Start with the part that surprises people: at the tiers most Gravity Forms jobs need, Claude and OpenAI cost the same. Anthropic lists Claude Haiku 5.5 at $0.10 per million input tokens and $0.50 per million output tokens, and Claude Sonnet 5.5 at $2 and $10 (Anthropic pricing). OpenAI lists gpt-6-luna at $0.10 and $0.50, and gpt-6.1-sol at $2 and $10 (OpenAI pricing).
So for most form work, price doesn’t decide it. What decides it is how each model handles your entries, which provider the client already has an account with, and what the output has to do next. Below are five jobs AI does in Gravity Forms builds, which tier to start each one on, and what it costs per 1,000 entries at list price. The token counts are assumptions, stated for each job. Your forms will differ, so treat every number here as an estimate and replace it with your own after the first week.
The prices side by side
| Tier | Claude model | Claude price | OpenAI model | OpenAI price |
|---|---|---|---|---|
| Small | claude-haiku-5-5 | $0.10 / $0.50 | gpt-6-luna | $0.10 / $0.50 |
| Mid | claude-sonnet-5-5 | $2 / $10 | gpt-6.1-sol | $2 / $10 |
| Large | claude-opus-5-5 | $4 / $20 | gpt-6-astra | $10 / $50 |
1. Classification
The job: read a submission and pick a label from a fixed list. Billing, bug or question. Spam or not. Which department. The answer lands in a drop-down or hidden field that conditional logic acts on.
Start on the small tier. Haiku 5.5 or gpt-6-luna. Picking one of five labels doesn’t need a large model, and at volume the difference adds up.
Estimate: 600 tokens in (instructions plus the ticket), 20 out (the label). About $0.07 per 1,000 entries on either provider.
How to choose between them: give both the same prompt with the allowed labels spelled out, run them on 50 past entries you’ve already labelled by hand, and count how often each one agrees with you. Whichever wins on your entries wins. Tell the model to answer with the label only, and check its answer against the list before you write it to the entry. Anything off the list goes to “needs review”.
2. Triage
The job: classification plus a short note for whoever picks it up. Category, priority, and two sentences on what the person actually needs. That’s one feed writing to three fields; your notification routing or Asana feed does the rest.
Start small, move up if the notes are thin. The label is easy. The summary is where a mid-tier model earns its price.
Estimate: 1,000 tokens in, 150 out. About $0.18 per 1,000 entries on the small tier and about $3.50 on the mid tier, on either provider.
3. Drafting
The job: a first-pass reply, a project brief, a cleaned-up bio. Text a person will read.
Start on the mid tier. Sonnet 5.5 or gpt-6.1-sol. This is where the difference between models shows most, and part of it is taste. Put your client’s tone guide in the prompt, read ten drafts from each, and keep the one that needs fewer edits.
Estimate: 1,500 tokens in, 500 out. About $8 per 1,000 entries on either mid-tier model. On the large tier, about $16 with Opus 5.5 and about $40 with gpt-6-astra.
Decide before launch who reads a draft before a customer does. For anything that goes out under the client’s name, a person should, at least until you’ve watched a few hundred go by.
4. Image generation
The job: a header image, a social card or a product mock-up from what someone typed into the form.
This one is OpenAI. Current Claude models read images but only output text (Anthropic’s model overview), so image feeds run on OpenAI’s image models.
Estimate: OpenAI prices image output by the token, $30 per million output tokens for gpt-image-2.5 at the standard tier, and the token count depends on the size and quality you ask for. Run the per-image calculator OpenAI links from its pricing page for the size you’ll actually use before you quote a client.
Put image feeds behind conditional logic so they only run when someone ticks the box. An image nobody asked for is the most expensive mistake on this list.
5. Approval review
The job: in Gravity Flow, before a reviewer approves an application, a model reads it and writes a summary and a recommendation into the entry. The reviewer still decides. They just don’t start from a blank page. In AI Feed this is an AI Text step (Premium) placed before the approval step.
Start on the mid tier, and consider the large tier when the decision is expensive. Applications are long and the reviewer will lean on the summary, so this is where a stronger model is easiest to justify. Current Claude models accept up to a million tokens of context, which covers any application form plus the documents pasted into it.
Estimate: 6,000 tokens in (the full application), 400 out. About $16 per 1,000 entries on Sonnet 5.5 or gpt-6.1-sol, about $32 on Opus 5.5 and about $80 on gpt-6-astra. Even the top figure is 8 cents an application, small next to what the reviewer’s time costs.
What matters more than the model
- A fixed output format. Labels from a list, a set number of sentences. Free-form output is where both providers drift.
- Conditional logic on the feed. The cheapest call is the one that doesn’t run.
- A budget and a ledger. Cap monthly spend per form and look at real token counts after the first week. Then throw away the estimates above and use yours.
- A failure plan. Decide what the entry looks like if the call fails, and who finds out.
Test on your own entries
The honest answer to “which model?” is: test on your own entries. In GravityOps AI Feed text feeds run on Claude, so you can put two feeds on the same form with the same prompt, one on Haiku 5.5 and one on Sonnet 5.5, each writing to its own field, and the usage ledger shows what each one cost. Keep the winner and delete the other feed. To compare against OpenAI’s text models, run the same 50 entries through OpenAI first; AI Feed uses OpenAI for images only.
If you’re still deciding where AI belongs in the build at all, start with AI belongs in the feed, not the form builder. For the other plugins on the market, see our comparison of every AI option for Gravity Forms.
Discuss this post
Questions, corrections or your own approach? Reply in the community thread.
