New on the blog: why AI belongs in the feed →
View Categories

Text Feed Settings

Text feeds are created under Form Settings -> AI Text.

Provider #

Chooses the AI service that runs the feed: Anthropic (Claude) or OpenAI (ChatGPT). Each provider uses its own API key from the Credentials tab. Changing the provider limits the Model list to that provider and selects its recommended model. Feeds created before this setting existed keep using Anthropic.

OpenAI text feeds use the Responses API. Requests are sent with storage turned off, so OpenAI does not keep the response for later retrieval; OpenAI’s standard abuse-monitoring retention still applies under your OpenAI account terms.

The same feed settings, output modes, triggers, and execution sources work with either provider. The differences are in the advanced controls:

  • Effort is available for both, on Pro and higher. OpenAI models also offer None (no reasoning) where the model supports it.
  • Thinking and Thinking Budget are Claude controls and are hidden for OpenAI models.
  • Speed: Fast on OpenAI uses the fast service tier, which costs more and must be enabled for your OpenAI project. The ledger records the speed the provider actually used.
  • OpenAI counts reasoning toward Max Tokens. If reasoning uses the whole limit the feed fails with no answer, so raise Max Tokens or lower Effort.

With JSON multi-field mapping, OpenAI models listed in the plugin use strict structured output. A custom model ID falls back to a prompt that asks for JSON.

OpenAI-compatible provider #

On Premium and higher, OpenAI-compatible provider appears in the Provider list once an enabled OpenAI-compatible profile exists on the Credentials tab. The Model list then shows the models from those profiles, grouped by profile, and the model you choose also selects the profile.

Effort, Thinking and Speed are not offered for these models. JSON multi-field mapping is available only when the profile has Structured output or JSON mode enabled. See Provider Profiles.

Fallback Model #

Requires Premium or higher. Fallback Model names a second model, on any provider or profile, to use when the first provider is unavailable.

The fallback runs only after one of these failures:

  • the request timed out or the provider could not be reached;
  • the provider answered with a server error (HTTP 5xx) or said it was overloaded;
  • the provider rate limited the request (HTTP 429);
  • a streamed answer ended before it was complete and no text had been shown yet.

It never runs after a refusal, a content-filter block, a token-limit stop, an invalid request, a rejected key, an account that is out of credit, or a token cap. Those would be the same decision on any provider.

What to expect:

  • Each provider is tried once. The plugin does not retry the first provider before falling back, and a fallback has no fallback of its own.
  • The fallback uses its own provider’s default effort, thinking and speed. The feed’s Effort, Thinking and Speed settings apply to the first model only.
  • Both attempts are recorded in the ledger. The failed first attempt has the outcome provider_failed_fallback; the second row shows the provider and model that answered. Both count toward token caps.
  • With a streaming button, a fallback is not used once text has started to appear, because the answer would start over.
  • When the feed uses JSON multi-field mapping, the fallback must also support it. This is checked when the feed is saved.
  • A fallback on a disabled provider or a retired model is skipped.

Model lifecycle #

Claude Sonnet 5.5 is the recommended default for new feeds. Existing selections are preserved. Older Claude models that Anthropic still lists as active may remain selectable with the plugin’s Superseded label and a recommendation to compare their newer replacement’s behavior, capabilities, and pricing. Legacy is reserved for an official provider status; the plugin does not currently apply it to any selectable Claude model.

Haiku 5.5 is available for new feeds. Haiku 4.5 remains supported and selectable as Superseded, with Haiku 5.5 as its suggested replacement; saved feeds keep their selected model and settings.

Claude Fable 5.2 is available for new feeds. Fable 5.1 remains current; Anthropic did not mark it superseded or retired.

Deprecated models are retained only for feeds that already use them. A retired model blocks every execution path until an administrator saves a supported replacement.

Effort and Thinking #

Effort, Thinking, Thinking Budget and Speed require Pro or higher. The free version does not show them and always uses the provider’s defaults for the selected model.

The controls and their choices update immediately when the model changes, including recognized Custom Model IDs. Unsupported explicit settings reset to supported defaults, with a visible explanation. Initial rendering preserves invalid saved values for correction. The server validates these combinations again before generation, including feeds imported or edited programmatically. Unknown custom IDs use provider defaults and have no reviewed advanced controls.

