AI SDK 7.0.0: what breaks when you upgrade ai from 6.x

AI SDK 7.0.0 removes exports that earlier renames left behind as deprecated aliases, and deletes a long tail of internal types alongside them. Four experimental-prefixed functions, two names v6 already deprecated, and the types for streamText’s finish callbacks are gone or renamed. Anyone importing the old names from ai needs a typecheck to find every site before upgrading. Many more break code whose imports are already correct: four experimental_ option keys on the two generation calls, the two token counts v7 moved under detail objects, the property the response helpers take their stream on, the options object handed to a tool’s execute, and a middleware written against the v3 provider types.

What changed

• Two experimental-prefixed exports are gone outright. experimental_customProvider and experimental_generateImage are no longer published; customProvider and generateImage are the names.
• Two more keep working under the old name, for now. experimental_transcribe and experimental_generateSpeech are still exported in v7, but deprecated: transcribe and generateSpeech are what the documentation uses.
• Two names v6 already deprecated are deleted in v7. ToolCallOptions is gone; ToolExecutionOptions is the type a tool’s execute function receives. isToolOrDynamicToolUIPart is gone; isToolUIPart means what it meant.
• Two more are renamed with the old name kept. stepCountIs is isStepCount now, and CallSettings is LanguageModelCallOptions, which also split request-level options into a separate RequestOptions type.
• streamText’s finish callbacks were retyped and renamed together. StreamTextOnFinishCallback, StreamTextOnStepFinishCallback and TelemetrySettings are no longer exported; onFinish became onEnd and onStepFinish became onStepEnd in the same release.
• Two keys on the telemetry options object stop type checking. metadata and tracer under experimental_telemetry are both gone in v7, which renamed that type from TelemetrySettings to TelemetryOptions. These fire at the call rather than at an import, on generateText, streamText, generateObject, streamObject, embed, embedMany and rerank.
• The experimental_ prefix comes off four options on the generation calls. experimental_activeTools, experimental_output and experimental_context are gone from both generateText and streamText, and experimental_prepareStep from generateText, which is the only one of the two that ever declared it. Three of them simply lose the prefix. experimental_context becomes runtimeContext, and v7 types that value through the call rather than as unknown. experimental_output is gone from generateText’s result as well, where output is the read.
• The two token counts moved under detail objects. usage.reasoningTokens is usage.outputTokenDetails.reasoningTokens now, and usage.cachedInputTokens is usage.inputTokenDetails.cacheReadTokens. Both are matched on the usage and totalUsage of a generateText or streamText result rather than on the property name alone, so a provider’s own usage object beside them is not reported. Tokens written to the cache are a separate count at v7, cacheWriteTokens, which 6.0.282 did not report at all.
• createTextStreamResponse and pipeTextStreamToResponse renamed textStream to stream. The value is the same ReadableStream<string>, and status and response beside it are unchanged.
• generateObject and streamObject no longer take a timeout. v7 takes Omit<RequestOptions, 'timeout'> for those two and no other member of the options intersection declares the name, so the option is gone rather than moved. generateText and streamText keep it, which is why the two object calls are named one at a time.
• Two telemetry values are no longer copied onto each step. functionId and metadata are gone from StepResult. v7’s callId is not the same value: it identifies the generation call the step belongs to rather than carrying your grouping key. A step’s experimental_context is runtimeContext there too, matching the option.
• streamText’s result names its parts stream. The result declares stream at 7.0.0 and marks fullStream deprecated in its favour, with the same TextStreamPart type and the same order. Nothing is removed: fullStream is still declared at 7.0.105. streamObject’s result keeps fullStream with no marker, so a loop over that one is not a finding.
• A tool’s execute requires a context in its options. v7 takes its tool types from @ai-sdk/provider-utils 5.0.0, whose ToolExecutionOptions declares context as required; 6.0.282 took them from 4.0.51, which declared an optional experimental_context and no context. The SDK supplies the options for its own calls, so this lands on tests and wrappers that call execute themselves and build the literal by hand.
• LanguageModelMiddleware is declared against the v4 provider types. v7 aliases it to LanguageModelV4Middleware where v6 aliased LanguageModelV3Middleware. The seven members keep their names, but transformParams, wrapGenerate, wrapStream, overrideProvider, overrideModelId and overrideSupportedUrls now receive and return LanguageModelV4 values. v7 keeps specificationVersion optional and typed as a string on purpose, so a middleware declaring 'v3' still checks; what stops checking is code inside the middleware that names a v3 type or reads a call option or result field v4 no longer carries.
• A long tail of internal type exports, gone or newly deprecated. Telemetry hooks (TelemetryIntegration, bindTelemetryIntegration, registerTelemetryIntegration), prepareToolsAndToolChoice from ai/internal, OutputInterface, and the callback and event types the on*Event naming replaced are removed or marked deprecated. The table below names each one.

