Model Benchmark

The AI CLI's accuracy is measured — not assumed — by a dedicated benchmark: 100 tasks in three complexity tiers, judged through layered verdict gates up to real job execution against Dockerized data sources, across 7 mainstream LLMs. This page summarizes the methodology, the results, and what they mean for choosing a model.

Comparing CLI Revisions

Save baseline and candidate benchmark runs in separate directories, then compare their results.json files without making additional model calls:

cd seatunnel-cli
python -m benchmark.compare benchmark/baseline/results.json benchmark/candidate/results.json \
    --out benchmark/comparison.md

The Markdown report pairs model/task/trial identities and shows aggregate deltas alongside every first-attempt and repair-budget pass-to-fail/fail-to-pass transition. Matching recorded model configuration, requested gates, trial count, repair budget, CLI revision stamps, and task-definition fingerprints are required. Missing, incompatible, skipped, or incomplete results are visibly excluded from both denominators, not treated as improvements. Inspect exclusions before interpreting the paired subset as the full suite.

New runs include task_sha256, covering the prompt, assertions, and execution probes. Older result files lack this evidence and are excluded; collect fresh runs with the fingerprint-enabled harness rather than backfilling hashes from current tasks. Existing single-run reports and inputs remain unchanged. The comparison refuses to overwrite an output file; omit --out to print to stdout. Exit code 0 means a report was produced, not that an accuracy gate passed.

This is an offline, descriptive comparison, not a statistical or CI acceptance gate. Keep provider environment variables, model serving state, validation code, connector metadata, engine version, and test data constant when isolating a CLI change. Task fingerprints alone do not establish that these other inputs match.

Results below were measured in July 2026 against seatunnel-cli v0.1.0 (commit 59ada4ec0) with models served by AWS Bedrock. Accuracy drifts as models and the CLI evolve; treat the numbers as a snapshot and re-run the benchmark for current values.

Methodology

Task suite: 100 tasks — 20 simple (single source→sink), 45 medium (type mapping, CDC, transform chains, multi-table), 35 complex (multi-source DAGs, fan-out, conditional routing) — across 12 ETL scenario categories, with 10 Chinese-language prompts and 18 rule probes targeting known LLM failure modes (conditional option misuse, BATCH/STREAMING inference, routing-label wiring).

Verdict gates, mirroring the CLI's own check → dry-run → run pipeline:

Gate	Verdict	Catches
L1 static	HOCON parse + connector metadata + assertions	Syntax, wrong connector, missing options
L3 real execution	Job runs on the official apache/seatunnel image against real MySQL/PostgreSQL/Kafka/ClickHouse/Elasticsearch; batch = exit code, streaming = 60s health	Everything "looks right but doesn't run"

Repair loop: on failure, the failing gate's real error output is fed to the CLI's own repair agent, up to 3 rounds. All verdicts are deterministic — no LLM-as-judge.

Headline Result: Static Rankings Invert Under Real Execution

Model	Static gate (L1)	Real execution (L1+L3)
Claude Opus 4.8	89% (rank 3)	85% (rank 1)
GPT-5.6 Sol	90% (rank 2)	81% (rank 2)
GPT-5.6 Terra	93% (rank 1)	74% (rank 3)

The static leader wrote configs that looked correct but failed at runtime (−20pp static→real decay); the static #3 wrote configs that ran (−6pp decay). "Looks correct" and "actually runs" are different model capabilities — which is why this benchmark executes configs for real, and why accuracy claims based on static checks alone should be treated with caution.