Provider default omits the corresponding request parameter. The page shows the reviewed model defaults. Opus 5.5 and Haiku 5.5 default to medium effort; Sonnet 5.5 defaults to high. Where supported, explicit effort levels are Low, Medium, High, Extra high (xhigh), and Maximum (max). Higher levels may increase latency and token usage. Haiku 4.5 has no effort control; Sonnet 4.6 does not accept Extra high.

Thinking choices depend on the model:

  • Opus 5.5 and Fable models support adaptive thinking only, which is also their default. The Thinking control is hidden unless an incompatible saved value needs correction.
  • Sonnet 5.5 supports adaptive or Between tools only, which produces no upfront thinking in these single-turn feeds without tools. Between tools only requires low, medium, or high effort.
  • Opus 5 and Haiku 5.5 permit disabled thinking only with low, medium, or high effort. Haiku 5.5 defaults to adaptive thinking, supports all five effort levels, and has no manual thinking budget or Fast mode.
  • Sonnet 5, Opus 4.8/4.7, and Sonnet 4.6 support adaptive or disabled thinking. Older models may default to disabled. Deprecated manual thinking on adaptive-capable models is not offered.
  • Haiku 4.5 supports a manual thinking budget. The budget must be at least 1,024 tokens and strictly less than Max Tokens. Recognized older custom IDs follow their reviewed control contract.

Thinking also hides when the selected effort leaves only adaptive thinking and it is the model’s default. Choosing a compatible effort makes the alternative modes available again.

Adaptive thinking may skip reasoning for simple requests. Thinking uses part of Max Tokens and is included in billed output tokens. The plugin does not display thinking summaries to visitors.

If thinking spends the whole Max Tokens limit, no answer text is left and the feed fails; a frontend visitor sees the generic error, and the usage ledger records why. This is most likely with a low Max Tokens and Thinking left at Provider default. For frontend button feeds, set Thinking to Disabled or Between tools only (no upfront thinking), or raise Max Tokens.

Speed #

Fast mode is an explicit opt-in, off by default. It is generally available for Claude Fable 5.2 and Opus 5.5, and in research preview for Opus 5. Opus 4.8 fast mode is deprecated and removed on November 30, 2026; feeds saved with Fast on that model must switch to Standard. It requires provider account access and has premium pricing. Access errors are reported as failures. Requested speed and provider-reported actual speed are recorded separately; actual speed may differ or be unavailable. Historical ledger rows remain unknown.

See the official thinking, effort, and fast-mode documentation for provider behavior and pricing.

Max Tokens #

The maximum number of output tokens the provider may return. Tokens are pieces of text rather than words; one word can use one or several tokens. A larger value allows a longer answer and can cost more. It does not reserve or guarantee that much output.

Enter a positive whole number. A blank value uses 4,096. For listed models, the plugin rejects a value above the known model output limit. Recognized custom model IDs and pinned aliases use the same reviewed ceiling. For an unknown custom model, consult the provider because the plugin does not know its ceiling.

Some models always reason before they answer, and Claude Fable 5.2 and Claude Fable 5.1 cannot be asked not to. That reasoning is charged against this same limit, so a value sized only for the answer can be spent before the answer begins. Allow headroom on those models rather than trimming to the shortest value that once worked.

Choose a limit suited to the task. A category label may need tens of tokens, a short summary a few hundred, and a long draft considerably more. A JSON multi-field response needs room for the whole object, including every property name and value. If Anthropic reaches Max Tokens before finishing, the response is incomplete and cannot be applied safely. The plugin records a max_tokens diagnostic under GravityOps -> AI Feed -> Ledger; raise the limit and retry.

System Prompt #

Optional standing instructions describing how the AI should behave: role, tone, rules, boundaries, audience, and output style. Merge tags are supported, although stable instructions usually belong here and entry-specific content usually belongs in User Prompt.

Example: You are an internal support assistant. Be concise, do not invent facts, and do not include personal contact information in the summary.

User Prompt #

The required task and content for this request. Use the merge-tag button to insert form values.

Example: Summarize this request in three bullets for the support team: {Request details:5}

Anything inserted through a merge tag is sent to Anthropic. Do not insert {all_fields} without understanding how much and what kind of entry data it includes.

