Arena brings Max routing to Agent Mode
Arena released agentic Max on October 12, 2026. Select Max in Agent Mode for limited time routing between Claude Fable 5.1 and OpenAI Astra.
The company introduced its original text router in February and expanded it to search, vision and creative tasks in May. The new release targets tasks spanning several turns.
Read the comparison
During Arena’s test, Max recorded a 14.7 percent net improvement score at a median cost of $3.27 per task. Arena reports 37.5 percent lower cost than Fable 5.1 Max. Its training data predates September 17, before Opus 5.5. This historical comparison cannot establish leadership against today’s alternatives.
Read the score before the ranking
Arena’s signal definitions separate explicit confirmation of success from praise, complaints, steering effort and tool errors. For confirmed success, a user must answer the specific completion prompt. A silent ending contributes no confirmation. A casual positive comment is also different from that explicit response.
Steerability considers how much correction a task needs, including repeated corrections. Tool hallucinations concern calls to capabilities that are absent from the agent’s available tools. These measures describe different parts of an interaction. None makes the headline net improvement figure a simple percentage of tasks completed.
For a buyer, that means deciding which outcome matters before choosing a ranking. A document assistant can produce a useful draft while still needing substantial correction. A coding agent can receive positive feedback before its changes have been independently tested. Those are different review problems and deserve separate checks.
The command error measure has changed
A separate October 12 methodology update replaced Bash Recovery with Bash Reliability. Arena now uses an LLM judge to assess whether a command failed because of the agent’s mistake. The earlier rules could flag valid tests or empty searches and miss errors hidden inside piped commands.
Scoring now considers failures per command rather than how many turns followed an error before a successful call. Arena says the change shifted model results. It also describes a public dataset backfill as forthcoming. Comparisons across that change need the scoring version recorded alongside the date.
Understand what a median task costs
Arena’s cost methodology, published in August, divides sessions into individual tasks and counts at most the first three tasks when measuring cost. This limits the effect of accumulated context in long conversations. Its cost views include the median and other points in the distribution.
The same system separates Code, Chat and Work, with tasks allowed to belong to more than one category. These are observations from activity inside Arena. A median task cost should therefore be read with the task mix and evaluation setting, rather than used as a quote for running a particular business workflow.
Test a bounded piece of work
API access remains an interest signup. Permanent terms remain unspecified.
A useful trial would start with a task whose result is easy to inspect. For a research brief, agree on the questions and check whether the cited sources support the answers. For a code change, define the expected behavior and inspect the tests and diff before accepting it. Record review time as well as the machine’s output.
Repeat that task with the alternative you already use under comparable conditions. Treat the public comparison as a reason to investigate routing, then let the quality of the finished work and the effort needed to verify it determine whether the approach fits.
Illustrative desk photograph by Mashfik Upal Aumy. CC0 via the WordPress Photo Directory.
