案例研究：Descript 如何借 OpenRouter 让新模型接入不再排队等工程师

Descript’s video editing agent, Underlord, runs on frontier models, and the set of models it picks from changes every few weeks. Getting a newly released model through evaluation used to take more than a week, and the testing itself was not the slow part. With OpenRouter as its production inference layer, Descript now evaluates models multiple times a week, runs 13 in production, and took a recent Anthropic release from announcement to live in a couple of hours.

Results

• Industry: Video and audio editing software
• Headquarters: San Francisco, California, United States
• Employees: 125+

About Descript

Descript is an AI-powered video and audio editing platform. It combines transcript-based editing with AI tools for audio enhancement, voice cloning, and generative media, so teams can produce studio-quality content without specialized expertise. Enterprise customers in healthcare, financial services, media, and technology use Descript to scale content production while keeping human control over every creative decision.

Underlord, Descript’s video editing agent, pairs LLMs with a harness that tells the model how to act and how to edit video.

Every new model waited for an engineer

Aleks Mistratov, who heads AI product at Descript, says Descript specializes in talking-head video. That covers YouTube creator video, marketing video with someone speaking, and video podcasts. The common thread is narrative, transcript-driven content, and public benchmarks do not answer how well a model handles it.

Descript maintained three direct integrations, with OpenAI, Anthropic, and Google. The team wrote and owned the fallback logic between them, and that logic worked by swapping models. If a GPT request failed, an Anthropic model answered instead.

The technical friction was parameter drift. Each model developer expected information passed a different way, and every new release required checking which parameters were now missing, which were now required, and which had become unsupported. Bringing a new model online cost at least a couple of hours, most of it spent debugging why something was not working.

The organizational friction was larger. The person who wanted to test a new model was not the person who could add it, so testing meant finding time on an engineer’s calendar. Days would pass before anyone got to it. The work itself was always a couple of hours. Getting those hours onto an engineer’s queue turned a question about a promising model into a week-long wait before anyone knew whether it was worth pursuing.

From one model to 13 in production

Descript was serving essentially one model when it decided to try Claude, and OpenRouter is how the second one arrived.

“We were just serving one model, basically. And then when we started to serve a second one, we also realized that we could add 10 more for basically very little work.”

Aleks Mistratov, Head of AI Product, Descript

Production as of August 2026 runs 13 models from four model developers.

• Anthropic: Fable 5, Opus 5, Opus 4.8, Opus 4.7, Opus 4.6, Sonnet 5, Sonnet 4.6, Haiku 4.5
• Google: Gemini 3.1 Pro, Gemini 3.5 Flash
• OpenAI: GPT-5.4, GPT-5.5
• xAI: Grok 4.5

Five of those Anthropic models are top-tier versions running at the same time.

Production went from three direct integrations to one connection through OpenRouter. Descript still keeps a direct Anthropic connection because it is an early Anthropic partner and some early-access models are available only through that API. That path is not production.

Underneath each model sits a choice of inference provider, and Descript configures that per model. OpenRouter exposes throughput, uptime, and cost for each provider, and Descript uses those to set priority order. One provider comes first for Sonnet 5, and Gemini uses a different order. Data retention policy is part of the same decision, and Descript routes only to inference providers that are within its policy.

That is also where the second kind of fallback lives. Descript’s old logic swapped models. OpenRouter’s provider fallbacks swap inference providers and keep the model the same, so a failure at one host does not change which model answers.

The team brings its own key to Baseten for a dedicated deployment of an open-weight model. OpenRouter routes to Baseten as the primary provider for that model, with other providers configured as fallbacks for the same model, all of them inside the same data policy constraint. Descript keeps the direct commercial relationship and its own key, and still has cover against single-provider rate limits and outages.

Evaluation stopped being a project

Descript now evaluates models multiple times a week. Most of them never reach production, which is the point.

