Strip away the framework debates and most production agents are a model wired to the same short list of tools. I route traffic across roughly 30 providers in 9 categories, and seven of those categories show up in almost every serious agent I see. Here is the baseline toolkit, with the actual routed prices from our catalog (provider list plus 20%; the pricing docs explain the margin) so you can budget instead of guess.
1. Web search
The tool that turns a model with a training cutoff into something that knows what happened this morning. The category splits into a SERP tier (Google-shaped links, cheap) and a neural tier (semantic matching, pricier). The floor is Serper at $1.20 per 1k queries routed; the ceiling is Exa at $8.40, which also returns page text with results. Most agents fire search more than any other tool, so this price matters at volume.
2. Scraping
Search tells you where the answer lives; scraping goes and gets it. The job is turning a URL into clean markdown a model can read, ideally through JavaScript walls. Jina Reader and Spider both cost $0.36 per 1k pages routed, with Firecrawl at $1.92 buying stronger JS rendering and the highest quality score in our scrape catalog (85). Fun fact from running failover: an empty response body should be treated as a failure here, not a success.
3. Document parsing
Scraping handles the web; parsing handles the PDFs, and every business process eventually produces a PDF. Parsers turn documents (tables, scans, multi-column layouts) into structured markdown for RAG. Routed prices run from LlamaParse at $1.50 per 1k pages to Mistral OCR at $2.40, with Reducto at $1.80 holding the category’s top quality score of 92. The LlamaParse vs Reducto breakdown covers the tier math.
4. Embeddings
The quiet workhorse. Embeddings power retrieval, memory, deduplication, and clustering, and they are startlingly cheap: DeepInfra’s bge-large costs $0.012 per million tokens routed, OpenAI’s text-embedding-3-small $0.024, Mixedbread’s mxbai-embed-large $0.072. Embedding a 100-million-token corpus costs single-digit dollars. The catch is not price, it is lock-in: vectors from different models do not mix, so this is the one choice on this list you cannot casually swap later.
5. Code execution
Agents write code, and code an LLM wrote is untrusted by definition, so it needs a sandbox rather than your production shell. E2B and Daytona both price at $0.0000168 per vCPU-second routed, which works out to about six cents per vCPU-hour. That number is small enough that the constraint is safety design (timeouts, teardown), not budget. Code execution is what upgrades an agent from “describes the analysis” to “ran the analysis.”
6. Speech (text-to-speech)
The moment an agent talks to a human, it needs a voice. TTS is priced per 1k characters and spans a 3.3x spread in our catalog: Deepgram Aura-2 at $0.036, LMNT at $0.06, ElevenLabs at $0.12 with the category’s top quality score of 93. Voice quality is the most subjective rating we maintain, which is exactly why the price axis is the trustworthy one.
7. Transcription
The other half of the voice loop, and home to the single best bargain in the entire catalog: Groq’s hosted Whisper large-v3-turbo at $0.00084 per audio minute routed. Deepgram Nova-3 ($0.00516) and ElevenLabs Scribe ($0.00804) charge 6-10x more and earn higher quality scores (88 and 91) plus managed features. For drafts and search indexing, cheap wins; for captions and compliance, pay up.
What the whole toolkit costs
The surprise, when you put the seven side by side, is how lopsided the spend is:
| Tool | Cheapest routed price | Unit |
|---|---|---|
| Search | $1.20 | 1k queries |
| Scrape | $0.36 | 1k pages |
| Parse | $1.50 | 1k pages |
| Embed | $0.012 | 1M tokens |
| Code | $0.0000168 | vCPU-second |
| Speech | $0.036 | 1k characters |
| Transcribe | $0.00084 | audio minute |
For a typical research-style agent, search and scrape dominate the bill because of call frequency, parsing spikes with document-heavy workloads, and embeddings plus sandboxes are nearly free. Voice categories only matter if you ship voice, and then they matter a lot.
The other lesson from running failover across all seven: every one of these categories has at least two viable providers, and every provider has bad days. The toolkit is not seven picks; it is seven categories with a fallback in each. One key on route.tools covers all of them, with automatic failover and the exact price attached to every response.
If you are assembling the stack now, start with the side-by-side numbers on the search and scraping comparison pages and work down the list from there.