Ploy, a website and marketing platform, launched an AI Gateway on November 19, 2026 and says it lets Ploy "complete bulk tasks up to 8x faster and 20x cheaper than Claude Code and Codex." The evidence is a single job on the company's own data, described in its launch post. This post reads the vendor's table closely, shows how the headline ratios were built and gives you a way to test the idea on your own workloads.
The idea behind the claim is sound and worth knowing. The numbers in the headline are less solid.
TL;DR: the claim and what supports it
| Claim | Source | Status |
|---|---|---|
| Up to 8x faster than Claude Code and Codex | Ploy's launch post | Holds against Codex only (8.7x); 3.8x against Claude Code |
| Up to 20x cheaper | Ploy's launch post | 22.6x against Claude Code's estimated cost; 6.8x against Codex's estimated cost |
| Same job for all three | Ploy | Self-run, single job, 300+ of Ploy's own sales calls |
| Rival costs | Ploy | Marked "est." in the table, method not given |
| Output quality equal | Ploy | All agreed on the top three use cases; quote counts differ and scoring is not explained |
| Independent replication | None found | Open |
What did Ploy launch?
Ploy AI Gateway lets Ploy call AI models directly. Its agent gives each item in a batch, such as a meeting recording, a blog post or an image, its own small AI workflow, runs hundreds in parallel and gathers the results. Ploy says it builds those workflows from one prompt, picks a model per step and bills against Ploy credits. The same gateway powers AI features on published Ploy sites, with no API keys for the site owner and a spend limit set on a usage screen.
The post lists building blocks with rough prices: language models at 0.01 to 0.4 cents per item, decision models at about 0.005 cents per call, images at 3 to 8 cents each, speech to text at about 0.4 cents per minute, web search at about 0.7 cents and a short video clip at about 30 cents. These are vendor figures for Ploy's own usage.
Ploy also offers a downloadable skill that works in any coding agent, provided you supply a key for the Vercel AI Gateway, a separate routing service. It covers picking models per job, running items in parallel, checking results and watching cost.
Why would per-item workflows be cheaper than an agent?
A ribbon passing through a window with its far end fading, showing an agent context window that keeps growing as it reads each call
Ploy's argument is that an agent reading 300 calls behaves like one patient person. Everything it reads stays in memory, so each next step is slower and costs more than the last. "By call 200, it re-reads a small library before it takes its next note." A workflow-per-item design reads one call and a few lines of instructions, then forgets it.
This matches how language-model APIs bill. Input tokens are charged on every request, and an agent loop resends the growing conversation each turn unless prompt caching absorbs it, which cuts the price of repeated prefixes but does not remove the growth. Caching, subagents and context summarization all reduce the effect, which is why harness defaults matter. We covered the mechanics in our Claude Code token-efficiency and prompt-cache guide, and the same idea shows up in AWS's AgentCore harness, which summarizes older turns and offloads large tool results.
The general lesson holds: for a bulk job where items are independent, a fan-out of small calls is often cheaper and faster than one long agent session. Whether it is 8x or 20x is the question.
The table, read line by line
Ploy gave itself, Claude Code and Codex one job: read more than 300 recorded sales calls and rank the use cases prospects like most, with counts and word-for-word quotes.
| Ploy | Claude Code | Codex | |
|---|---|---|---|
| Time | 4.5 min | 17 min | 39 min |
| Total cost | $2.87 | $65 (est.) | $19.40 (est.) |
| Read every call in full | Yes | Yes | Yes |
| Quotes found word for word | 18 of 18 | 23 of 25 | 48 of 48 |
Three things stand out when you do the division.
- The two headline ratios come from different rivals. Eight times faster compares Ploy to Codex (39 divided by 4.5 is 8.7). Twenty times cheaper compares Ploy to Claude Code ($65 divided by $2.87 is 22.6). Against Claude Code, the speed gap is 3.8x. Against Codex, the cost gap is 6.8x. No single competitor loses by both margins.
- Rival costs are estimates. Both the $65 and the $19.40 carry "est." and the post does not say how they were estimated, for example from token counts, plan pricing or billing data. Ploy's own $2.87 is described as including "Ploy's own work, not only the model calls."
- Quality is not comparable from the table. All three read every call and agreed on the top three use cases. But the quote counts differ widely: Codex returned 48 verified quotes, Claude Code 23 of 25 and Ploy 18 of 18. The numbers show how many returned quotes matched the source, not how useful the ranking was. Codex produced more than twice as many verified quotes as Ploy, which may reflect a more thorough answer.
What a fair test would need
A single run on a vendor's own data cannot establish a general ratio. A test you can trust would include:
- Several tasks, not one. Translation, alt text, scoring and summarization behave differently.
- Repeated runs, because agent time and cost vary a lot between attempts.
- The same models and settings disclosed for each tool, including whether subagents and prompt caching were on.
- Measured, not estimated, cost from billing records for the competitors.
- A quality rubric scored blind, such as accuracy of the ranking and the share of correct quotes.
- Independent hands. A third party running it removes the incentive question.
None of that is in the launch post. We found no independent replication, and the Hacker News submission of the post had 7 points and no comments when we checked, so there is no developer testing to report yet.
The other numbers in the post
Ploy's usage table lists six jobs with time and total cost: translating a whole site into 35 languages (6 minutes, $1.51), fixing alt text where 1,271 images were sorted and 360 described (12 minutes, $3.60), re-voicing 177 lines in 20 training videos (9 minutes, $2.21), building 573 lesson audio clips (25 minutes, $5.71), screening 693 research abstracts and reading the top 40 closely (2 minutes, $0.73), and scoring 240 website visitors (13 minutes, $6.21).
These are interesting as examples of what bulk automation can do at low cost, and they come without a competitor baseline, so they do not support the "faster than Claude Code" claim. They also come from Ploy's own sites and tasks. Treat them as plausible ballparks for similar jobs, not guarantees.
What this means for what you build or pay
For builders, the takeaway is design, not brand. If a task is a batch of independent items, write it as a fan-out: one small prompt per item, a cheaper model where quality allows, parallel execution and a check step. Reserve a long agent session for work that needs memory across steps. You can apply this in any coding agent without Ploy. For orchestration ideas, see how an agent harness pins tools and hooks and our look at a model-factory training and deployment loop.
For founders and marketers buying a tool, ask for a rerun on your own data and a measured cost per item before you sign. The cheap per-item prices are attractive for alt text, translation and lead scoring, but the comparison with general agents depends on a test you have not seen. Also check the credits model: the launch post says Ploy builds the workflow and bills Ploy credits, and it states the platform shows an estimate and waits for your go-ahead before a big job. That is a good control to confirm in the product.
How to check it yourself in an afternoon
- Pick one bulk task you actually run, such as summarizing 200 tickets.
- Run it three ways: in your current agent, as a script that makes one API call per item, and in Ploy if you use it.
- Record wall-clock time and the real bill for each, not an estimate.
- Score 20 outputs per method blind.
- Repeat once to see the variance.
If the fan-out is cheaper and equal in quality, you have your own number. It may be 3x or 30x, and it will be yours.
Related reading
- Claude Code token efficiency and prompt caching guide
- AgentCore harness on Strands: defaults, tools and hooks
- Prime Intellect Model Factory
- Higher-scoring coding models and merge rates
- GPT-6.2 Astra Sol power-user reviews versus benchmarks
- Kong Volcano Agent Cloud LavaFlow: claimed vs verified
Figures are as published by Ploy on November 19, 2026. Competitor costs are Ploy's estimates. Check current pricing and test on your own data.
