Caspian Labs

The search engine for opportunities.

Discover and develop new ideas through deep AI research. Create something new or improve what you’re building.

Explore CASP
A NEW WAY TO EXPLORE

Your next idea could
be a much better one.

Bring a goal, an idea or something you’ve built. We’re building CASP to investigate unfamiliar approaches and develop them into designs, software, scientific proposals and business improvements. Guide the work at any point, or let CASP develop it for you.

MEET CASP

Discover what’s possible. Develop what comes next.

Explore different approaches, work through what each requires and carry the strongest into work you can use.

Find a new way to keep city homes cool.

Deep researchRun example

Interactive example · real sources · illustrative opportunities

CASP / RESEARCH EXAMPLE

A cooling business hidden in a noise problem.

What it changes

What remains to be tested

Read original source
THE EXPERIENCE WE’RE BUILDING
01 ↗

Find a different way forward.

Discover unfamiliar approaches and new applications that could improve what you’re building or make a new idea possible.

02 ↗

Develop the idea into useful work.

Work through the alternatives and requirements, then develop the design, software, scientific proposal or business improvement the project needs.

03 ↗

Keep building on what works.

Continue the same project as requirements and evidence change, preserving useful decisions while developing the affected work further.

BUILT ON DEEP RESEARCH

CASP Continuum

The research and learning engine behind CASP.

We’re developing Continuum to bring continuous research and accumulated knowledge into every project, helping CASP discover new possibilities and develop stronger ideas over time.

WANDR 45-task results

  1. CASP
    0.5833
  2. Perplexity
    0.447
  3. Anthropic
    0.262
  4. OpenAI
    0.153
  5. Exa
    0.111
  6. Parallel
    0.080
  7. Gemini
    0.074
CASP development scores · Published system references
WANDR DEVELOPMENT RESULTS

45 tasks.
Preserved development scores.

0.583Research quality & coverageSoft F1
0.493Fully evidenced resultsHard F1
ONE ENGINE. MORE POSSIBILITIES.

Wherever you do
your best thinking.

Expert research and development for your own projects, your team and the tools you already use.

In development. Join us early.
THERE’S MORE AHEAD

What could you
create next?

EARLY ACCESS

Bring your
next possibility.

Tell us what you’d like to create or improve. We’ll review your goal and get in touch about early access.

CASP’s commercial experience is in development. We’ll use these details to reply about your request. No account or payment required.
THE RESEARCH FOUNDATION

Results and references.

CASP’s preserved development evaluation alongside separately published results from WANDR’s 45-task table.

CASP EVALUATION

45 recorded task results. Per-task scores, receipt hashes, and evaluation notes.

Results & methodology Task-level scores
PUBLISHED REFERENCES

The source for the 19 published system configurations below.

WANDR Table 6
Ordered by Soft F1Higher is better ↗

Swipe across for scores

CASP preserved development results and all 19 published configurations from WANDR Table 6.
SystemSettingCompletedSoft F1Hard F1
CASPDevelopment portfolio45 / 450.58330.4926
Perplexity Search as Codexhigh44 / 450.4470.224
Perplexity Search as Codehigh45 / 450.3970.156
Perplexity Search as Codemedium45 / 450.2950.149
Anthropic managed agenthigh45 / 450.2620.099
OpenAI web agenthigh45 / 450.1530.073
OpenAI web agentxhigh45 / 450.1270.060
Perplexity Search as Codelow45 / 450.1210.053
Exa Agentxhigh45 / 450.1110.036
OpenAI web agentmedium45 / 450.0910.053
Parallel Tasksultra8x45 / 450.0800.035
Gemini Deep Researchmax45 / 450.0740.028
Exa Agenthigh45 / 450.0730.029
Parallel Tasksultra4x45 / 450.0670.025
Parallel Tasksultra2x45 / 450.0630.026
Gemini Deep Researchspeed45 / 450.0480.020
OpenAI web agentlow45 / 450.0380.024
Exa Agentmedium45 / 450.0130.004
Exa Agentlow45 / 450.0060.003
Parallel Tasksultra40 / 450.0050.002

Soft F1 measures research quality and coverage. Hard F1 credits results that meet their full evidence requirements. CASP reports the mean of its best preserved official evaluator results across 45 tasks, collected during development. Reference scores come from WANDR Table 6. These are separately reported evaluations, not a controlled head-to-head test.

Built by Atrin Farnamfar.
THE EXPERIENCE WE’RE BUILDING

Planned product experience. Availability is not yet announced.