Right Tool, Right Job: You Still Pay to Read the Web

The frontier models crossed "good enough" for most business tasks a while ago, and enterprises have noticed. Over four quarters the market went from unlimited spend to routing every task to the cheapest model that can finish it. Stripe paying a reported $7.5 billion for OpenRouter is the clearest signal yet that model choice is now a process someone owns. Routing is the right instinct. But it only fixes half the bill, and the half it misses is the web.

Four quarters, one correction

Q4 '25, the honeymoon. Claude Code lands across the enterprise and spreads faster than anyone budgeted for.

Q1 '26, the immune response. CFOs start asking for justifiable ROI and cap the spend.

Q2 '26, the hunt for efficiency. "Tokenomics" and "routers" enter the enterprise vocabulary.

Q3 '26, the strategic signal. Stripe agrees to acquire OpenRouter, a gateway that routes token usage across 400 or so models from more than 80 providers. Sources put the price at $7.5 billion, against a $1.3 billion valuation three months earlier.

This exact story is playing out everywhere, and it's growing. It comes down to choosing the right tool for the right job.

Good enough is now the default

Need a quick pointer to the right part of your codebase? Want to bounce ideas for fleshing out a blog post? Looking for a quick translation of a note full of abbreviations and industry jargon?

These are solved by the majority of models out there.

There's no reason to reach for the McLaren every time you want to deliver cookies to your neighbor down the street.

Routing fixes the model bill, not the web bill

Here's what routing doesn't touch.

When your agent works on the web, you pay twice. Once for the model you routed to. Once for everything that model has to read.

The second bill is the one nobody put in the forecast. A page comes back with the navigation, the sidebars, the cookie banner, the ad slots and the trending widget attached, and your model pays full price to reason over all of it. We measured this on 45 URLs across five providers. One Daily Mail article came back through Fetch at 4,737 characters, 87% of it the actual article. The noisiest provider in the test returned 171,095 characters of the same page, 19% of it the actual article. Same article. Same answer at the end of it. Roughly 36 times the token cost to get there.

Then it fails, and you send the whole thing again.

You can route that workload to a model that costs a tenth as much and still be paying to read 167,000 characters of navigation menu. Junk context doesn't just fail to add signal. It reduces it.

Artificial Analysis scores it the same way. Their Search API leaderboard splits the cost of a task into two lines: what the search provider charges you, and what the model spends reading what came back. On the same benchmark, with the same answer model, the model line swings from $0.03 to $0.072 per thousand tasks depending only on which search provider fed it. That spread has nothing to do with which model you routed to.

For the record, TinyFish Search sits at $0 on the search line and $0.035 on the model line, the lowest total on the board as of August 31, 2026. Right tool, right job.

Web access shouldn't be coupled to your model

This is the thesis TinyFish has been building on.

The operations of the world continue to flow through the interface of the web. That infrastructure isn't going away. Agents acting on your behalf need reliable access to it, and that's what we've spent the last two years building: an access layer for your agents.

Operating the web doesn't require live, on-the-fly reasoning every single time you attempt the same task or click the same button on a page. It sure as hell doesn't require the knowledge to disprove math conjectures and theorize cancer cures. So don't pay that premium.

And that layer shouldn't be coupled to the model you choose. You want to switch between Kimi, GLM, Gemma or whatever ships next quarter. They all need the same reliable access to the web underneath.

Your agent shouldn't be worse at observing the web just because the task is simpler, just like a cheaper car shouldn't come with foggier windows.

What the right tool actually looks like

For us it means the web layer is priced and built as infrastructure, not as intelligence.

Search and Fetch are free. Not discounted, free, because reading the web is the most fundamental thing an agent does and fundamental things shouldn't meter. Web Agent handles the part that genuinely needs operation: logging in, working through filters and pagination, filling the form, completing the task. That's metered per step, because steps are real work.

Underneath, Mako is purpose-built for operating live sites rather than being a frontier model wrapped in scaffolding. It isn't trying to out-reason Claude or Gemini. It's trying to click the right thing, on a page that changed last Tuesday, ten thousand times in a row.

TinyFish Agent ships with this model included, so you don't pick the model inside the agent. If model choice inside the automation layer is the thing you care about most, that's a genuine trade-off and other vendors handle it differently. And you don't have to take the whole stack. Take Search as a tool inside your own harness, take Fetch in front of your RAG pipeline, or hand the whole job to Agent. One platform, one API key, and you decide how much of it you use.

Route your models. Route them aggressively. Just stop paying frontier prices to read a navigation bar.

Your agents need to operate the web. We're the best tool for that job.

Start building at agent.tinyfish.ai or read the Agent docs.