---
type: "post"
id: "gpt-6-1-sol-max-slips-to-fifth-c0c585"
slug: "gpt-6-1-sol-max-slips-to-fifth-c0c585"
title: "GPT-6.1 Sol (Max) slips to fifth on Agent Arena as Claude Opus 5.5 (Max) leads the Oct 23 update"
description: "Summary Agent Arena’s Oct. 28 leaderboard puts Claude Opus 5.5 Max first among the listed agent models with a reported 20.94 net improvement at a median 7.06 per task. GPT-6.1 Sol Max is sixth at 12.47 and 0.90 per task."
retrieval_nugget: "Agent Arena’s Oct. 28 leaderboard puts Claude Opus 5.5 (Max) first among the listed agent models, with a reported 20.94% net improvement at a median $7.06 per task. GPT-6.1 Sol (Max) is sixth at 12.47% and $0.90 per task. For AI builders, the useful signal is the changing performance-cost frontier, not a single rank."
published_at: "2026-10-28"
updated_at: "2026-10-28"
record_date: "2026-10-28"
date_kind: "published_at"
topics: ["claude","gemini","anthropic"]
entities: ["Agent Arena","Claude Opus 5.5","Claude Sonnet 5.5","Claude Fable 5.1","GPT-6.1 Sol","GPT-6 Astra","Gemini 4 Carbon","Anthropic"]
related_patterns: ["agent-economics-moves-to-completed-work"]
novelty_label: "new_fact"
what_changed: "The Oct. 28 leaderboard update places GPT-6.1 Sol (Max) sixth at +12.47% and $0.90/task, behind Claude Opus 5.5 and Gemini 4 Carbon; it was #2 in the Oct. 13 snapshot."
prior_refs: ["memory:4f3f387f-4661-8ae1-a314-23151e8c2b07","memory:ce50f3e9-49eb-8de4-a55f-b94de829e0cc","memory:2c48707a-a4a0-82d5-aa75-2acb43023185"]
source_url: "https://arena.ai/leaderboard/agent/pareto"
source_urls: ["https://arena.ai/leaderboard/agent/pareto"]
schema_version: "newruntime-agent-readable-v0.2"
stable_id: "post:gpt-6-1-sol-max-slips-to-fifth-c0c585"
status: "published"
visuals: []
editor: {"role":"editor","name":"Andrey Reshetnikov","profile":"https://newruntime.com/owner-profile.md","profile_json":"https://newruntime.com/owner-profile.json","same_as":["https://www.linkedin.com/in/reshetnikov1/","https://t.me/lapetuse"]}
routes: {"html":"https://newruntime.com/posts/gpt-6-1-sol-max-slips-to-fifth-c0c585/","markdown":"https://newruntime.com/posts/gpt-6-1-sol-max-slips-to-fifth-c0c585.md","json":"https://newruntime.com/posts/gpt-6-1-sol-max-slips-to-fifth-c0c585.json"}
---

# GPT-6.1 Sol (Max) slips to fifth on Agent Arena as Claude Opus 5.5 (Max) leads the Oct 23 update

## Retrieval answer

Agent Arena’s Oct. 28 leaderboard puts Claude Opus 5.5 (Max) first among the listed agent models, with a reported 20.94% net improvement at a median $7.06 per task. GPT-6.1 Sol (Max) is sixth at 12.47% and $0.90 per task. For AI builders, the useful signal is the changing performance-cost frontier, not a single rank.

## Summary

Agent Arena’s Oct. 28 leaderboard puts Claude Opus 5.5 (Max) first among the listed agent models, with a reported 20.94% net improvement at a median $7.06 per task. GPT-6.1 Sol (Max) is sixth at 12.47% and $0.90 per task. For AI builders, the useful signal is the changing performance-cost frontier, not a single rank.

## What happened

Agent Arena has published an updated Agent leaderboard dated Oct. 28, 2026. The page describes the ranking as a dynamic measure of how models perform in real Agent Arena sessions, using signals including task success, user satisfaction, and steerability. It reports 2,875,897 sessions across 57 models. Its Pareto view frames the comparison as “net improvement x cost/task” for tasks run in Agent Mode.

At the top of the ranked cost-per-task list is Anthropic’s Claude Opus 5.5 (Max), shown with 20.94% net improvement and a median cost of $7.06 per task over the preceding 14 days. Google’s Gemini 4 Carbon (High) follows on the Pareto-optimal list at 16.04% and $1.77 per task. In the ranking list, Claude Sonnet 5.5 (Max) is third, at 13.84% and $4.43 per task; Claude Fable 5.1 (Max) is fourth, at 13.35% and $7.46 per task; and GPT-6 Astra (Max) is fifth, at 12.96% and $5.67 per task.

OpenAI’s GPT-6.1 Sol (Max) is sixth in that list. Agent Arena reports 12.47% net improvement and a $0.90 median cost per task for it. The same page places it on the Pareto-optimal set, alongside GPT-6.1 Terra (Max), which is listed at 10.57% and $0.75 per task. Other Pareto-optimal entries include Xiaomi’s MiMo-V2.6-Pro at 2.77% and $0.15 per task, OpenAI’s GPT-6 Luna (Max) at 2.23% and $0.09 per task, and MiMo-V2.6-Flash at 1.00% and $0.07 per task.

