Service no. 098
Billed Friday, August 14, 2026
Meter read: about 5 min
Load dispatch bulletin
Mixedbread trained a specialist that searches like a frontier model at a tenth of the price, Cursor prebuilds the rooms its agents wake up in, and DeepSeek added an effort dial. The week's tooling all pulls one direction: dispatch the cheapest source that can carry the load.
Today's art direction
A mid-century public power page: night mural, merit order, rate schedules.
Nineteen-thirties electrification programs sold infrastructure the way startups now sell APIs: monumental murals, plain tariffs, and a public promise of cheap power for everyone. This page borrows the whole apparatus, a painted night hero, a dispatch chart, schedule-lettered sections, a bill stub for a masthead, because the day's news is a price war over metered intelligence. The register is civic rather than corporate: squared corners, typewriter labels, and one amber lamp of an accent against prussian cobalt.
Dispatch notice no. 1
Mixedbread released Toast 1, a model trained to do exactly one job: agentic search. Handed a query, it decomposes, greps, inspects, and curates the evidence before returning context, and it matches Claude Opus 5 and GPT-5.6 Sol across retrieval benchmarks while running up to 10x cheaper and 12x faster. The launch post names the shift in plain words: intelligence is metered now, so the routine turns of the loop should go to the cheapest source able to carry them.
Bar height: rough cost per unit of work
The week's releases arranged the way a grid operator prices generation: cheapest capable source first, and the expensive plant runs only when nothing else meets the load.
Schedule A: Tooling
Z.ai released GLM-5.3, calling it the most capable open-weights coding model: a 50 percent jump over GLM-5.2 on its in-house bench and open-source records on Terminal Bench 3.0. The note is unusually frank about a side effect, cyber capability grew faster than expected during post-training, so the weights ship only after two more weeks of safety hardening.
DeepSeek moved V4-Pro to general availability with agent upgrades and a flexible effort switch: low for simple tasks, high for daily agent work, max for the hard ones, plus native Responses API support tuned for Codex. It lands a day after the company's harness preview led issue 097.
Cursor's cloud agents now boot from builds, ready-made copies of your development environment prepared in the background, the way a grid keeps spinning reserve. Environments start 10x faster with 3x faster time to first token, and a broken commit never becomes the active build.
Schedule B: Technique
NanoNets open-sourced Graft, which compiles a repository into a code graph and feeds the matching nodes to Claude Code, Cursor, or Codex through hooks. Its 162-run benchmark reports 42 percent fewer tokens and 60 percent less wall-clock time, with SWE-bench Verified correctness up 12 points over a cold agent.
The latest What's !important roundup collects the Custom Highlight API, CSS navigation matching, and a working fix for text-stroke. Each one retires a small script somewhere in your stack.
Schedule C: Workflow
Vercel's AI Gateway now sets up coding agents in one command, routing any of them to 200 plus models with a shared spend dashboard, per-key budgets, and team policy an agent cannot route around. Metered intelligence, with an actual meter.
Stack Overflow's No Dumb Questions series takes on context architecture: the discipline of deciding what an AI system sees, when it sees it, and when buying that layer beats building it.
When the things a client compares differ mainly along one axis, price per unit, response time, carbon, plot them as a single dispatch curve: ascending labeled bands ordered by that axis, each carrying a name and one honest number. It replaces the three identical feature cards on pricing and plan pages, and it fits anywhere the visitor's real question is which one first rather than which one at all.
What keeps it honest: real units on the axis, one cited number per band, and label text that clears WCAG AA against every band color. If you cannot source the numbers, use a plain table instead of faking a chart.
Service application
Recreate today's visual system with an AI design or build tool. Paste the application below and attach your own generation fleet.
Build a landing page for a public power utility in a mid-century public-works register, in the spirit of WPA electrification posters. Structure: a full-bleed painted night mural hero (transmission pylons marching toward a glowing powerhouse) under a deep cobalt scrim, with the headline and deck bottom-left; a perforated bill-stub masthead under the nav carrying account number, date, and read time as stacked typewriter lines; one merit-order dispatch chart (ascending labeled cost bands over a solid baseline, with an axis caption); content sections headed as rate schedules (Schedule A, B, C) with a large amber schedule letter; and a cream service-application panel for the closing form or code block. Palette: prussian cobalt ground #1B2447, panel navy #22315E, cream text #F0E7CE, amber accent #E8A13C, vermilion #C6472E used at most twice, steel-blue muted #8D9AC4. Type: League Spartan bold caps for display, Bitter for body text, Courier Prime for labels and stubs. Guardrails: body text at least 19px with 1.7 line height and 60 to 75 characters per line; cream on cobalt and every label color must meet WCAG AA; squared corners (6px maximum radius); no neon glow, no gradient text, no fake toggles or search fields; murals and paintings only, with no readable text inside images; hover states around 150ms with a prefers-reduced-motion guard.
Runs in Figma Make, v0, Lovable, Claude Code, and Beaver Builder AI
Schedule D: Closing
A Firefox extension called SlopScan now stamps a slop score on public repos as you browse. The meters are being read in both directions; assume your work is on one.
Schedule E: Record