Entry Note #

Available on Pro, Premium, and Agency plans. Adds the AI output to the entry as a note, so the result is kept on the entry even when no field is changed.

  • Off adds no output note. This is the default.
  • Add a note in addition to the field output writes the fields as usual and also adds a note.
  • Note only (no field output) adds a note and writes no field. Output Mode, Target Field, Output Map, and Existing Values are hidden because the feed has no destination.

Notes are authored by “GravityOps AI Text Feed”, contain plain text with line breaks, and are only ever added. An administrator rerun adds a new note that starts with “Rerun” and leaves earlier notes unchanged.

A note-only feed cannot use a frontend-button trigger, because a frontend button needs a field to show its result in. With Add a note in addition to the field output, a frontend button keeps the generated text on the server and adds the note when the form is submitted. The note records the text as generated, even if the visitor edits the field before submitting.

If the note cannot be added, the field output is still saved and the failure is recorded in the ledger. A note-only feed fails in that case.

Note Template #

Shown when Entry Note is on. Sets the text of the note. The default is {output}.

  • {output} inserts the AI output.
  • With JSON multi-field mapping, {output:name} inserts one generated value by its response name, and {output} lists every value as name: value, one per line.

Tools #

Premium or higher. Lets the model call tools on an MCP server while it answers. Select the tools from the servers added under GravityOps -> AI Feed -> Credentials; leave empty for a feed that only writes text.

A feed with tools runs after form submission only, always in the background, and cannot use a frontend button, JSON multi-field mapping or a Fallback Model. Rerunning it repeats its actions. See MCP Servers And Tools.

Output Mode #

  • Single target field saves the complete response to one selected Gravity Forms field.
  • JSON multi-field mapping asks the model for named, typed values and saves each through the selected destination field’s storage rules.
  • Actions only is for a feed with Tools whose purpose is what the tools do. No field is written; the model’s closing summary is added to the entry as a note. Premium or higher.

Target Field #

Shown for single-field output. Choose the field that receives the generated text. The selector lists only fields that can store generated output, so a field type that cannot receive it is not offered; the same check runs when the feed is saved and again before generation. Text, Paragraph, and Hidden destinations require Pro or higher. Other supported field types require Premium or higher and are shown with an upgrade explanation when the current plan cannot use them.

Regular Paragraph Text fields receive plain text. HTML output sent to one of them is flattened rather than dropped: paragraphs and line breaks become blank lines and newlines, entities are decoded, and the tags are removed. If Use the Rich Text Editor is enabled on a Paragraph Text field, the field receives sanitized HTML: output that already contains HTML tags is kept as written, and Markdown output is converted first. Either way the result passes through wp_kses_post(), so unsafe markup such as scripts is removed. Write the prompt for one format or the other; output that mixes HTML tags with Markdown syntax is treated as HTML, leaving the Markdown characters visible.

Output Map #

Shown for multi-field output. Add one row per generated value:

  • Response name is the exact JSON property requested from the model.
  • Destination field is the supported whole Gravity Forms field that receives the value.
  • Output instructions describe the content required for this one destination.

Keep common context and behavior in the System and User Prompts. Put type-specific or destination-specific requirements in Output instructions so fill-empty execution can remove both an already-populated property and its instructions from the provider request. A destination can be used by only one row; after it is selected, the other rows disable that option.

Existing Values #

Shown for after-submission text feeds and used by synchronous, asynchronous, and Gravity Flow execution:

  • Fill only empty targets asks the provider only for destinations that are logically empty. This is the default for a new feed and can avoid unnecessary provider usage.
  • Overwrite configured targets requests and replaces every configured destination.

An administrator rerun always force-overwrites, regardless of this setting. Frontend generation has no saved entry yet, so Existing Values does not apply there.

Execution Mode #

Shown only for after-submission feeds.

  • Synchronous waits for generation during submission processing. Use it when the generated value must be available to an immediate confirmation or another immediate operation.
  • Async queues generation after submission. Use it for slower requests when the visitor does not need the result in the immediate confirmation.

Notification Options #

Premium/Agency settings can opt completion/failure events into administrator reruns and Gravity Flow runs. Normal post-submission feeds already emit their feed event. See the notifications and reruns reference.

Stuck? Ask in GravityOps Help #

Search past answers or post your question, tagged "ai-feed".