Ling 3.1 Flash: API Pricing, Tasks and Model Choices

For an application, create an account with the chosen provider and keep its API key in your server environment. For repository work, OpenCode Zen lists Ling 3.1 Flash in its coding-client model menu. The client supplies file access and runs tools; the model supplies text and tool-call requests.

Vercel’s September 30 announcement gives October 13 as the end of the offer. Select the regular ID only when you are prepared to use a paid route afterward. Select the -free ID when the intended behavior is to stop instead. The two names and their cutoff behavior below apply to Vercel; use OpenRouter’s own billing terms for its route.

In a Node.js project with the AI SDK installed, set AI_GATEWAY_API_KEY in the server or terminal environment and run the example below. It uses Vercel’s -free ID and selects Novita, the listed provider. Save the returned text and usage; this request only asks for a verification plan and gives the model no repository or shell access.

After the text request succeeds, use the same model in a tool-capable client for one testable issue. Require the patch and actual test output. A 4,096-token generation limit keeps this first response bounded, but reasoning also consumes output tokens, so an exhausted budget can end before the final answer. Increase it only for a task that needs the extra generation.

A free trial can answer whether Ling handles your task, but it cannot establish your ongoing invoice. On October 7, Vercel’s announcement does not give its post-promotion rate. Artificial Analysis’s current provider page lists $0.30 input and $0.90 output per million tokens and labels the provider InclusionAI. Those evaluator-recorded rates are separate from the promotional routes and do not establish Vercel’s future quote.

The comparison below uses 1,000 requests, each with 20,000 uncached input tokens and 3,000 billed output tokens. That is 20 million input and 3 million output tokens in total. Include reasoning in billed output, and add tool fees, retries and cache storage when the workflow uses them.

For text-only code and tool work, compare Ling with DeepSeek V4.1 Flash and GLM 5.3 Flash first. Add Gemini 3.5 Flash-Lite when the job is lightweight extraction, translation or intake from audio, video and PDFs. The choices below follow the documented inputs, controls and billing conditions; they are starting points for evaluation rather than a ranking of coding quality.

Token rates alone can miss a large difference in generation. Artificial Analysis reports about $0.99 per weighted Intelligence Index task and 220 million output tokens across its Ling 3.1 Flash evaluation. That task-cost result belongs to its weighted evaluation protocol. It is a reason to record total generated tokens when testing Ling, not a dollar quote for your coding ticket or Vercel’s free route.

Keep the repository, prompts and acceptance checks identical when comparing models. Record billed input, cached input, all billed output, tool calls, retry count and elapsed time. Divide total spend, including failed runs, by accepted results. A short final answer can still come after a long reasoning trace or several unsuccessful tool turns.

For repeated prefixes, DeepSeek’s direct cache-hit rate is $0.006 per million tokens during peak periods, while Z.AI lists $0.03 for GLM 5.3 Flash. Google lists $0.03 for Gemini 3.5 Flash-Lite cache reads plus $1 per million cached tokens per hour of storage. The invoice example above uses no caching. DeepSeek’s off-peak rates and Google’s Batch/Flex tiers are separate discounts, not the Standard/peak rates shown there.

MindStudio’s October 1 hands-on report describes an approximately hour-long Blender-to-Godot attempt. Ling created files and a runnable panel, but the author found misoriented assets and no coherent playable level. The same article reports a usable single-prompt SVG result.

For a visual build, open the result and test the intended interaction before accepting the task. A generated file, successful tool return or long uninterrupted session is not enough. The MindStudio article does not provide the exact provider, full settings or downloadable experiment artifacts, so use it to choose what to inspect rather than predict the outcome of your own client.

The current Vercel and OpenRouter routes list a 262,144-token context window and a 32,768-token maximum completion. Use those limits for requests on these services. For long-document extraction, require the source passage next to each important value and check it in the original; check retrieval accuracy on the passages you need. Novita also advertises 1M context elsewhere on its page, and Artificial Analysis lists 1M for InclusionAI’s route; that does not enlarge the current Vercel or OpenRouter request limit.

The checked Novita route accepts text and returns text. Use a multimodal alternative when the source is a screenshot, scanned page, audio recording or video. Generating code for a website or animation does not mean the endpoint can inspect images.

OpenRouter supports tools and tool_choice for Ling, but lists response_format as unsupported. If you need structured ingestion, validate returned JSON in your application and bound the repair loop, or choose an endpoint with the required structured-output contract. GLM’s documented JSON object mode also needs application schema validation; JSON syntax and strict schema enforcement are different guarantees.

Vercel describes Ling 3.1 Flash as a 560B-total, 25B-active hybrid-reasoning model. Those figures are its launch description. The API route is the available starting point here; no public 3.1 checkpoint or model-specific weights license was established by the sources checked on October 6.

For a downloadable Ling checkpoint, the separate Ling 3.0 Flash guide covers its MIT-licensed weights and reported runtime setups. Its 124B-total/5.1B-active architecture, quantizations and local-memory results belong to 3.0. Keep the exact version when deciding between an API trial and self-hosting.