Genspark Super Agent in 2026: The $1.25B AI Agent That Makes Phone Calls, Builds Websites, and Actually Gets Things Done
You tell Genspark Super Agent: “Find a family-friendly hotel in Barcelona for September 20-22, compare three options, then call the best one to check if they have connecting rooms.” It researches, builds a comparison table, dials the hotel in Spain, navigates the front desk, asks about room availability, and returns a transcript. All while you go make coffee. This is not a demo video. This is what the product does today.
Super Agent is the reason Genspark went from a relatively obscure AI search engine to a $1.25 billion company in under a year. It is also the reason Genspark’s credit-based pricing drives some users insane.
The pivot that created a unicorn
Genspark launched in June 2024 as a Perplexity competitor: an AI search engine that generated “Sparkpages,” Wikipedia-style summary pages with citations and an embedded chatbot. It worked well enough. But by late 2024, the founders noticed something. Users were not just asking questions anymore. They were asking for deliverables. Slide decks. Video scripts. Follow-up emails. Tables comparing mortgage options.
The product was solving the wrong problem. Users did not want better search results. They wanted work products.
In April 2025, Genspark killed its search product, the thing 5 million users had signed up for, and pivoted entirely to an agentic AI platform built around the Super Agent. The company had raised a $60 million seed round and $100 million Series A (led by an unnamed major investor, at a roughly $530 million valuation). By November 2025, it closed a $275 million Series B at a $1.25 billion valuation. Total funding: $435 million and counting.
The growth numbers:
- $10 million ARR in the first 9 days after Super Agent launched
- $36 million ARR within 45 days
- $50 million+ ARR within 5 months
- 88 to 92% first-month paid user retention
- A team of roughly 20 people at launch, with zero paid marketing
Genspark co-founder Kay Zhu laid out the pivot decision in a blog post titled “Why I Killed Our AI Search Product With 5 Million Users.” His argument: traditional AI search is too constrained by predefined workflows. The real value is in agentic systems that independently reason through problems and take action on the user’s behalf.
What makes the Super Agent different
Most AI tools are one-model, one-response. You type a prompt. One LLM generates one answer. If you want an image, you switch to a different tool. If you want a spreadsheet, you switch again.
Super Agent works differently. It breaks your task into sub-tasks, routes each to the most relevant model or tool, runs them while monitoring and adjusting mid-stream, then assembles everything into a single output.
The system orchestrates 9 specialized LLMs and over 80 integrated tools. Under the hood, it uses what Genspark calls a “Mixture-of-Agents” architecture: when you make a request, the system runs it through multiple models in parallel, evaluates their outputs, and synthesizes the best result. A reflection step discards hallucinations and keeps what the strongest models got right.
The core models include GPT-4.1 (which handles the heaviest research and structured output workloads with its 1-million-token context window), o3 and o4-mini-high for complex reasoning, Claude Sonnet 4 and Opus 4, Gemini 2.5 Flash and Pro, and Grok 4. OpenAI’s GPT-image-1 handles image generation, and the Realtime API powers voice interactions.
This is not a chatbot with a tool-calling plugin bolted on. It is an orchestration layer that turns a natural language request into a multi-step, multi-model, multi-tool workflow. You do not need to know which model does what.
The three-tier agent structure
Genspark is not one monolithic agent. It operates in three tiers.
Tier 1, the Super Agent, is the top-level orchestrator. It handles open-ended, multi-domain tasks by autonomously calling on lower-tier agents. Feed it a vague goal and it figures out the plan.
Tier 2 is Advanced Agents: specialized agents for complex but well-defined output types. These include Agentic Deep Research (Sparkpages), AI Slides, AI Sheets, Agentic Fact Check, Agentic Data Table, and Call For Me.
Tier 3 is Basic Agents: single-purpose tools. AI Chat handles multi-model chat with web search. Image Studio covers 13+ image models with remix and editing. Video Generation spans 15+ video models including Veo 3 and Runway. Translation uses 4 engines with Mixture-of-Agents consensus.
You can access each tier independently. For a one-off spreadsheet, skip the Super Agent and use AI Sheets directly. It is faster and cheaper.
