Inkling 2 Small on Workers AI: Pricing, Paid Access and 20/50 RPM Limits

What is available, and which capabilities should you check first?

Cloudflare announced on November 12, 2026 that Thinking Machines Lab’s Inkling 2 Small was available on Workers AI. This guide checks the official announcement, model page, pricing, limits and compatibility documentation on November 13 in Beijing time. It covers Cloudflare-hosted access.

Use the complete model name @cf/thinkingmachines/inkling-2-small in requests. The official model page describes text and image inputs, text output and function calling. Function calling means the model can request a tool invocation; the application still integrates and executes the tool.

The documented context window is 1,048,576 tokens: the amount of information a request can accommodate, rather than a maximum output length or a guarantee of long-document accuracy. We did not run inference and have not independently verified a region-by-region availability table, latency commitment or maximum output.

Choose standard billing or prepaid gateway credits first

This model requires paid access and cannot be called through standard Workers Free billing alone. Cloudflare provides two paths: upgrade to Workers Paid, or use prepaid AI Gateway credits. General Workers AI free-usage information does not replace this model’s access requirements.

For the prepaid path, have prepaid AI Gateway credits, set the target gateway’s Workers AI billing setting to Unified billing, and specify that gateway in the Workers AI binding or REST request. Changing only the model name, having a gateway, or holding another provider’s key does not satisfy these conditions.

A Workers AI binding is the interface through which a Worker calls AI; the official example uses env.AI.run(). The model page also provides REST and OpenAI-compatible access. This guide checks public documentation only. It did not purchase credits, connect an account or send billable inference requests for readers.

How do token prices and cached-input rates work?

The current Cloudflare price table lists $0.300 per million input tokens, $0.100 per million cached input tokens and $1.000 per million output tokens. Tokens are the model’s measurement units for inputs and outputs such as text. Do not read per million as per thousand or per request.

The cached-input price applies only to input that meets the platform’s cached-billing conditions. Sending the same text again does not guarantee a cache hit. The complete cache-eligibility rules, retention window and any per-request cache-hit guarantee for this model were not verified here.

As a hypothetical pricing example, suppose the billable usage is one million uncached input tokens and 100,000 output tokens. The table gives an equivalent model cost of $0.30 + $0.10 = $0.40. This is not a measured request bill. It excludes base-plan charges, any fees associated with buying credits, taxes and other product charges; check actual usage classifications and billing separately.

The pricing page presents equivalent prices in tokens and Neurons, the Workers AI usage-billing unit. These are alternative expressions of the same price. Do not add the two unit representations together for the same usage.

Why is the limit 20 or 50 requests per minute?

The paid-model limits apply per account, per model: 20 requests per minute with standard Workers AI billing, or 50 with prepaid AI Gateway credits. The limits documentation also requires loading prepaid credits and selecting Unified billing on the gateway to receive the higher limit.

The general Text Generation limit of 300 requests per minute explicitly excludes models that require Workers Paid. It therefore does not apply to Inkling 2 Small. The November 12 model announcement and the current limits body agree on the 20/50 limits. The limits page still displays September 17 as its update date, so this guide checks the actual body against the announcement.

A per-minute request limit is not a concurrent-request count, a response-time promise or a guarantee of sustained throughput. Long reasoning and tool loops can still involve multiple requests. Plan request pacing and error handling for the actual access path; changing a proxy exit does not replace account eligibility or platform limits.

Check four values when reusing an OpenAI client

The compatibility documentation specifies the base URL https://api.cloudflare.com/client/v4/accounts/{account_id}/ai/v1. Replace {account_id} with your own Cloudflare account ID and authenticate with a Cloudflare API token suitable for that endpoint. Use the full model ID @cf/thinkingmachines/inkling-2-small.

The compatible text-generation endpoint is /v1/chat/completions. In an existing client, check the account ID, Cloudflare token, base URL and model ID separately. Compatible request formatting does not mean you continue using an OpenAI account balance or OpenAI key. Keep real tokens out of public examples and troubleshooting records.

Do not assume an existing client.responses call works after changing its model name. The current documentation lists only two GPT-OSS models for /v1/responses, with non-streaming support. It does not list Inkling 2 Small for that endpoint. Check any other OpenAI-specific client features individually as well.

Reasoning effort and a minimal evaluation before rollout

Reasoning is enabled by default, with reasoning_effort set to high. The announcement lists alternatives including max, medium, low, minimal and none; none disables reasoning. Changing this request setting does not establish a particular improvement in latency, cost or quality.

In an existing eligible test environment, first check one nonsensitive text request through the chosen endpoint, then examine image-input or tool-call formats. Record the model ID, request path, reasoning effort, return status, actual usage and an answer a human can assess. This is a suggested evaluation procedure; we did not execute those inference tests.

If the application requires a fixed processing region, particular data-retention terms or strict response deadlines, verify the applicable formal product conditions before integration. These sources do not provide a universal region-by-region commitment for every account. Public model availability, documentation returning HTTP 200 and paid eligibility do not independently prove that an actual account request succeeds.