Together Link points existing coding agents at open models

Together AI published Together Link on 5 October 2026, by Will Van Eaton and Hassan El Mghari. The post says one install command points the coding harness you already use at open models on Together’s serverless API, and that going back to the native closed models takes one command. It names Claude Code, Claude Desktop, Codex in the ChatGPT app and the CLI, OpenCode, and Pi. Billing, it says, stays on a Together API key, as serverless pay-as-you-go or credit packs, with no separate contract.

This is a feature explainer from those pages and from Together’s current docs. It is not a test of the installer.

What the command sets up

The post and the docs both give this installer:

curl -fsSL https://link.together.ai/install | bash

That is a remote script piped into a shell. Read the script before you run it. The docs say it also installs Bun if Bun is missing, puts commands on ~/.local/bin, and is aimed at macOS or Linux. They say Together Link does not install Claude Code, Codex, OpenCode, or Pi. Those apps have to be present already. The docs call the current release a beta and say commands, routing, and the model list may change. A separate CLI guide calls it a private beta.

Launch commands on the docs page are togetherlink claude (tclaude), togetherlink codex (tcodex), togetherlink opencode (topencode), togetherlink pi (tpi), togetherlink claude-desktop, and togetherlink chatgpt. OpenCode must be version 2. Pi Code must be 0.80.8 or newer, which the docs say needs Node.js 22.19 or newer. The Claude Desktop command also covers Cowork. ChatGPT Desktop is marked beta on macOS and Linux. Terminal launches get a temporary configuration and, the docs say, do not rewrite the agent’s own config. ChatGPT Desktop uses a separate profile at ~/.codex-togetherlink.

The blog’s one-command return matches togetherlink claude-desktop off and togetherlink chatgpt off. For a terminal agent, the docs say to launch the original tool and stop using togetherlink.

Where Auto is described two ways

The blog says to set the model to Auto. It says the router reads the session’s first task, sends quick fixes to a fast, cheaper model and harder problems to a more capable one, and chooses once per session so prompt caching still works. With an Anthropic key, it says Auto routes between Opus 5.5 and GLM 5.3. Without one, it says Auto routes between GLM 5.3 and GLM 5.3 Flash.

The docs, opened the same day, describe a different timing. They say the cloud gateway classifies each request. Straightforward requests go to Together models such as GLM 5.3. More difficult requests, in Claude Code or Claude Desktop, and only if an Anthropic key is configured, go to Claude Opus and are billed to that Anthropic account. Codex, OpenCode, Pi Code, and ChatGPT Desktop, the docs say, stay on Together models either way. The CLI guide agrees that Auto classifies each request, names Claude Opus 5.5 as the complex route when an Anthropic key is present, and says utility calls such as compaction and token counting stay on Together-hosted models.

The same two pages disagree on Claude Code’s menu. The docs map Opus to Kimi K3 (moonshotai/Kimi-K3), Fable to GLM 5.3 (zai-org/GLM-5.3), Sonnet to GLM 5.3 Flash (zai-org/GLM-5.3-Flash), and Haiku to DeepSeek V4.1 Flash (deepseek-ai/DeepSeek-V4.1-Flash). The CLI guide maps Opus 5 to Kimi K3, Fable 5.1 to GLM 5.3, Sonnet 5 to DeepSeek V4.1 Flash, and Haiku 4.5 to MiniMax-M3 (MiniMaxAI/MiniMax-M3). Both say togetherlink models is the live catalog. The docs catalog opened on 5 October was Auto plus those four models, each listed at a 1 million token context. Pin a model with --main before the tool name (togetherlink --main zai-org/GLM-5.3 claude). The docs say a --model flag after the tool name is refused or ignored.

The savings Together claims

The headline and the summary say Together Link cuts spend by over 50 percent. The post does not show the workload, the prices, or the closed-model baseline behind that figure. Treat it as Together’s claim.

The post says a per-session tracker shows spend next to what the same session would have cost on Opus 5.5. The docs say each session prints a cost on exit, the Claude Code status line compares that spend with Claude Opus on the same tokens, and togetherlink usage --last 7d shows seven days of gateway spend.

Together also says that, as of 30 September 2026, it served the largest share of OpenRouter tokens for DeepSeek V4.1 Flash (40.8 percent), GLM 5.3 Flash (28.2 percent), and Kimi K3 (23.1 percent). The sentence is marked as a footnote, and the retrieved article has no footnote body. Those shares are Together’s claim.

Practical takeaway

The documented path is a Together API key, the curl installer above after you have read it, and a togetherlink launch of an agent you already installed. Auto is the default both pages describe. Rely on the blog if you need the once-per-session account, and on the docs if you need the per-request account, because they were both live on 5 October and they do not match. Prices and the menu mappings are behind togetherlink models. The over-50-percent saving is a vendor sentence without a published method.