Baseten Previews Coding Model Router

Baseten introduced Baseten Router on November 4, 2026, as a public preview on Baseten Grid. The system selects an open or closed model for each coding task. In tests spanning DeepSWE, SWE Atlas QnA, and Terminal-Bench, it matched Opus 5.5’s pass rate to within 0.8 percentage points while cutting costs by 72%. Baseten says it was the only commercial router it evaluated to beat random routing by a statistically significant margin at the same cost.

The router currently chooses between Opus 5.5 and GLM 5.3 Flash. Baseten selected GLM 5.3 Flash because it is a strong coding model that costs approximately 80% less than Opus 5.5 across relevant workloads. Users can adjust a routing threshold to choose different points on the cost-performance Pareto frontier (trade-offs where neither cost nor performance can improve without worsening the other). The selection score estimates each model’s chance of success on a task and subtracts a penalty proportional to its predicted cost in U.S. dollars.

Baseten compared its system with commercial routers from Fireworks, Together AI, and Applied Compute. It also reported that its results sat above the range containing 90% of random routers and ahead of OpenRouter, FireRouter, Together, and Applied Compute; the company described this as near-Opus accuracy at about a quarter of the cost. Baseten said it is working with design partners to tailor the router to production coding workloads and plans to share those results later. Opendoor’s head of engineering, Chintan Parikh, said automatic routing sends employees to a suitable model without requiring them to choose, while reducing payment for work that does not need frontier prices.

For training, Baseten ran Opus 5.5 and GLM 5.3 Flash at high reasoning effort on public coding datasets covering bug fixes, feature implementation, codebase questions, and terminal tasks, recording verifier results and model costs. Its held-out tests covered long-horizon coding with DeepSWE 1.1, codebase questions with SWE Atlas QnA, and terminal work with Terminal-Bench 2.1. The benchmarks were also included in Artificial Analysis’s Coding Agent Index v1.4 and OpenRouter’s initial Router Benchmark release. Baseten corrected grader errors and broken task environments, added safeguards against cheating, and excluded training tasks whose repositories appeared in the test set.

To address variation between repeated model runs, Baseten measured signal-to-noise ratio (SNR), a comparison of task-to-task differences with run-to-run variation. With one run per model, pooled test SNR was 0.27; averaging four Opus runs and eight GLM runs per task raised it to 1.51. Reported test pass rates average those runs. To test whether routing beat chance, the company simulated 10,000 random routers for each share of tasks assigned to GLM. The comparison band covered the middle 90% of random routers, from the 5th to 95th percentile; a result above it was treated as statistically significant.

In early experiments, Baseten embedded task instructions and trained predictors for success and cost. Predicting cost alone captured most, but not all, initial gains over random routing; combining predicted cost and success advantage performed best in cross-validation. A linear probe predicted benchmark identity from embeddings with 89% balanced accuracy, revealing a potential shortcut that might not transfer to other benchmarks or production. Baseten then explored removing benchmark cues, adjusting similar-task clusters, and extracting structured task attributes such as scope, required domain knowledge, and ambiguity, using Base Labs’ elephant-1 and mammoth-1 decision models.

Separately, the team tested whether GLM’s intermediate activations (internal model signals between input and output) could help predict cost. A prediction head estimating response steps reached approximately 0.6 Spearman correlation on validation data; the signal strengthened with model depth before leveling off and declining in deeper layers. Baseten used quantization techniques and smaller Qwen models as proxies for GLM’s internal layers to reduce routing latency while preserving performance. Its final router combines structured task attributes and internal activations to predict success and expected cost. The supplied article text ends as the results section begins, so it provides no further final benchmark figures.

Baseten introduced Baseten Router on November 4, 2026, as a public preview on Baseten Grid. The system selects an open or closed model for each coding task. In tests spanning DeepSWE, SWE Atlas QnA, and Terminal-Bench, it matched Opus 5.5’s pass rate to within 0.8 percentage points while cutting costs by 72%. Baseten says it was the only commercial router it evaluated to beat random routing by a statistically significant margin at the same cost.

The router currently chooses between Opus 5.5 and GLM 5.3 Flash. Baseten selected GLM 5.3 Flash because it is a strong coding model that costs approximately 80% less than Opus 5.5 across relevant workloads. Users can adjust a routing threshold to choose different points on the cost-performance Pareto frontier (trade-offs where neither cost nor performance can improve without worsening the other). The selection score estimates each model’s chance of success on a task and subtracts a penalty proportional to its predicted cost in U.S. dollars.

Baseten compared its system with commercial routers from Fireworks, Together AI, and Applied Compute. It also reported that its results sat above the range containing 90% of random routers and ahead of OpenRouter, FireRouter, Together, and Applied Compute; the company described this as near-Opus accuracy at about a quarter of the cost. Baseten said it is working with design partners to tailor the router to production coding workloads and plans to share those results later. Opendoor’s head of engineering, Chintan Parikh, said automatic routing sends employees to a suitable model without requiring them to choose, while reducing payment for work that does not need frontier prices.

For training, Baseten ran Opus 5.5 and GLM 5.3 Flash at high reasoning effort on public coding datasets covering bug fixes, feature implementation, codebase questions, and terminal tasks, recording verifier results and model costs. Its held-out tests covered long-horizon coding with DeepSWE 1.1, codebase questions with SWE Atlas QnA, and terminal work with Terminal-Bench 2.1. The benchmarks were also included in Artificial Analysis’s Coding Agent Index v1.4 and OpenRouter’s initial Router Benchmark release. Baseten corrected grader errors and broken task environments, added safeguards against cheating, and excluded training tasks whose repositories appeared in the test set.

To address variation between repeated model runs, Baseten measured signal-to-noise ratio (SNR), a comparison of task-to-task differences with run-to-run variation. With one run per model, pooled test SNR was 0.27; averaging four Opus runs and eight GLM runs per task raised it to 1.51. Reported test pass rates average those runs. To test whether routing beat chance, the company simulated 10,000 random routers for each share of tasks assigned to GLM. The comparison band covered the middle 90% of random routers, from the 5th to 95th percentile; a result above it was treated as statistically significant.

In early experiments, Baseten embedded task instructions and trained predictors for success and cost. Predicting cost alone captured most, but not all, initial gains over random routing; combining predicted cost and success advantage performed best in cross-validation. A linear probe predicted benchmark identity from embeddings with 89% balanced accuracy, revealing a potential shortcut that might not transfer to other benchmarks or production. Baseten then explored removing benchmark cues, adjusting similar-task clusters, and extracting structured task attributes such as scope, required domain knowledge, and ambiguity, using Base Labs’ elephant-1 and mammoth-1 decision models.

Separately, the team tested whether GLM’s intermediate activations (internal model signals between input and output) could help predict cost. A prediction head estimating response steps reached approximately 0.6 Spearman correlation on validation data; the signal strengthened with model depth before leveling off and declining in deeper layers. Baseten used quantization techniques and smaller Qwen models as proxies for GLM’s internal layers to reduce routing latency while preserving performance. Its final router combines structured task attributes and internal activations to predict success and expected cost. The supplied article text ends as the results section begins, so it provides no further final benchmark figures.