SubQ Snap Claims Constant Compute Per Token at 384K Context

The decision model scores 61.7 on Decision Index v0.2.1 against Jev's 57.9, with output free.

@alex_whedon introduces SubQ Snap as the first long-context decision model for real enterprise work. He says it is the only decision model or LLM where compute per token stays constant as context grows. Snap carries 384K tokens of multimodal context, 12x larger than Jev, and scores 61.7 on Decision Index v0.2.1 against Jev's 57.9, outperforming on 13 of 16 enterprise benchmarks.

Speeds run 80ms at minimal context, 250ms at 32k, 660ms at 100k, and 2.5 seconds at 384k. The model processes 100K tokens and returns a decision in under one second. Pricing is $0.05 per million input tokens, with output free, and the first 10,000 sign-ups get 100M free input tokens.

Asked for baselines, Whedon says the 13 wins came against Jev on 16 benchmarks from the Decision Index, drawn from more than 50 evaluated in total. @LLMJunky presses on timing, asking when the LLM actually releases since it was introduced months ago. @rohanpaul_ai says 80ms at minimal context could sit in front of almost every agent call. @omarsar0 says staying useful and efficient at long context is what matters for decision models.

From @alex_whedon

Agents & Coding

Uncle Bob Edits Video With Grok Bot, Camtasia Builds an MCP

@unclebobmartin has the bot listen, sort and place clips in the Camtasia timeline for him.

@unclebobmartin uses Grok Bot for video editing: he uploads the clips, the bot listens to them, sorts them against the script, flags irregularities, and places them in the Camtasia timeline. @JRRDearman, who works on Camtasia, says his team built a desktop MCP where agents write Lua to build projects and check the result in a QA loop, now in tech preview. @omerokumusx recommends Gemini for the same job.

From @unclebobmartin

One Reused Prompt Forces Agents to Think Parallel-First

@acoyfellow's operating rule: a prompt that expects exactly one answer back is a mistake.

@acoyfellow shares one of his most re-used prompts for forcing agents to think parallel-first. The first operating rule reads: “Default to parallel. A prompt that expects exactly one answer back is a mistake.” If a task can be tried two ways, try both and compare, because two competing attempts that an eval can score is the best reason to fan out. @KingBootoshi says he now sends an image of the idea to his engineering agents and tells them to parallelize work.

From @acoyfellow

Models & Hardware

SudoingX Asks Nvidia for Two More Sparks

A full encode ties up both Sparks for 16 to 18 hours; kernels ran on a laptop.

@sudoingX writes an open letter to Nvidia: his two DGX Sparks run a 552B deepseek v4.1 at 103 tok/s. The kernel work happened on his laptop 5090 because both Sparks were busy encoding. The build went from 39 tok/s on October 8 to 60.1 the next day, then to 103.0 with two CUDA kernels and 5 draft tokens, in under 4 days. @niklaslenz_ai mentions four Sparks on his desk, and @machinegenie offers a pair of idle DGXs by DM.

From @sudoingX

AI in Accounting & Tax

SlowBooks Pro Positions Itself as the Free QuickBooks Exit

The self-hosted bookkeeping app reports 500 daily server calls from unique IPs and costs nothing.

@308Greenfield pitches SlowBooks Pro as better than QuickBooks Online and QuickBooks Desktop, free forever, self-hosted, with no telemetry. He counts over 500 daily calls back to the server from unique IPs and says it adds dozens of users a day. In replies he says it imports and exports to most major ledgers, has 560+ REST APIs built for AI use, and ships with Stripe and PayPal built in. Bank integration runs through SimpleFin at $1.50 a month, and he funds the project with his day job.

From @308Greenfield

Sources

Monday, October 12, 2026 Back to the Archive Wednesday, October 14, 2026