# SPARKIT — Full reference

> SPARKIT independently evaluates AI models, agents, and products, publishes public benchmarks, and helps teams integrate AI systems into real work.

This document is generated from the same benchmark data used by the public website. Canonical company information lives at https://sparkit.science.

## What SPARKIT is

SPARKIT has two primary services: independent evaluations of AI systems and integration of selected AI models and agents into customer software and workflows. These services can be purchased separately or together.

SPARKIT also operates an asynchronous research API for evidence-heavy questions. A caller sends a question to `POST /v1/research`; the service returns a job identifier immediately, then the caller polls the job endpoint or supplies a callback URL for a webhook.

The agent can search public web sources and scientific literature, read relevant material, compare claims, perform calculations or code-backed analysis, and produce a Markdown report. When literature evidence is relevant, the response can include inline citations and a structured source list.

SPARKIT is primarily designed to be embedded in another agent, backend, or research workflow. It is not evaluated here against broad product categories. Public benchmark claims compare only the exact systems and modes named below.

## Workflow

1. Accept a research question.
2. Search web and literature sources in parallel where useful.
3. Read and compare relevant evidence.
4. Run calculations or code-backed analysis when the question requires it.
5. Return an inspectable report with citations when source evidence is relevant.

## Benchmark results

### Artificial Biological Intelligence Index beta

The ABI beta compares nineteen AI models on 115 biology questions from LABBench2, HLE-Gold, and CGBench. It publishes combined scores, estimated cost, and response time. These are early automatically graded results, not a universal model ranking.

- Canonical ABI page: https://sparkit.science/benchmarks/abi

### Research agent benchmarks

Measured August 2026. These are dated evaluation snapshots, not permanent product rankings.

### HLE-Gold

SPARKIT answered 64.4% of the 149 evaluated HLE-Gold questions correctly, compared with 53.0% for direct Claude Opus 5 and 39.0% for direct GPT-5.6-Sol.

| System | Evaluation mode | Accuracy |
|---|---|---:|
| SPARKIT | SPARKIT research-agent run | 64.4% |
| Claude Opus 5 | Direct model call | 53.0% |
| GPT-5.6-Sol | Direct model call | 39.0% |

Methodology summary:

- The evaluation uses 149 biology, medicine, and chemistry questions from the gold subset of Humanity's Last Exam.
- Each system is scored against the benchmark answer key, and the reported value is the percentage of evaluated questions answered correctly.
- The direct-model baselines are intentionally shown without the SPARKIT search, reading, and analysis loop; they measure the value of the research workflow rather than a model-family comparison.

Known limitations:

- HLE-Gold is a difficult scientific subset, not a comprehensive measure of every research domain or production workflow.
- Model and agent behavior can change as providers update their systems, so these results are a dated snapshot rather than a permanent ranking.
- Representative run artifacts are published below, but the complete run-level benchmark output is not yet public.
- Benchmark accuracy is not evidence of clinical, legal, or regulatory reliability; cited sources and conclusions still require expert verification.

Supporting material:

- [Canonical HLE-Gold page](https://sparkit.science/benchmarks/hle-gold)
- [Public benchmark source](https://arxiv.org/abs/2501.14249)
- [Representative HLE-Gold run artifacts (JSON)](https://sparkit.science/benchmarks/data/hle-gold-representative-runs.json): Three difficult example questions with run metadata and outputs. These examples are not the complete benchmark dataset.
- [End-to-end HLE-Gold walkthrough](https://sparkit.science/blog/sparkit-101-bio-example): A worked example showing the question, research trace, report, and grading outcome.

## API and integrations

- API docs and Python quick start: https://sparkit.science/docs
- Integration guides for the SDK, MCP, and HTTP API: https://sparkit.science/docs/integrations
- Python package: https://pypi.org/project/sparkit-science/
- MCP package: https://pypi.org/project/sparkit-mcp/
- MCP source: https://github.com/SPARKIT-science/sparkit-mcp

## Output and citations

Completed research jobs return a Markdown report and job metadata. When a question warrants literature evidence, claims can include inline citations and the response can include structured source metadata. Purely computational or definitional questions may not require literature citations.

## Latency

Median end-to-end completion time is approximately 110 seconds. Actual time varies with question complexity, retrieval, reading, and analysis requirements.

## Commercial information

Independent evaluations and AI integration are scoped directly with customers through https://sparkit.science/contact. Research agent prices and commercial terms can change; the canonical current pricing page is https://sparkit.science/pricing.

## Safety, privacy, and limitations

SPARKIT screens queries under its safety policy and refuses prohibited requests. It does not use customer queries or returned reports to train models. Details are available at https://sparkit.science/privacy and https://sparkit.science/terms.

SPARKIT can still be wrong. Citations may be misattributed, sources may be summarized inaccurately, calculations may contain errors, and conclusions may overstate the evidence. Users should inspect the report and verify source material, especially in high-stakes settings.

## Publisher and corrections

Product notes and evaluation writeups are published by the SPARKIT Research Team. Authorship and correction information is available at https://sparkit.science/about#research-team. Benchmark corrections can be reported to info@sparkit.science.

## Canonical page index

- Homepage: https://sparkit.science/
- Benchmarks: https://sparkit.science/benchmarks
- ABI beta: https://sparkit.science/benchmarks/abi
- Pricing: https://sparkit.science/pricing
- Services: https://sparkit.science/#services
- Documentation: https://sparkit.science/docs
- Integrations: https://sparkit.science/docs/integrations
- FAQ: https://sparkit.science/faq
- Blog: https://sparkit.science/blog
- About: https://sparkit.science/about
- Privacy: https://sparkit.science/privacy
- Terms: https://sparkit.science/terms