Problem

Every Anthropic request on the unified-billing run path (env.AI.run) fails as soon as it has a system prompt:

7003: Model execution failed (User Input Error): Invalid value at system: Invalid input: expected string, received array

@ai-sdk/anthropic always serializes system as an array of text blocks ([{ type: "text", text }]), which Anthropic's own API accepts. The run path validates system as a string only. Since runWireFormat: "anthropic" sends the AI SDK's native body through makeRunFetch verbatim, the request is rejected.

This breaks Think's string-model routing for every Claude model (getModel() { return "anthropic/…" }), because Think always sends a system prompt.

Repro (Worker with an AI binding, workers-ai-provider@4.0.0):

const provider = createWorkersAI({ binding: env.AI, providers: [openai, anthropic] });
await generateText({ model: provider("anthropic/claude-sonnet-5.5"), system: "Be terse.", prompt: "Say ok." });
// → 400, 7003 "Invalid value at system: … expected string, received array"

Fix

In makeRunFetch, when the run wire format is anthropic, join a text-only system array into a single string before calling binding.run. Bodies with non-text blocks are left alone. The gateway path (gw.run, provider-native endpoints) is unchanged, because it accepts arrays.

Block-level cache_control can't be expressed in a string, so it is dropped on this path. Those requests were rejected outright before, so this is strictly an improvement. If the run path later accepts arrays, this can be removed.

Testing

• Unit tests: run path flattens (including a block with cache_control), gateway path leaves the array untouched, and helper edge cases (string, absent, non-text blocks). vitest: 22 files / 460 tests pass; tsc --noEmit clean.
• Live, against the real gateway with the patched dist: anthropic/claude-sonnet-5.5 and anthropic/claude-sonnet-5, with a system prompt and a tool, now succeed via both generateText and streamText (all previously failed with 7003).

Changeset included (workers-ai-provider: patch).

🤖 Generated with Claude Code

https://claude.ai/code/session_01HX9h9yQ3A7rkgDtZzotEPx