Browser Use's rl-env-use Lifts Qwen 3.6 Past Claude Opus for $8

Browser Use has released rl-env-use, an open-source CLI that converts a one-line skill description into a dataset, grader, and interface for reinforcement learning or prompt optimization. Browser agents collect paired inputs and later outcomes from the web, while the pipeline filters the cases, computes reference answers, and separates training data from held-out tests.

Browser Use reports that a prompt-optimized Qwen 3.6 35B-A3B scored 0.78 on its clinical-trial forecasting benchmark, close to Claude Opus 5.5 at 0.77. The prompt hillclimb cost $8 after a $72 environment build. Those figures come from 16 held-out tasks, and the reported run changed the system prompt while leaving model weights fixed.

Fresh outcomes become answer keys

The method extends an Anthropic workflow that proposes changes, measures them against train and test splits, estimates evaluation noise, and reverts regressions. Anthropic sourced cases from production traffic or synthetic generation. Browser Use replaces that input stage with cases mined from live websites.

Clinical-trial forecasting illustrates the design. An earlier trial protocol supplies the model’s input, and a subsequently published result supplies the reference answer used for scoring. Selecting outcomes published after a model’s stated training cutoff reduces the chance that its parameters already contain the answers.

Interactive websites make browser automation relevant because useful records often sit behind search forms, tabs, pagination, or client-rendered interfaces. Browser Use reported the following internal comparison against Claude Code equipped with WebSearch and WebFetch:

The repository presents this as a vendor-run comparison under one configuration. Independent runs would be needed to establish broader tool rankings.

One retained edit moved Qwen

The example in envs/trial-forecast contains 35 forecasts drawn from 29 ClinicalTrials.gov trials whose results appeared after the selected training cutoffs. Browser Use reports the following costs:

Qwen 3.6 35B-A3B is a mixture-of-experts model with 35 billion total parameters and about 3 billion active for each token. Across 16 held-out tasks, its score rose from 0.53 to 0.78. Claude Opus 5.5 scored 0.77, while GPT-6 Astra scored 0.81.

The paired 95% confidence interval for Qwen’s improvement was +0.07 to +0.45. The interval is wide because the test set is small. The hillclimb evaluated 12 proposed prompt changes and retained one.

A constant 50% forecast scores 0.75 under this benchmark’s probabilistic scoring rule, so 0.78 should not be read as 78% classification accuracy. Browser Use attributes most of the gain to calibration: the revised prompt discouraged predictions of exactly 0 or 1 and introduced base-rate priors for different question types.

Four stages, two commands

rl-env-use assembles each environment through four connected stages:

1. Mine paired records. Browser agents collect an earlier input and a later outcome, such as a trial protocol and its result, or a regulatory filing and a subsequent event.
2. Generate and grade tasks. The pipeline writes prompts and uses code to compute the reference answers. Tasks solved too easily by a panel of frontier models can be revised.
3. Reject weak cases. Filters look for leaked answers, ambiguous wording, memorized outcomes, and unstable grading.
4. Optimize or export. The system creates train and test splits, measures candidate prompt changes with paired confidence intervals, reverts regressions, and can export the data for an external RL trainer.

The CLI exposes environment creation and prompt hillclimbing through two commands:

rl-env-use create "predict whether a completed trial met its primary endpoint from its registered protocol; ClinicalTrials.gov" --budget 40
rl-env-use climb envs/<name> --model qwen-35b

The first command creates an environment with a configured budget of 40. The second evaluates prompt changes for the selected model. Python applications can also load the generated environment through an Env interface.

• Runtime: Python 3.11 or later with uv
• Credentials: a Browser Use API key from Browser Use Cloud and an OpenRouter key
• License: MIT

When web-mined environments fit

Environment construction remains a practical constraint for agent training. Hand-authored benchmarks can saturate or enter later training corpora, while synthetic cases may omit the irregular interfaces and records found in production work. Anthropic’s earlier workflow improved its own claude-api skill from roughly 66% to 88%, but it still depended on an existing supply of suitable cases.

Web mining can expand that supply in domains where an input appears first and a verifiable outcome arrives later. Candidate areas include trial results, earnings reports, regulatory decisions, sports results, and genetic-variant annotations.

Post-cutoff publication dates reduce pretraining contamination without guaranteeing a valid forecast test. If the evaluated model can browse the current web, it may retrieve the outcome directly. Environment authors must therefore record source timestamps, preserve evidence, validate the reference-answer code, and control tool access according to whether retrieval is part of the task.

The $8 figure covers prompt search for one model on one narrow environment. The complete environment cost $72, and the model’s weights remained unchanged, giving full RL and fine-tuning a different cost profile. The project’s broader contribution is a repeatable pipeline for building fresh, auditable cases that teams can use for prompt optimization, evaluation, or subsequent reinforcement learning.