Skip to content

EcoCoder — 70% lower API cost, 74% less GPU energy.

Sign up for a $10 credit
LeanLM / EcoCoder

AI Coding Endpoint · Metered by Loupe

Opus-level brownfield coding. One third the cost and energy.

EcoCoder (api.leanlm.ai) is a drop-in Anthropic, OpenAI, and Codex coding endpoint. Match Claude Opus 5.5 on multi-file repositories while keeping code clean and cheap on the next feature.

Python .py TypeScript .ts / .tsx JavaScript .js / .jsx Java .java

No credit card required · Hard $10 cap. Already invited? Open Portal ↗

Works with Claude Code, Codex CLI, pi & delegated subagents
# Zero-dependency Python 3 CLI (leaves your daily claude / codex config untouched)
curl -fsSLO https://api.leanlm.ai/ecocoder && chmod +x ecocoder
./ecocoder login --key "eco_live_..."
./ecocoder claude -- --model ecocoder-turbo
Five SlopCodeBench Python repositories (522 tests): Claude Opus 5.5 passed 510 in 31 min for $7.95 and 1,329 kJ estimated; EcoCoder Turbo passed 511 in 35 min for $2.36 and 345 kJ measured claude-opus-5-5 (510/522) $7.95 · 1,329 kJ* ecocoder-turbo (511/522) $2.36 · 345 kJ
Five Python repositories from SlopCodeBench (522 tests, tuning suite; 12- & 19-repo results →). *Opus energy is estimated; EcoCoder energy is measured on the GPU.

511 / 522

Tests passed on ecocoder-turbo (matching Claude Opus 5.5: 510)
35 min vs. Opus 31 min

−70%

Lower API cost ($2.36 vs. $7.95 on Claude Opus 5.5)
5-repository suite

−74%

Less GPU energy (345 kJ measured vs. 1,329 kJ* estimated)
NVIDIA NVML metered

4 langs

Supports Python, TypeScript, JavaScript & Java repositories
Plus SQL, Shell & configs

Matches Opus 5.5 on brownfield repositories.

Head-to-head across five multi-file Python repositories from SlopCodeBench (522 hidden unit and integration tests):

Model Tests Passed Cost GPU Energy Time
ecocoder-turbo 511 / 522 (97.9%) $2.36 345 kJ 35 min
claude-opus-5-5 510 / 522 (97.7%) $7.95 1,329 kJ* 31 min

Five Python repositories from SlopCodeBench (tuning suite; see non-tuning 12- and 19-repository results). *Opus energy is estimated because Anthropic doesn’t publish watts; EcoCoder energy is measured on the GPU. And because repositories stay clean, shipping the next feature on the same codebase cost 35–43% less (details →).

Python, TypeScript, JavaScript & Java.

Backend & Data

Python (.py) & Java (.java)

Multi-package Python services, CLIs, and data pipelines (pytest, uv, poetry) alongside Java backend repositories (Maven, Gradle).

Web, Node & Full-Stack

TypeScript (.ts, .tsx) & JavaScript (.js, .jsx)

TypeScript and JavaScript monorepos, Node/Edge runtimes, React UI components, and test suites (Vitest, Jest, npm/pnpm), plus SQL, Shell, and YAML/TOML configs.

Drop-In Compatibility

Use standalone or delegate subtasks from Opus & Codex via @ecocoder.

Speak Anthropic /v1/messages (./ecocoder claude), OpenAI /v1/chat/completions (./ecocoder pi), or Codex /v1/responses. Run ./ecocoder install-subagent --max-usd 1.50 to keep Opus or Codex as your lead session while delegating bounded implementation and test-repair tasks to EcoCoder.

Choosing between ecocoder-turbo and ecocoder

Use ecocoder-turbo unless cost or energy matters more than pass rate. Across 12 SlopCodeBench repositories (1,352 tests) it passed 25 more tests and finished 13% sooner. ecocoder cost 23% less and used 14% less energy, and all of that saving came from the seven harder repositories:

12 Repositories (1,352 Tests) Tests Passed Cost GPU Energy Time
ecocoder-turbo Recommended 1,282 (94.8%) $7.42 1,176 kJ 117 min
ecocoder 1,257 (93.0%) $5.74 (−23%) 1,013 kJ (−14%) 135 min

Frequently asked questions

Which profile should I use?

Use ecocoder-turbo unless cost or energy matters more than pass rate. Across 12 SlopCodeBench repositories (1,352 tests) it passed 25 more tests (1,282 vs. 1,257) and finished 13% sooner (117 min vs. 135 min), while ecocoder cost 23% less ($5.74 vs. $7.42) and used 14% less GPU energy (1,013 kJ vs. 1,176 kJ)—all from the seven harder repositories.

How does EcoCoder compare with Claude Opus 5.5?

On our 5-repository SlopCodeBench suite (522 tests), ecocoder-turbo matched Claude Opus 5.5 on pass count (511 vs. 510) at 70% lower cost ($2.36 vs. $7.95) and 74% less GPU energy (345 kJ measured vs. 1,329 kJ estimated); Opus was four minutes faster (31 min vs. 35 min).

Which languages and tools are supported—and which are benchmarked?

EcoCoder supports Python, TypeScript, JavaScript, and Java across Claude Code, Codex CLI, pi, and any Anthropic- or OpenAI-compatible SDK. Published benchmark results so far are on Python repositories; TypeScript, JavaScript, and Java suites are in progress.

How does the $10 credit work?

Enter your email and we send an API key with $10 of usage, hard-capped with no credit card required and no expiration. Keys are provisioned same-day; the full 522-test benchmark run cost $2.36, so $10 covers it roughly four times.

Sign up for a $10 credit.

Enter your email and we’ll send your eco_live_... API key and portal invite. No credit card required, hard $10.00 spend cap.

Request received — watch your inbox for your $10 EcoCoder key.

Optional: which client will you connect first?

Already invited? Sign in at api.leanlm.ai/portal →  ·  Evaluating for a team? [email protected]