Please rotate your device

To continue exploring this website, please rotate your device to a vertical orientation

How Unblocked shifted to multi-model intelligence for lower cost per task

Unblocked partners with Baseten for their inference infrastructure to power code review.

Industry
Developer tools
Workload
LLMs
Primary result
20xlower price per token

Company overview

Unblocked is the context layer for agentic software development. It connects a team’s code, docs, tickets, and Slack conversations into a single knowledge graph, so developers and AI agents can understand not just the code but the decisions behind it. Webflow, Drata, Cribl, and Fetch use Unblocked to get answers faster and get it right with less back and forth from engineers.

As AI coding tools generate more code, the development bottleneck is now reviewing it. Unblocked Code Review automatically reviews every pull request when it’s opened or updated, flagging logic errors, security risks, and unsafe patterns with inline comments. Because Code Review draws on Unblocked’s context engine, it reviews how the team actually works, not generic best practices. It can catch when a change technically works but goes against a past decision.

Impact highlights

  • 3.1x lower cost per review

  • 20x lower price per token 

  • 4x lower stall rate vs. other providers

  • 100+ TPS on Baseten

Inference for agentic code review

Challenge

To write a useful review comment, Unblocked’s agent needs more than the code change. It needs to know why the code looks the way it does, what the team decided last quarter, and what past incident the change might repeat. The agent gathers that context over many model calls per review, and each call carries a large prompt. 

Originally, Unblocked powered their code review agent with closed models given the task required a high degree of accuracy. Their agent needed to thoughtfully balance accuracy versus flagging false issues or nitpicks to increase developer productivity and trust. However, as customers scaled and pushed more commits, it became clear closed models were not pareto efficient for every task (the right mix of intelligence to cost). Unblocked paid a premium to have closed models review each commit, whether or not the review found anything worth fixing. Even small routine steps that typically require less reasoning were running on closed models. 

In order to get more efficient intelligence, the Unblocked team wanted to explore whether open models could handle these tasks at the same high bar for quality, with closed models working in parallel only for the steps that required it. 

“Our customers trust our review comments for their high level of accuracy. We wanted to be cognizant of ensuring any changes in models continued to meet that bar. We’ve found a mix of open and closed models enable us to get the same intelligence, but faster and cheaper.” - Richie Bresnan

Shifting to multi-model intelligence for lower cost per task

Testing open model quality

While public benchmarks showed that open models had the potential to meet Unblocked’s quality bar, they needed to ensure developers would trust an open model’s comments, so Unblocked tested its own team first. For a few months, both models reviewed each commit in Unblocked’s own repositories and posted their comments side by side on the pull request. Unblocked’s engineers reacted to each comment with an emoji, without knowing which model wrote it. The open model passed the blind review on quality. 

Once Unblocked rolled the open model out to customers, the results were mixed. On one large customer’s reviews, the open model posted 0.46 comments per review against the closed model’s 0.31. The open model had better recall, catching more of the real issues in the code. However, the closed model had better precision: a higher share of its comments were correct, and developers accepted them a little more often. 

A newer open model closed that gap. In Unblocked’s internal benchmarks, this new open model scored the highest defect recall of any configuration the team had tested and remained the most cost-effective option. Unblocked made it the default engine for code review. This open model now handles most calls, and Unblocked routes the hardest reasoning steps to the closed model. 

Scaling open models in production

To start, Unblocked ran open models across several inference providers for redundancy. However, these providers had frequent failures, so Unblocked introduced a router to prioritize reliability over price. If a provider failed, the router sent a request to the next one, and a shared circuit breaker pulled a failing provider out of rotation within seconds.

Baseten joined the pool through Model APIs in July 2026. After initially running Baseten alongside other providers on live traffic, Unblocked made it the primary provider, the first one the router calls for each request. 

“I’ve found Baseten to be the best inference provider when adopting open models. And they have much better charts.”

  • Peter Werry, Founding Engineer at Unblocked

Unlocking pareto efficiency

On live traffic, Baseten runs at about 115 tokens per second against about 97 on the alternative, and a typical review on the open model finishes in about a minute. Baseten has a quarter of the stall rate of the other providers, and on Unblocked’s benchmark, the open model on Baseten hasn’t stalled once. Its prompt prefix cache also outperformed the rest of the pool by 10%. Additionally, with 99.9% uptime, there is little to no reliance on the circuit breaker within Unblocked’s router. 

Unblocked cut costs by about 20x per token by adopting closed models on Baseten - that’s $0.25 per million tokens versus $5 (includes effect of input, output, and cached tokens). That means for every code review, Unblocked saves 3.1x. The day after Unblocked’s largest consumer switched to the open model, daily costs dropped 67.9%. 

“The number that matters to us and our customers is the cost of a finished review. On Baseten, that’s 3.1x lower than with our previous closed model.”

  • Richie Bresnan, Founding Engineer at Unblocked

What’s next

Baseten is now the default for AI code review for customers without regional restrictions. Region-locked customers stay on the closed model, and on-prem deployments run where they did before. 

Unblocked is moving its largest remaining workload onto Baseten as part of a zero-day retention rollout. The team is also refining how it classifies review steps, so only the steps that need the closed model go to it and every other call runs on Baseten. 

When the next open model ships, Unblocked will put it through the same tests: a blind run on its own commits, then internal benchmarks, then customer traffic. Models that pass go into production on Baseten.