Wiredrag
Menu

Four tests for an every-place answer, set before judging any tool, ours included

One question asked of four tools, one plain line each: grep's hits, where the type checker stopped, three slices disagreeing, a retrieval tool's top chunks.

Passing all four costs us speed, and we accepted it. Every question reads some 40 million tokens, and that takes minutes.

The same question put to four tools: grep's hits, where the type checker stopped, three slices disagreeing, a retrieval tool's top chunks.

See what a question costs

Every file read - the whole repository in view, not a slice

The value followed - across languages and formats, under every name

What was not read - each skipped file and boundary, named

A price before asking - fixed when sent, not counted in tokens afterward

Where the alternatives beat us, in ways a five-month migration will feel

A repository that fits one 1M-token window needs none of ours.

A larger one asking us still gives up each of these.

Grep and the type checker

run again on every commit, free, for as long as the change takes. Each re-ask here after the first is $45.

A retrieval tool

hands back a ranked answer, likeliest place first. Ours has no ordering by likelihood to start from.

Augment Code

flat $20 or $100 a month, up to 50 seats, that much usage included. Our $180 question costs more than its Business month.

Why forty slices hand back forty answers that disagree

Three slice answers in the order-id scenario, drawn up to show the failure rather than observed. They differ on the count, and on whether the export reads the id as 32 bits.

Three slice answers side by side, disagreeing on the count of places and on whether the export reads the id as 32 bits.

Each slice - a fortieth of the path, counting its own part

The export - declared in one slice, loaded in another

Long inputs - degrade even on simple tasks, and worse with distractors

Claimed length - in a 2024 test of 17 models, almost all fell short of it

Several at once - found less reliably than one; every place is several

2024September 2026TodayIn five months

Magic's stated window holds ten million lines, more than ours

If it ships, it holds repositories we turn away. Your sequence crosses in five months, and Magic has published no date for access.

Magic states 100 million tokens, about ten million lines of code. Ours holds 60 million. It has raised $768 million and runs thousands of GB200s.

Its site offers no product, price, terms or access. Its careers page lists research, kernels, inference and security roles, and no sales or support.

In 2024 it called its own code synthesis not good enough yet. Its newest post, from September 2026, is a pretraining research update, not a release.

Your request has been received.

Expect a message from Wiredrag. It goes to the address you gave.