Full Static-Gate Ranking (7 models)

Model	Pass (≤5 repairs)	First-try pass	Repaired
GPT-5.6 Terra	93%	79%	+14
GPT-5.6 Sol	90%	80%	+10
Claude Opus 4.8	89%	87%	+2
Claude Sonnet 5	82%	78%	+4
Claude Fable 5	80%	73%	+7
Qwen3-Coder-Next	67%	53%	+14
DeepSeek V3.2	58%	58%	+0

Model Behavior Profiles

Generation ability and repair ability are independent dimensions — and family traits are consistent:

• First-try-correct (Anthropic family): Opus 4.8 has the highest first-try rate (87% static / 77% real) and the smallest static→real decay — what it writes tends to run. But its repair contribution is minimal (+2): it rarely recovers from its own failures.
• Iterative-repair (GPT-5.6 Terra): unremarkable first-try, but the strongest repairer — it fixed 48% of its real runtime failures, climbing from 55% to 74%. Its drafts carry the most "looks-right-but-won't-run" risk.
• Repair-disabled (DeepSeek V3.2): zero successful repairs in 100 tasks — its repair attempts return the original config or give up. With a product built around an auto-repair loop, this effectively halves the product's mechanism.
• Notable: the deep-reasoning flagship (Fable 5) ranked only 5th — it over-asks clarifying questions where the task calls for sensible defaults. And on the 10 Chinese-language tasks the Chinese-developed models scored lower (Qwen 5/10, DeepSeek 4/10) than GPT-5.6 Sol (10/10).

Model Selection Guide

Scenario	Recommendation	Why
Interactive use (user waiting)	GPT-5.6 Sol / Terra	Fast (25–58s/task), high combined accuracy
Unattended / batch generation	Claude Opus 4.8	Highest first-try rate; output most likely to run unmodified
Budget-sensitive regression runs	Qwen3-Coder-Next	Lowest cost tier; acceptable for smoke testing, not production generation
Avoid for auto-repair workflows	DeepSeek V3.2	Repair success rate of zero in measurement

Known Weak Scenarios (all models)

Thirteen tasks failed for every top-3 model. If your pipeline matches one of these shapes, review the generated config before production use:

Scenario	Failure root cause
Doris / StarRocks sinks	Connector option knowledge gaps (fenodes, load ports, save modes)
PostgreSQL-CDC	Missing prerequisites (replication slots, publications) not reflected in config
Conditional routing (one source split to multiple sinks by predicate)	plugin_output/plugin_input wiring with parallel SQL transforms
Wide DAGs (5+ blocks, mixed transforms)	Label pairing and block-ordering mistakes

These clusters are engineering targets, not permanent limits: they are being addressed through connector-metadata injection and golden-example coverage, with improvements verified against this same benchmark.

Repair Loop: Measured Effectiveness

Feeding real engine errors back to the repair agent recovered 47% of runtime failures (top-3 models combined). The same model repairs structured validation errors at ~2× the rate of raw Java stack traces — evidence that structured error parsing, not a smarter model, is the highest-leverage next improvement for the repair loop.

Running the Benchmark Yourself

The benchmark harness ships in the main repository under seatunnel-cli/benchmark/ — 100 declarative tasks, the layered verdict gates, the Docker data environment, and the report generator.

cd seatunnel-cli

# Credentials via provider-standard environment variables
export OPENAI_API_KEY=sk-...          # or ANTHROPIC_API_KEY / AWS credentials

# One command: installs deps, preflights the environment, runs, reports
./benchmark/run_benchmark.sh --provider openai --model gpt-4o

# Multi-model comparison
./benchmark/run_benchmark.sh --models benchmark/models.json

# Optional deeper gates:
#   L2 (engine dry-run)   — set SEATUNNEL_HOME to a dev-branch build
#   L3 (real execution)   — docker compose -f benchmark/docker/docker-compose.yml up -d --wait

Missing infrastructure degrades gracefully: without an engine or Docker you get a static-gate (L1) report; trials whose requested gates could not execute are excluded from every pass metric and flagged in the summary, so results from differently-equipped machines are never silently compared.

Every report is stamped with the CLI version and git commit under test — rerun the same model on a new CLI build to measure the impact of any prompt, metadata, or repair-logic change on identical tasks. See benchmark/README.md for the full methodology, task-suite layout, and metric definitions.