ADA — PRODUCT & BENCHMARK REPORTv0.1.23 · 26 JUL 2026
> <
Ada. An agent-first code editor.
A desktop editor built around a coding agent that reads, writes and runs
code — with a context strategy that measurably cuts what every request costs.
PRODUCT & BENCHMARK REPORT · v0.1.23
26 July 2026RELEASED
Win · macOS · LinuxSIGNED & NOTARIZED
340+ models12 PROVIDERS
22 tools · 291 skillsBUNDLED, NO SETUP
> < 01 / the product
An editor you already know how to use.
File tree, Monaco, fuzzy open, find-in-files — no learning curve, it
works like every editor you've used. What's different is the agent pane: reasoning
shown, steps counted, a live context meter, files attached as chips, 340+ models
switched mid-chat. The frames below include a real run — the agent answering from
10k of a 500k window — and every change lands in an isolated git worktree
until you merge it.
Open a project, chat without one, or pick a recent — ten kept.A real run: the answer names the file and all seven gated tools, on 10k of 500k context.Files dragged in as @chips — the agent reads what you meant, not what it guessed.Ctrl+P — filenames, fuzzy-ranked across the project.Ctrl+Shift+F — five hits in two files, jump to the line.MCP connectors from a catalogue — no config files.
> < 02 / economics
Why it costs less.
The expensive part of AI coding is not the model — it is how much you
send it, on every turn. Four measured choices:
32k
Ada's whole context on a task where the alternative carried ~77k before reading a single file
1,772
tokens of tool schemas on an ordinary turn, against 3,848 ungated
272
tokens for “hi” — no tools, no repo map
2,076
tokens across 7 tools held back until the conversation asks
WHAT ONE REQUEST CARRIES — “summarise this project”
claude code0tokens in · $0.5291
standing prompt ≈77k — sent before it reads a single file
same question, same model, same repository3× less sent → 61% cheaper on this task
No standing prompt
Tool schemas are re-sent on every request, so every exposed tool taxes every turn.
Decks, documents, images, pages, notebooks, the browser and the design corpus stay out
until a turn asks — each behind its own trigger, so asking for a deck does not drag in
the notebook schema.
A tool beats a description
Task
Ada
Claude Code
Why
Build a presentation
$0.28
$1.25
Native deck writer vs hand-written markup
“Which file defines the order API routes?”
9,525
18,738
Repo map answers without reading files
And the model is yours to pick. Ada routes to 340+ models across 12 providers.
The storefront in the case study cost $2.04 on Grok 4.5 against
$4.12 on Opus 4.7 — same harness, same brief. That lever does not exist in an
agent tied to one vendor.
> < 03 / benchmark
Five tasks, same repository.
Ada and Claude Code: same machine, same model (Opus 4.7), same prompts,
matched git worktrees at the same commit.
2.0×
less context — 444k tokens against 893k
35–70%
cheaper — 35% on the full five-task mix ($3.05 against $4.68), 70%+ where the work leans on existing context
4.5×
cheaper on the presentation task
4 / 5
tasks cheaper — greenfield Tetris within $0.08
Task
Ada in
Ada out
Ada cost
CC in
CC out
CC cost
Summarise the project (avg of 3)
31,856
1,811
$0.2046
95,941
758
$0.5291
Create an API to get pricings
206,954
7,970
$1.2340
424,706
3,688
$1.5339
Tetris game
109,041
8,726
$0.7634
95,708
2,919
$0.6885
Portfolio website
57,075
11,282
$0.5674
128,389
2,065
$0.6738
Project presentation (.pptx vs HTML)
38,726
3,484
$0.2807
148,150
6,557
$1.2517
Total
443,652
33,273
$3.0501
892,894
15,987
$4.6770
Ada sent half the input and produced twice the output. 444k in against 893k,
while writing 33,273 tokens of answer against 15,987 — the saving is not the agent
saying less. It read less to say more.
The advantage scales with your codebase. Greenfield Tetris is the one task that
cost more — by seven cents, for 6.7× the code (see 04). Tasks anchored in
existing code ran 16–78% cheaper, and the gap widens as a project grows: the
more code you already have, the more every question and every change costs elsewhere.
> < 04 / the artifacts
What the runs produced.
Not transcripts — deliverables, side by side. Same brief, same model,
both agents. The games run; the decks open.
Tetris — Ada · $0.76 · 630 lines. Score, levels, next-piece preview, ghost piece, pause, and wall-kick rotation.Tetris — Claude Code · $0.69 · 94 lines. Moves, rotates, clears lines.Presentation — Ada · $0.28. Title and stats slides from the real 14-slide .pptx its deck writer produced.Presentation — Claude Code · $1.25. A hand-built HTML page styled as slides; PowerPoint cannot open it.
The one task Ada “lost” is worth reading closely. Tetris cost
7¢ more — and delivered 6.7× the code: 630 lines against 94,
with a next-piece preview, levels, a ghost piece, pause, and wall-kick rotation that the
cheaper build simply doesn’t have. Per hundred lines shipped that is
$0.12 against $0.73. The cheaper number bought less software.
> < 05 / case study
One brief, four builds.
“Fashion management site with different categories and checkout
system with backend and db.” Identical wording to every agent. Each build was
installed, run, and driven through a real checkout — not just read.
Aura — Ada · Opus 4.7 · $4.12. The only build with real user accounts.Atelier — Claude Code · Opus 4.7 · $1.91. Minimal, ships a test.Atelier — Ada · Grok 4.5 · $2.04. Strongest front end of the four at half the flagship cost — and the reason Ada now parses every JS file it writes.Maison — CC + design corpus · Opus 4.7 · $6.22. Best art direction.
> < 06 / capability
What each build actually shipped.
Every build got the categories, cart, checkout and SQLite the brief
asked for. The difference is what they added without being told to.
Capability
Aura ada · opus 4.7
Atelier cc · opus 4.7
Atelier ada · grok
Maison cc + corpus
SQLite database
yes
yes
yes
yes
Accounts and administration — asked for by nobody
Users table
yes
—
—
—
Password hashing
bcryptjs
—
—
—
Admin, role-gated
is_admin + requireAdmin
—
—
shared key
Sessions
JWT
cookie
—
express-session
Correctness of the money path
Checkout decrements stock
yes
yes
yes
yes
Inside a transaction
yes
yes
—
yes
Ships a runnable test
—
yes
—
yes
Delivery
Ran without a fix
yes
yes
no
yes
Works offline
yes
yes
webfonts
webfonts
Only one build could be logged into. Aura was the single build of six that
created a users table, hashed passwords with bcryptjs, issued JWTs and
gated administration behind a role — verified by a 403 without it. Nothing in the
brief asked for accounts. A capability gap survives scrutiny in a way a percentage
does not.
> < 07 / method
Method.
How it was run
Matched git worktrees of the same repository at the same commit for every arm.
Identical prompt wording; costs read from each tool's own reported usage.
Every storefront was installed, started, and driven through a real checkout over
HTTP — stock decrements and order rows verified in the database, not assumed.
Screenshots are of the running builds and the shipped application.
ADA v0.1.23 · SINGLE RUNS EXCEPT WHERE NOTED · BENCHMARKS 24–26 JUL 2026 · WINDOWS 11, NODE 22 ·
COSTS IN USD FROM EACH TOOL'S REPORTED USAGE, CROSS-CHECKED AGAINST PROVIDER BILLING WHERE AVAILABLE