Before and after

The clearest deterministic patch is the experimental-prefix rename, on a file that calls four of them and re-exports one under another name:

import {
  experimental_customProvider,
  experimental_generateImage,
  experimental_generateSpeech,
  experimental_transcribe,
} from 'ai';

export const registry = experimental_customProvider({ languageModels: {} });
export const image = () => experimental_generateImage({ model: null as never, prompt: 'a cat' });
export const speech = () => experimental_generateSpeech({ model: null as never, text: 'hello' });
export const text = () => experimental_transcribe({ model: null as never, audio: new Uint8Array() });

export { experimental_transcribe as transcribeAudio };

After npx emendant fix, every import and every reference is renamed, including inside the re-export:

import {
  customProvider,
  generateImage,
  generateSpeech,
  transcribe,
} from 'ai';

export const registry = customProvider({ languageModels: {} });
export const image = () => generateImage({ model: null as never, prompt: 'a cat' });
export const speech = () => generateSpeech({ model: null as never, text: 'hello' });
export const text = () => transcribe({ model: null as never, audio: new Uint8Array() });

export { transcribe as transcribeAudio };

The other four deterministic patches are the same specifier rewrite: ToolCallOptions to ToolExecutionOptions, isToolOrDynamicToolUIPart to isToolUIPart, stepCountIs to isStepCount, and CallSettings to LanguageModelCallOptions.

Fourteen more rename a property rather than an import specifier. One of them rewrites fullStream to stream, and only where the receiver is what streamText returned, so the read on streamObject’s result beside it is left alone:

const result = streamText({ model, prompt: 'hi' });
for await (const part of result.fullStream) console.log(part);
const object = streamObject({ model, prompt: 'hi', schema: undefined as never });
for await (const part of object.fullStream) console.log(part);

After the patch the first loop reads result.stream and the second is unchanged. A pipeThrough or a tee over the same read is the same site and needs no other edit.

The largest group of property renames is the four experimental_ options, reported at the key in the options literal:

return generateText({
  model,
  prompt: 'hi',
  experimental_activeTools: ['weather'],
  experimental_output: undefined,
  experimental_prepareStep: () => undefined,
  experimental_context: { tenant: 'acme' },
});

After the patches, three keys have lost the prefix and the fourth has a new name:

return generateText({
  model,
  prompt: 'hi',
  activeTools: ['weather'],
  output: undefined,
  prepareStep: () => undefined,
  runtimeContext: { tenant: 'acme' },
});

The token counts move a path rather than a name, so usage.reasoningTokens becomes usage.outputTokenDetails.reasoningTokens and usage.cachedInputTokens becomes usage.inputTokenDetails.cacheReadTokens, on the usage and the totalUsage of both results. generateText’s result.experimental_output becomes result.output. The two response helpers take stream where they took textStream, and a shorthand property is rewritten as a pair: createTextStreamResponse({ status: 200, textStream }) becomes createTextStreamResponse({ status: 200, stream: textStream }).

Thirty-eight changes are assisted rather than deterministic, and two of them are worth showing. The first is the streamText callback types. Emendant reports the import of StreamTextOnFinishCallback, StreamTextOnStepFinishCallback or TelemetrySettings, but writing the replacement means typing the callback from the option it fills (onEnd or onStepEnd) rather than from a name that no longer exists, and that needs a named model provider.

The second is the tool options literal, which is reported at the literal rather than at an import, because the key that would be the site is the one that is missing:

await weather.execute({ city: 'Exeter' }, { toolCallId: 'a', messages: [] });

The value to add is the tool’s context, an empty object where the tool declares no contextSchema and otherwise a value of that schema’s type. toolCallId and messages stay as they are. An options object passed by name, or a literal holding a spread, is not reported, since the key may be there at runtime.

The other thirty-six assisted changes are the removed and newly deprecated exports the table names, and the option keys on consumeStream, experimental_generateVideo and the two internal helpers. Emendant reports the site and asks a model for the name that belongs there, because the name is documented by the release and is not in the two published packages the entry was derived from.