Mistratov describes the pattern. He is on a walk, sees an announcement post that a new model has shipped, and posts into Slack asking for evals on it, with a link to the model on OpenRouter. An hour or two later the evals have run in Descript’s own harness.

“I’ve never even touched a keyboard for that to happen.”

Aleks Mistratov, Head of AI Product, Descript

The mechanism is Claude Tag, Anthropic’s Slack integration, which Descript set up in a channel with a set of skills for this loop. It runs the evals, opens a pull request adding the model to Descript’s harness, and opens a second one putting the model behind a feature flag. A human reviews and approves.

OpenRouter removed the per-model integration work that had been putting these evaluations on an engineer’s queue. Descript’s own agent and harness automate everything around that interface. The work no longer needs to be assigned to anyone.

On a recent Anthropic release, announcement to live took a couple of hours. The steps that remain are Descript’s own. They run internal evals, decide whether the model beats the incumbent, configure the model picker, set the feature flag, and deploy. David Dodero, who works on the agent team’s harness and evals, noted that most of that remaining work comes from how Descript’s code is structured rather than from the connection to the model, and that it would be easier if that were set up differently on their end.

What weekly evaluation turned up

Keith Simmons, the technical lead of Descript’s AI Enablement team, had been watching frontier-tier model costs become an issue. Open-weight releases were arriving fast enough that the only question was whether any of them held up against Descript’s own test cases. Running that question multiple times a week is how the open-weight model on the Baseten deployment surfaced. Simmons puts Descript’s total model spend at under half what it was, in part because of that process, and says quality held against Descript’s own test cases.

Simmons initially read routing as work Descript should not be spending engineering time on. That has changed.

“It’s a perfect outsourcing thing, how to route between models. But with this explosion of open-weight models, the ability to easily integrate new models has become critical to our business.”

Keith Simmons, Technical Lead, AI Enablement, Descript

Provider failures land in the logs

Fallbacks fire and the team finds out afterward by reading logs, or does not find out at all.

In the OpenRouter logs, Mistratov can watch traffic move. The primary provider serves during slow periods, 429s appear during spiky ones, and traffic then moves to plan B, plan C, and plan D with minimal disruption to what users get. When someone posts a screenshot of a Claude status page in Slack asking whether Descript is affected, the answer is no, because the Claude traffic moves to Vertex AI as another inference provider. Simmons notes that the team cannot use Claude Code during those windows, while Underlord keeps working.

Between incidents the pattern holds.

“The on-call for the agent team is extremely quiet. There are built-in default fallbacks and we don’t have to do anything. Our internal models are way more noisy. We spend a lot more time maintaining them, but anything that goes through OpenRouter is rock solid.”

Keith Simmons, Technical Lead, AI Enablement, Descript

This was true even while Descript was standing up the Baseten deployment.

“When we were rolling out this dedicated deployment on Baseten there were some config issues and they were throwing an exception, but OpenRouter just fell back to another provider. We didn’t know it was happening. None of our users complained because they didn’t see any of the issues.”

Keith Simmons, Technical Lead, AI Enablement, Descript

The agent team is not the only one on the connection. The Media Understanding team routes uploaded video through Gemini via OpenRouter to analyze content, and Google has announced that it is deprecating the version they use. Moving off it is a configuration change rather than a project.

Another team used the existing connection to prototype across four to six models for visual layout generation, including models the agent team does not run in production.

That now affects how Mistratov evaluates anyone selling him inference.

“People ping me all the time saying they have the best throughput on this model, you should use us. My first question is always: are you on OpenRouter? If the answer is no, I might just wait on that. It would need to be extremely advantageous in some way to overwhelm the ease of use I get by using OpenRouter.”

Aleks Mistratov, Head of AI Product, Descript

Descript on OpenRouter

Descript has been active on OpenRouter since June 2025. At the time of writing, Descript ranks #1 in our Video Generation category and #2 in Creative. Current usage is on the Descript app page.