zig engine that loads, tokenizes and runs inference with no Python process at runtime. @niklaslenz_ai benchmarked 1.0.1 against the Python 0.6.6 engine on a Lenovo PGX Spark running Nemotron 3.5 Lightning. Prose decode rose from 125.5 to 137.1 tok/s and code from 184.8 to 186.5. That is 1.09x on prose and 1.01x on code, inside the 1.01x to 1.10x range in the release notes. The new engine boots in under 8 seconds and gave bit-identical answers on all 8 of his test prompts.

From @niklaslenz_ai

Uncensored 180B Qwen Quant Hits 60 Tok/s on One Spark

@ViC305 flags CYBER-FROST-3.8, an uncensored 180B Qwen 3.8 Flash Next, at 60 tok/s on one Spark.

@ViC305 praises @yume_arasaki’s two-part run of @Blackfrost_AI’s CYBER-FROST-3.8, an uncensored 180B version of Qwen 3.8 Flash Next, on a single DGX Spark. The EXL3 quant hit 60 tok/s on the one box, and the six-hour test focused on cybersecurity capabilities. @ViC305 credits @turboderp_’s quantization format and links the recipe and the 3.87bpw model weights.

From @ViC305

Sources

  1. @robinebers on X
  2. @niklaslenz_ai on X
  3. @ViC305 on X
Tuesday, October 6, 2026Back to the ArchiveThursday, October 8, 2026