Announcements and writeups

Perplexity, ChatGPT Deep Research, Gemini Deep Research, Claude Research, Elicit, SPARKIT — the deep-research category has converged on the basics. The remaining differences are about audience, deployment surface, and how each tool treats citations. A map of who builds what for whom, and where SPARKIT fits.

Drop a five-gene hereditary-cancer panel into SPARKIT and get back the unifying syndrome, lifetime cancer risks, mechanism, and the current surveillance guideline — every claim cited and pulled from primary sources, not from the model's training memory.

sparkit-mcp ships today on PyPI. One config block in claude_desktop_config.json (or the Cursor / Claude Code equivalent) and your LLM client can call SPARKIT mid-conversation. No more leaving the chat to curl the API.

A real workflow: take twenty DepMap-validated selective dependencies, send them to SPARKIT in a single 1,200-character prompt, get back a 105-source synthesis in under thirty minutes. The headline: ATR, EZH2, MDM2, PRMT5, and WEE1 are the most crowded synthetic-lethal fields; WRNIP1, SLC7A11, selective DNMT1, BRD9, PARG, PKMYT1, and selective CDK12 are the clearest open commercial opportunities.

When research agents run at scale, two failure modes dominate: invisible hallucination at industrial volume, and agentic uplift for harmful research. Here's what SPARKIT has in place to engineer against both, and what we deliberately don't claim.

Verified academics get 20% off any SPARKIT subscription, applied automatically at checkout from your academic-domain email. No paperwork, no annual reverification, no separate plan to choose.

We ran a single GAIA question through SPARKIT, direct Claude Opus 4.7, and direct GPT-5.5. SPARKIT fetched Nature's archive, counted 1,002 articles, and answered correctly. Both direct LLMs invented different article counts and confidently landed on the wrong answer.

We ran the same hard scientific question through Perplexity, ChatGPT Deep Research, Gemini Deep Research, Elicit, and SPARKIT. All five got it right. Here is what was actually different, and why only one of them runs inside your codebase.

An end-to-end walkthrough using a real HLE-Gold question: install the SDK, send the call, read the report, and see the agent trace behind a 53% benchmark answer.

A scientific research agent in an API. One call deploys an agent that retrieves and synthesizes the relevant literature, performs any analyses the question requires, and returns a Markdown report.