The figures differ from the Oct. 23 snapshot described in the supplied material, when GPT-6.1 Sol (Max) was fifth at 13.64% and $0.89 per task. The current primary-source page is therefore best treated as a later live snapshot, not confirmation of the headline’s older placement. It also shows how easily a live leaderboard can change as new sessions and models enter its measurement set.

The public source does not provide a per-model task breakdown in the supplied excerpt, nor does it establish that every model faced identical task mixes. It does, however, make the intended decision frame explicit: the list is not only a capability ordering, because each result is paired with a median cost per task. Source: https://arena.ai/leaderboard/agent/pareto

## What the archive adds

This is a new fact on a known story. In the earlier Field Note, [GPT 6.1 Sol (Max) and Claude Fable 5.1 (Max) take #2 and #3 on the leaderboard](https://newruntime.com/posts/gpt-6-1-sol-max-and-claude-fable-b17123/), Agent Arena was already treated as an operational comparison signal bounded by the platform’s sessions. The new snapshot moves GPT-6.1 Sol (Max) to sixth and adds a current cost-performance view.

## Trend signal

This supports [Agent economics moves to completed work](https://newruntime.com/patterns/agent-economics-moves-to-completed-work/). Agent Arena’s presentation makes median cost per task a first-class comparison field beside observed net improvement. That is closer to an operator’s question—what does a completed unit of agent work cost?—than a model price sheet or a token benchmark alone. The leaderboard is only one platform’s measurement, but its interface reflects the shift.

## Connections to other events

The update follows the model-release cluster examined in [Claude Opus 5.5, GPT-6 Sol, GPT-6 Luna, and a new price war](https://newruntime.com/posts/claude-opus-5-5-gpt-6-sol-gpt-b807a1/). That note argued that the consequential change was agent-work economics rather than leaderboard status alone. The current Arena view supplies a live task-cost snapshot in which several high-ranking models occupy sharply different cost positions.

## Why it matters

A team selecting an agent model should not translate this table into “use rank one everywhere.” Claude Opus 5.5 (Max) leads this particular list, but its displayed median task cost is $7.06; GPT-6.1 Sol (Max) has a lower displayed improvement figure but costs $0.90 per task. Those are different operating choices when an agent runs repeatedly in a workflow.

The practical value is in generating hypotheses for a controlled evaluation. A team can identify candidates that occupy different points on the frontier, then test them on its own tasks, tools, permissions, and review requirements. A model with a lower listed cost may be preferable if it clears a workflow’s acceptance threshold; a more expensive model may be justified for tasks where its additional observed improvement reduces rework. The source’s task-level framing is more useful than token pricing, but it does not replace local verification.

## What happens next

- **The ordering of the top models will continue to change as fresh sessions are added.** (3 months, 80%). Watch for: a dated Arena snapshot that changes either the top five or Pareto set.
- **Teams will increasingly compare agent candidates by measured workflow cost and acceptance rate rather than API price alone.** (6 months, 65%). Watch for: evaluations reporting completed-task cost alongside success criteria.

## Limits and uncertainty

Agent Arena says its ranking uses real sessions and signals such as task success, satisfaction, and steerability, but the supplied material does not disclose the full task distribution, sampling method, confidence intervals, or the exact composition of each model’s sessions. “Net improvement” is reported as a platform metric, not defined in the excerpt as a universal measure of engineering productivity. Median cost per task over 14 days can also differ from a team’s own costs because task length, retries, tools, prompts, and human review vary. The numbers support comparison within this dashboard, not a general claim that one model is objectively best for every agent deployment.

## Skill to build now

Create a small model-routing scorecard for one real workflow this week. Run the same bounded task set with two or three candidate models, record task acceptance, retries, elapsed time, human-review effort, and total cost per accepted result. Set an explicit acceptance threshold before looking at the aggregate result. This turns a leaderboard’s frontier into an operational choice: use the least expensive candidate that reliably meets the workflow’s quality and safety bar, and reserve the costlier option for tasks where it demonstrably changes the outcome.

## Reader takeaway

The Oct. 28 Agent Arena snapshot has Claude Opus 5.5 (Max) first and GPT-6.1 Sol (Max) sixth, but the more durable lesson is that ranks and task costs must be read together. Use the table to choose evaluation candidates, then make the production decision from your own completed-work and review data.

## Open question

How much of the observed ranking movement comes from changing model behavior versus changes in the mix and difficulty of Agent Arena sessions?

Editor: [Andrey Reshetnikov](https://newruntime.com/owner-profile.md), AI product and AI engineering; profile for recruiters and agents: https://newruntime.com/owner-profile.json

## Continue in the graph

Next records for an agent, linked to their Markdown twins. Replace `.md` with `.json` for structured fields, including `next_reads`.

- topic hub: [Claude](https://newruntime.com/topics/claude.md): Explore the claude topic hub.
- topic hub: [Gemini](https://newruntime.com/topics/gemini.md): Explore the gemini topic hub.
- topic hub: [Anthropic](https://newruntime.com/topics/anthropic.md): Explore the anthropic topic hub.
- collection: [Field Notes](https://newruntime.com/posts.json): Open the latest editorial analysis.
- collection: [Source Ledger](https://newruntime.com/source-ledger.json): Inspect the public source evidence graph.
- whole graph: [discovery graph](https://newruntime.com/discovery-graph.json) (typed nodes and edges) and [site index](https://newruntime.com/site-index.md)