The features worth paying for
Call For Me
This is Genspark’s most attention-grabbing feature. The agent dials a real phone number, navigates phone trees, talks to humans, asks questions on your behalf, and returns a transcript with a summary. It uses a dual-layer architecture: OpenAI’s Realtime API handles the live conversation, while a shadow model monitors and guides the interaction via a message queue.
In Japan, where Genspark sees roughly 48% of its user base, a widely reported case involved a user having the agent call their employer to submit resignation. This went viral. The more common uses (restaurant reservations, checking product availability, calling customer service) are the real daily value.
Availability is currently limited to US and Japan phone numbers. The feature burns 400 to 900 credits per call depending on duration.
Sparkpages (Agentic Deep Research)
The feature that started it all, now powered by the Super Agent’s orchestration. Sparkpages are dynamic research pages that synthesize multiple sources into a Wikipedia-style article with citations, data tables, charts, and an embedded copilot for follow-up questions. They beat raw search results because the agent filters, cross-references, and structures information rather than just retrieving it.
Genspark was GAIA benchmark leader as of mid-2025, scoring 87.8% on Level 1, 72.7% on Level 2, and 58.8% on Level 3. That put it ahead of Manus and OpenAI Deep Research on the more complex multi-step reasoning tasks.
AI Slides and AI Sites
AI Slides generates presentation decks with editable charts, speaker notes, and royalty-free images. Output formats include PPTX and PDF. A 10 to 15 slide deck costs roughly 250 to 500 credits.
AI Sites builds fully responsive landing pages with a live URL from a single prompt, complete with images, product tables, and interactive elements. A full landing page runs 600 to 1,200 credits.
Design quality is functional but not polished. Genspark co-founder Kay Zhu has called the output “80% done.” If you need pixel-perfect design, budget time for manual cleanup or external tools.
AI Browser, AI Pods, AI Drive
Newer additions: AI Browser lets the agent navigate the web, fill forms, and extract structured data. AI Pods converts text into professional-sounding podcasts. AI Drive and “Download For Me” let the agent fetch files from the web and store them in an integrated drive. Useful for assembling research dossiers before a deep dive.
The credit system is the real product decision
Genspark’s pricing is not subscription-based in the way ChatGPT Plus is. You subscribe for a monthly allocation of credits, and every action burns credits from your pool. How many credits are consumed depends on which models and tools the agent decides to use. You do not fully control the cost of any given request.
| Plan | Price | Monthly Credits | Best For |
|---|---|---|---|
| Free | $0 | 200/day | Testing, light research |
| Plus | $25/mo | 12,000 | Regular content, slides, research |
| Pro | $200/mo | 125,000 | Video, calls, websites, heavy automation |
Credit consumption examples:
- Simple Sparkpage: 30 to 80 credits
- AI Slides (10 to 15 slides): 250 to 500 credits
- AI Sites (landing page): 600 to 1,200 credits
- Call For Me (2 to 4 min): 400 to 900 credits
- AI Pods (10-min episode): 1,000 to 1,800 credits
- Deep research briefs: up to 1,000 credits
A Plus subscriber who does ten slide decks per month uses roughly 2,500 to 5,000 credits. Manageable. But start adding video generation or frequent phone calls, and you burn through the 12,000-credit allocation quickly. The Pro tier at $200/month is steep for individuals. Teams pay $30 per seat per month with 12,000 credits each.
Some users report the agent burning credits on unnecessary intermediate steps or failed attempts. There is no refund for credits consumed by a task that failed or produced garbage. Genspark’s Trustpilot rating sits around 2 out of 5, driven largely by billing and credit complaints. This is the product’s biggest unresolved UX problem.
A practical workaround from Daniel Nest’s Why Try AI guide: frontload context into a single detailed prompt to maximize first-attempt success, treat the Super Agent’s output as your “80% draft,” then refine in an external tool like ChatGPT or Claude rather than iterating expensively inside Genspark.
