Announced 5 Oct 2026 · Sources checked
What did Together announce on 5 October?
The dated object is the Together AI blog post of 5 October 2026. Schema on that page gives datePublished 2026-10-05 and dateModified 2026-10-05T16:05:24.426Z. The authors listed there are Will Van Eaton and Hassan El Mghari. The lead claim is that Together Link “connects the harness your team already uses to the best open models on Together AI and cuts your spend by over 50%.”
The post’s supported list is Claude Code, Claude Desktop, Codex in the ChatGPT app and the CLI, OpenCode, and Pi. Settings and logins “stay as they were.” Going back to native closed models “takes one command.” Billing, it says, uses an existing Together API key on serverless pay-as-you-go or credit packs, with no separate contract.
This is a routing and packaging story, not a new open-weight release. For the license distinction the blog blurs when it says “open models,” see open weights versus open source. For how a token price becomes a bill, see how to calculate AI API cost.
What do the 9 October docs say you actually install?
The documentation page “Configure Claude Code, Codex, OpenCode & Pi Code with OSS models,” which we opened on 9 October at docs.together.ai/docs/togetherlink, is the source of truth for commands. It opens with a beta warning: commands, routing and the model list may change, and issues go to GitHub. Requirements are a Together API key; macOS or Linux with Bash and curl; and the target agent already installed. On Linux, installing Bun also needs unzip. Windows is not on that list.
The vendor install line is curl -fsSL https://link.together.ai/install | bash. The installer, the docs say, can install Bun, puts commands in ~/.local/bin, and opens an interactive launcher. togetherlink and the tlink alias open that launcher. Shortcuts on the same page are tclaude, tcodex, topencode, tpi, tclaude-desktop, and togetherlink chatgpt. We did not run the installer.
“How it works” is explicit: Together Link “configures each tool to connect directly to its hosted gateway.” “No local proxy or daemon runs, and your normal configuration files remain unchanged.” Terminal tools get a temporary per-launch config that is removed when the session ends. Claude Desktop uses a Together Link-owned third-party profile with backups. ChatGPT Desktop uses ~/.codex-togetherlink and, the docs say, never rewrites ~/.codex.
Which models and prices are on the docs table?
The docs table we copied lists Auto plus four Together models, each with a 1M-token context window. Prices are per million tokens, as input / cached input / output. The page says to run togetherlink models for the live table.
Those IDs sit next to other coding-model releases, including JetBrains Mellum 2.1, which is a different product: an in-IDE model, not a Together gateway. If you are choosing an assistant rather than a router, start from how to choose an AI assistant.
| Model | Model ID | Claude tier | In / cache / out per 1M |
|---|---|---|---|
| Auto (default) | auto | Routes per request | Billed at the chosen model |
| Kimi K3 | moonshotai/Kimi-K3 | Opus | $2.70 / $0.27 / $13.50 |
| GLM 5.3 | zai-org/GLM-5.3 | Fable | $1.40 / $0.26 / $4.40 |
| DeepSeek V4.1 Flash | deepseek-ai/DeepSeek-V4.1-Flash | Sonnet | $0.30 / $0.01 / $1.20 |
| Qwen 3.8 | Qwen/Qwen3.8-2.4T-A95B | None | $2.00 / $0.25 / $6.00 |
How does Auto routing work on the two pages we opened?
The 5 October blog says Auto “reads each session’s first task and sends it to the right model,” that routing “happens once per session, so prompt caching keeps working,” and that an Anthropic key routes between Opus 5.5 and GLM 5.3, otherwise between GLM 5.3 and GLM 5.3 Flash. After each session a tracker “shows what you spent next to what the same session would have cost on Opus 5.5.”
The 9 October docs describe a different unit. “For each request, Together Link’s cloud gateway classifies the prompt.” Straightforward requests go to Together models such as GLM 5.3. In Claude Code and Claude Desktop, a configured Anthropic key sends “more difficult” requests to Claude Opus, billed to that Anthropic account. Codex, OpenCode, Pi Code and ChatGPT Desktop “always stay on Together AI models,” as do Claude sessions without an Anthropic key.
We are not choosing between those two descriptions. The blog is the 5 October announcement. The docs page is what Together was serving on 9 October. If you pin a model, the docs say to put --main <id> before the tool name on the full togetherlink command. The tclaude-style shortcuts expand with the tool name first and cannot select a model that way.
What else is in the docs, and what did the blog add as color?
Desktop notes from the docs: Claude Desktop must live in /Applications or ~/Applications on macOS, or claude-desktop must be on PATH on Linux. togetherlink claude-desktop off switches back; reset removes the profile after confirmation. ChatGPT Desktop’s Together profile disables native web search. Image generation is listed as a skill in Claude Code, Claude Desktop and ChatGPT Desktop sessions, billed separately in togetherlink usage.
Headless examples pass through each agent’s own flags. The docs insist that headless Claude Code append < /dev/null so an open stdin does not hang. OpenCode must be version 2; Pi Code must be 0.80.8 or newer on Node.js 22.19 or newer. Together Link does not install those agents.
The blog’s infrastructure paragraph is Together’s: as of 30 September 2026 it says Together served the largest share of OpenRouter tokens for DeepSeek V4.1 Flash (40.8%), GLM 5.3 Flash (28.2%) and Kimi K3 (23.1%). Those shares are a vendor footnote, not a measurement we ran.
What did we not test?
We opened the 5 October blog, the 9 October docs page, and link.together.ai/llms.txt. We did not run the installer, save an API key, launch Claude Code or Codex through Together Link, or check a session bill against Opus 5.5. “Over 50%” remains Together’s line. An older personal GitHub repository that shares a similar name is not the official CLI described on these pages.
Common questions
Does Together Link run a local proxy on my machine?
Not according to the docs we opened on 9 October. That page says the tools connect directly to a hosted gateway, no local proxy or daemon runs, and normal config files stay unchanged. Terminal sessions use a temporary per-launch configuration that is removed when the session ends.
Can I use it on Windows?
Not on the requirements list we opened. The docs name macOS or Linux, with Bash and curl. Windows is not mentioned there.
Is the “over 50%” saving a measured result we checked?
No. It is the 5 October blog’s claim. The docs describe a status-line estimate against Claude Opus and a togetherlink usage command. We did not compare a session bill.
What to remember
If your team already lives in Claude Code or Codex and already has a Together key, Together Link is the 5 October beta that aims to keep the harness and change the model bill. Read the 9 October docs for the real command list, treat Auto’s timing as unsettled between the two pages, and do not run the installer on our say-so.
Sources & further reading
How this story was made
Written by Kristian Kostov with AI assistance and checked against the linked sources. Company performance claims are attributed to the company. Analysis reflects AiLookout’s interpretation; we have not independently tested the products discussed. Cover photography is illustrative and does not depict the specific announcement or product.
Our editorial standards
