Baseten Router Nears Claude Opus 5.5 at 72% Lower Cost

Baseten has released its first model router for coding tasks. In its own benchmark, the router came within 0.8 percentage points of always using Anthropic’s Claude Opus 5.5 at 72% lower cost: $0.38 a task against $1.34, and a 69.6% pass rate against 70.4%, according to Baseten’s blog post of 4 November.

Per passed task, that is about $0.55 against about $1.90, or about 3.5 times as many passed tasks per dollar, by our arithmetic from Baseten’s figures.

Business Pill · Percentiles

A short Business Engineer explainer on percentiles, the measure behind the 5th-to-95th percentile band Baseten uses for random routers. It teaches the general idea only and says nothing about Baseten or its router.

What Baseten Announced

The post is titled “Introducing Baseten Router: a Pareto-optimal model router for coding tasks”. Baseten says the router “sends each coding task to the right model, open or closed, so you pay frontier prices only when the task needs it”. It is part of Baseten Grid, which the company says is now in public preview.

For now the router picks between two models: Opus 5.5 and the open-weight GLM 5.3 Flash. Baseten says it chose GLM 5.3 Flash because it is “approximately 80% cheaper than Opus 5.5 across relevant workloads”.

Sarah Guo, quoting Baseten’s post on X at 00:25 UTC on 5 November, called it “best in class (cheapest at every pass rate)”.

What Was Posted on X

Sarah Guo posted this at 00:25 UTC on 5 November, quoting Baseten’s X article on the router. She calls it best in class, cheapest at every pass rate. The benchmark figures in this piece come from Baseten’s own blog post and charts.

The Benchmark Numbers

Baseten tested the router on three benchmarks held out from training: DeepSWE 1.1, SWE Atlas QnA and Terminal-Bench 2.1. Its results chart puts always using Opus 5.5 at a 70.4% pass rate and $1.34 a task.

The router has three settings, which Baseten calls Cost, Balanced and Intelligence. Its chart shows three points: 0.8 points below Opus at 72% cheaper, 2.6 points below at 80% cheaper, and 3.3 points below at 84% cheaper. Always using GLM 5.3 Flash came in 5.9 points below Opus at 87% cheaper.

Baseten says it was the only router it evaluated that did better than random routing at the same cost, on these benchmarks, by a statistically significant margin. The routers it compared include those from OpenRouter, Fireworks, Together AI and Applied Compute.

Cost per Passed Task

Dividing cost per task by pass rate gives the cost of each task that passes. Always using Opus 5.5 works out at about $1.90 per passed task, by our arithmetic. The router’s top setting works out at about $0.55.

At the cheaper settings, from Baseten’s rounded percentages, the figure falls to about $0.40 and about $0.32. Always using GLM 5.3 Flash comes to about $0.27 per passed task, also our arithmetic. These figures are approximate because Baseten rounds its percentages.

How the Router Decides

Baseten says the router reads the input prompt and predicts each model’s chance of success and expected cost. Its final version combines “structured task attributes and internal model activations”.

The post says GLM 5.3 Flash’s intermediate activations predicted the number of agent steps with a Spearman correlation of about 0.6. On a single B200 GPU at 4 concurrent requests, Baseten measured p50 and p95 latencies of 273 ms and 429 ms.

Baseten’s Own Caveats

The post says “Relative router performance may differ on other benchmarks.” It also says model pools differ: “All but one router offered both Opus 5.5 and GLM 5.3 Flash. One router was limited to Anthropic models.”

Baseten says its next focus is production, where requests are less explicit, success is harder to measure and tasks run over many turns. It quotes Chintan Parikh, head of engineering at Opendoor, as saying that auto routing means “we stop paying frontier prices for work that doesn’t need it”.

Connecting the Coverage

Our related pieces on Baseten and the cost of AI coding:

• Baseten Set for a $30 Billion Valuation, 2.3x Its June Mark: Baseten’s funding, a week earlier.
• Cheapest OpenAI, Anthropic Main Models Cost 1% of Top Price: the model price ladder a router works across.
• Claude Code Subagent Status Updates Now Cost About 99% Less: another way coding-agent costs are being cut.
• Nvidia: 256K Context Costs Up to 7x; Opus Hit 70.2% at 16K: context length and the cost of agent runs.
• Claude Cut ChatGPT’s Lead in £1B UK AI Spend to 23 Points: where business AI spend is going.
• AI Token Prices 2026: what every major model charges per million tokens.

The Structural Read

As we read the post, the claim is about choosing a model per task, not about a new model: the router picks between Opus 5.5 and GLM 5.3 Flash for each task.

As we read Baseten’s chart, the cheapest cost per passed task in it is always using GLM 5.3 Flash, at about $0.27 by our arithmetic; what the router adds is a pass rate closer to Opus.

As we read the caveats, the comparison is as offered by each router: Baseten says model pools differ and that one router was limited to Anthropic models.

Three Implications

FOR TEAMS BUYING CODING AGENTS Baseten says the three settings can be passed in at request time to the same router, without retraining.

FOR MODEL MAKERS Baseten says it chose GLM 5.3 Flash because it is approximately 80% cheaper than Opus 5.5 across relevant workloads.

FOR BASETEN The post says it is working with design partners to specialise the router for production coding workloads, and gives no production results yet.

The Business Engineer Lens

This story maps onto the Business Engineer framework The CFO’s Guide to the Token Economy.

The framework says: “The new job of finance is to manage the ratio between the two — work produced per token burned.”

As we read it, cost per passed task is one way to write that ratio for coding agents, and Baseten’s benchmark puts it at about $0.55 for the router against about $1.90 for Opus 5.5, our arithmetic. The framework does not discuss Baseten or this router.

What Is Not Established

All results are Baseten’s own benchmark, run on three public coding benchmarks; we did not test the router. The post gives no production results and no price for the router itself. Baseten’s chart prints exact values only for always using Opus 5.5 and for the router’s top setting; the other points are given as rounded percentages. The $1.90, $0.55 and 3.5 times figures are our arithmetic.

The Bottom Line

Baseten’s first router came within 0.8 points of Claude Opus 5.5’s pass rate at 72% lower cost in its own coding benchmark, $0.38 a task against $1.34. Per passed task, that is about $0.55 against $1.90, or about 3.5 times as many passed tasks per dollar, by our arithmetic.