Genspark vs. Perplexity vs. Manus vs. ChatGPT
Genspark does not fit neatly into any single competitor bucket. It competes with AI search tools, autonomous agents, and general-purpose AI platforms at the same time.
| Genspark Super Agent | Perplexity | Manus | ChatGPT | |
|---|---|---|---|---|
| Core function | Task execution + research | AI search engine | Autonomous task agent | General AI assistant |
| Output formats | Slides, sites, sheets, calls, video, text | Text + tables | Code, reports, browser actions | Text, images, code |
| Model access | 9+ models, Mixture-of-Agents | Proprietary + select models | Multi-model backend | OpenAI models only |
| Real-world actions | Phone calls, file downloads, web publishing | None | Browser automation | Limited (browser in ChatGPT) |
| Pricing model | Credit-based subscriptions | Flat subscription ($20/mo Pro) | Invite-only / waitlist | Flat subscription ($20/mo Plus) |
| Best for | End-to-end task completion | Deep research with citations | Complex autonomous workflows | Conversational assistance |
For pure research with extensive citations, Perplexity’s flat $20/month pricing is more predictable and often more cost-effective. For complex autonomous workflows that need browser control, Manus competes directly but is harder to access. For everyday AI chat with occasional deep research, ChatGPT Plus is simpler and cheaper.
Genspark’s edge is breadth: one platform, one prompt, and the system picks the right tools, models, and output format. The tradeoff is the credit meter ticking in the corner.
What nobody tells you
The GAIA benchmark scores deserve more attention than they get. Scoring 72.7% on Level 2 multi-step reasoning puts Genspark ahead of most competitors in real-world task completion. The Mixture-of-Agents architecture is not marketing: running prompts through multiple models and keeping the consensus answer measurably reduces hallucination rates compared to single-model approaches.
The Japanese market dominance (48% of users) is also under-discussed. Voice-first agent workflows (phone calls, reservations, customer service) land harder in markets where phone-based service interactions are still the norm. Call For Me is not a gimmick. It is a real product differentiator in specific geographies.
Now the problems.
Credit consumption is opaque by design. The agent dynamically picks which models and tools to use, so you cannot predict the cost of a request before running it. A simple-sounding task that triggers video generation or multiple agent iterations can cost 5x what you expected.
Customer support is widely reported as slow and unresponsive. For a $200/month Pro subscription, that is not acceptable.
Privacy implications of voice calls are unresolved. Call recordings are stored. Genspark does not publish a clear data retention policy for voice interactions. If you use Call For Me for anything sensitive (medical appointments, financial inquiries, legal matters), you are trusting a startup with data most people would not share with their email provider.
Who should use Genspark Super Agent
Start with the Free tier if you are curious about agentic AI but have not tried it, you do occasional research that could benefit from Sparkpages, or you want to test the Mixture-of-Agents approach against your current tools.
The $25/month Plus plan makes sense if you produce slide decks, reports, or research briefs regularly, you want access to premium models (o3-Pro, Claude Opus 4) under one subscription, or you do 5 to 15 agent-driven tasks per month.
The $200/month Pro plan is only worth it if you generate video content regularly, use Call For Me for business operations, build landing pages or sites frequently, or treat Genspark as a core production tool rather than a supplement.
Skip Genspark if you already have ChatGPT Plus or Claude Pro and only need occasional deep research, predictable monthly pricing matters more to you than breadth of capability, or you need pixel-perfect creative output without manual post-processing.
The bottom line
Genspark Super Agent is the most ambitious attempt yet at a general-purpose AI agent that bridges the gap between finding information and getting things done. The Mixture-of-Agents architecture, the three-tier structure, and 80+ tools create a platform that does more than any single-model AI assistant can.
The credit pricing model is the Achilles’ heel. It pits the user against the product: every task comes with a mental calculator running in the background. Genspark would be a stronger product with flat-rate pricing and hard usage caps, or at minimum with transparent per-task cost estimates before execution.
If you want a single AI tool that can research, design, call, code, and publish without switching contexts, Genspark is the best option available in mid-2026. If you want predictable billing and are comfortable using four different tools for four different tasks, you can get 90% of the capability for less money elsewhere.
Try the free tier for a week. Track your actual credit consumption. Then decide whether the math works for your workflow. Do not subscribe to Pro on day one. The learning curve on prompt engineering for agentic systems is real, and you will waste credits learning it.