Migration guidance and its limits

Apply the twenty-two deterministic patches. Eight rename an import specifier and fourteen rename a property, which is an option key in a literal or a read on a result. The values either side are unchanged, so the call sites are correct once the name is, with one thing to check after: runtimeContext is typed through the call’s own runtime context parameter where experimental_context was unknown, so a context object that carried no type of its own may now need one.

The thirty-eight assisted changes need a model provider named on the command line or in emendant.json. Without one they are reported with their guidance and nothing is sent anywhere. Two of them are assisted because the edit means reading how your code builds an object: the callback types have to be written from the option they fill, and the tool context from the tool’s own contextSchema. The rest are assisted because the replacement name is documented by the release and is not in the two published packages, so the entry states the construct that broke and asks a model for what belongs there instead.

Read those patches before applying them. A deterministic patch is the same rewrite in every repository, from parameters a person curated; an assisted patch is one model’s answer to one prompt. Both are bounded to the site and both are proved the same way, in a copy of your repository moved to 7.0.0 and put through whichever of your own typecheck and tests it finds there, and a patch that fails is withheld. A repository declaring neither gets its patches graded structurally checked, which says the edits applied and the workspace matched again, and that nothing compiled or ran the code. Where the release documented no replacement at all, such as OutputInterface and the telemetry hooks, the guidance tells the model to decline rather than invent a name, so expect some of these to come back unfixed and stay yours to write.

Seven findings are report only, and each for its own reason. The two telemetry keys are the first: neither has a destination inside the call. A custom Tracer now goes to the OpenTelemetry constructor from @ai-sdk/otel and is registered once with registerTelemetry, which the migration guide places at application startup rather than in the file the finding is in. metadata has no replacement key at all; per-call attributes reach telemetry through includeRuntimeContext and includeToolsContext, which name properties of a runtime context rather than carrying a literal map.

timeout on generateObject and streamObject is report only because there is no option to rename it to. Pass an abortSignal from a timer instead, which both releases accept.

StepResult’s functionId and metadata are report only for the same reason as the telemetry keys: v7 keeps no per-step copy of what you passed in. Read them from the telemetry settings you handed to the call, which still declares both.

LanguageModelMiddleware is report only for a third reason: the edit is inside your middleware’s own body, wherever it names a v3 type or reads a field the v4 call options and results dropped. The finding says a value is built against the type, not which members that value names, which is why its confidence is medium and no patch is offered.

This page needs emendant@0.3.0 or later. The entries behind six of the rows below use a matcher kind or a receiver no earlier client reads, and an older CLI skips the whole file for this release with a coverage warning rather than reporting part of it.

This release also renamed the system prompt option to instructions, and moved onRerankFinish and onEmbedFinish to onRerankEnd and onEmbedEnd. Neither is a name the table below covers, so nothing here detects them. A clean scan says a file uses none of the constructs below, nothing more.

What Emendant detects, fixes and verifies

67 changes in ai 7.0.0. 22 patched by a transform, 38 patched only with a model provider you name, 7 report only. Every patch is proved in a copy of your repository before it is offered, and its header names the checks that passed. This table is generated from the feed entry, so it cannot claim more than the entry does.

emendant explain <change-id> prints any row's entry, guidance and sources at the terminal.

Coverage

Snapshot

2026-09-20.1, sequence 27, signed 20 September 2026

Minimum CLI

emendant@0.3.0

Feed entry

feed/npm/ai/7.0.0.json

emendant feed status shows the snapshot your machine holds, and emendant feed update fetches the latest.

Primary sources

Every claim above was checked against these, each pinned to the release rather than to a default branch.

• Migration guidevercel/ai at ai@7.0.0
• package.jsonvercel/ai at ai@7.0.0
• ai@7.0.0npm registry, the published package
• ai@6.0.258npm registry, the published package
• ai@6.0.279npm registry, the published package
• ai@6.0.280npm registry, the published package
• ai@6.0.0npm registry, the published package
• Changelogvercel/ai at ai@7.0.0
• ai@6.0.282npm registry, the published package
• ai@7.0.105npm registry, the published package
• @ai-sdk/provider-utils@5.0.0npm registry, the published package
• @ai-sdk/provider-utils@4.0.51npm registry, the published package
• @ai-sdk/provider@4.0.0npm registry, the published package
• @ai-sdk/provider@3.0.16npm registry, the published package

Scan your repository

Runs locally, reads your lockfile and source, sends nothing anywhere.

npx emendant scan

Then npx emendant fix writes the patches the table says exist. The getting started guide covers the flags and the patch grades.