# AI Coding Models: Opus 5.5 Leads Agent v2.0; Kimi K3.1 Pro Debuts

> As of Oct. 25, Coding Agent Index v2.0 puts Claude Opus 5.5 back on top. New Kimi K3.1 Pro debuts at fourth on the Intelligence Index.

- URL: https://roobrief.com/en/posts/ai-weekly-coding-models-oct-25
- Language: en
- Published: 2026-10-24T15:20:00.000Z
- Tags: ai
- Source: https://artificialanalysis.ai/leaderboards/models

**Claude Opus 5.5 retained the top spot in overall intelligence, and with the release of Coding Agent Index v2.0, Opus 5.5 overtook Sonnet 5.5 to reclaim the coding agent lead. For value, GPT-6.1 Sol remains strong, while newly launched Kimi K3.1 Pro joined the high-score alternatives.**

This is a weekly brief for choosing models across coding and agentic workflows. Rankings are based on Artificial Analysis (AA) leaderboards, with figures taken from a snapshot around Oct. 25, 2026.

The Intelligence Index evaluates overall capability across math, science, coding, and reasoning. The Coding Agent Index measures performance on realistic open-source repository tasks, interactive terminal operations, and technical software Q&A. This week's headline is the Oct. 23 launch of Coding Agent Index v2.0, which reordered the frontier standings. "Hot models" is an editorial roundup of new launches and default changes over the past week rather than an official rank.

## Intelligence Index top 3 and new entrant

AA Intelligence Index v4.3.2 evaluates models per configuration setting. The table below lists each model's best setting; second place remains tied at 56:

| Rank | Model (setting) | Score | Cost per task |
|---|---|---|---|
| 1 | Claude Opus 5.5 (max, fallback) | 58 | $5.98 |
| 2 | Claude Sonnet 5.5 (max, fallback) | 56 | $5.46 |
| 2 | Muse Wonder Full (max) | 56 | not listed |

The top three standings stayed level, but Moonshot AI's Kimi K3.1 Pro—shipped on Oct. 23—debuted at fourth with an Intelligence Index score of 55 and an estimated cost of $2.11 per task. Opus 5.5 launched Sept. 22 and Sonnet 5.5 on Sept. 28; both feature a 1-million-token context window. A token is a small piece of text, roughly part of a word, that AI models read and write. API pricing for Opus is $4 input and $20 output per million tokens, and Sonnet is $2 input and $10 output. Kimi K3.1 Pro matches Sonnet on headline token rates ($2.00 input, $10.00 output), but higher average output generation pushes its benchmark task cost closer to $2.

## Coding Agent Index v2.0 overhaul

The most consequential update this week was the Oct. 23 release of Coding Agent Index v2.0. Replacing the earlier DeepSWE v1.1 benchmark, the index now integrates Cognition's FrontierCode v1.1 (100 repository tasks), Terminal-Bench 4.0 (66 tasks), and SWE-Atlas-QnA-AA (124 tasks) across 290 total evaluation workloads. By requiring code modifications that human maintainers would actually accept and merge, the average score across 21 tracked configurations dropped by roughly 7.4 points (from 54.8 down to 47.4).

| Rank | Model and harness | v2.0 score | Notes |
|---|---|---|---|
| 1 | Claude Code + Claude Opus 5.5 (max) | 61.7 | ~$11.42 per task |
| 2 | Claude Code + Claude Sonnet 5.5 (max) | 59.3 | Replaced from last week's v1.0 lead |
| 3 | Muse Code + Muse Wonder Full (max) | 58.9 | Subscription only |
| 4 | Claude Fable 5.1 (max, fallback) | 57.6 | Consistent upper tier |

Under v1.0, Sonnet 5.5 on Claude Code (68.4%) and Muse Wonder Full on Muse Code (68.0%) outscored Opus 5.5 (66.0%). Under the stricter maintainer-merge standards of v2.0, Opus 5.5 regained first place at 61.7 points. Its deeper reasoning capacity appeared to hold up better across broader repository context and refactoring tasks. Further down the board, GPT-6.1 Sol on Codex scored 54.1 points (~$1.50 per task), Gemini 4 Argon on Antigravity CLI landed in the low-to-mid 40s, and DeepSeek V4.1 Flash reached 47.3 points at approximately $0.39 per task.

## Value top 3

AA does not maintain a dedicated "value top 3." The table below reflects an editorial comparison of Intelligence Index performance against per-task expenditure. Because Sol's scores vary by reasoning effort, max represents the upper tier and medium the lower-cost counterpart:

| Rank | Model (setting) | Intelligence Index | Cost per task |
|---|---|---|---|
| 1 | GPT-6.1 Sol (max) | 52 | $0.72 |
| 2 | GPT-6.1 Sol (medium) | 48 | $0.21 |
| 3 | Kimi K3.1 Pro (Moonshot AI) | 55 | $2.11 |

Sol remains the most economical choice within the 48–52 intelligence range. At the xhigh setting, it delivers an intelligence score of 51 at $0.39 per task. The newly introduced Kimi K3.1 Pro nears frontier intelligence (55 vs. 56–58 for Sonnet and Opus) while cutting task expenditure to $2.11—less than half of Anthropic's top-tier runs. For lower budgets, MiMo-V2.6-Pro (intelligence 46 at $0.13) and Claude Haiku 5.5 (max, intelligence 43 at $0.21) maintain strong baselines.

On the Coding Agent board, Codex with Sol (54.1 points at ~$1.50) and DeepSeek V4.1 Flash (47.3 points at ~$0.39) continue to serve as viable workhorse alternatives to high-cost Opus runs exceeding $11 per rollout.

## Hot models last week, and what shipped new

Four notable developments unfolded across coding and agent infrastructure between Oct. 19 and Oct. 25:

1. **Moonshot AI Kimi K3.1 Pro & Highspeed** — released Oct. 23. Built on a 2.8-trillion parameter architecture (104B active MoE) with a 1-million-token context window. It entered the Intelligence Index directly at 55 and has planned open-weights release. Open-weight means releasing model weights so others can download, run, or adapt them.
2. **NVIDIA Nemotron 3.5 Bolt E4B 4GB** — launched Oct. 23. Based on Google's Gemma 4 E4B and tuned to fit 4GB VRAM(memory dedicated to a graphics card) targets (Q4_K_M GGUF). It supports 64k context windows, multi-turn tool calling, and speculative decoding drafters for edge devices and local desktops.
3. **Coding Agent Index v2.0** — announced Oct. 23. Replaced DeepSWE with FrontierCode v1.1 to align agent evaluations with maintainer-level pull request approvals.
4. **OpenAI Decisions API General Availability** — GA on Oct. 23. Gives builders a specialized, low-latency endpoint for classification, routing, and fast discrete agent decisions.

New open and local weights released across the week:

| Date | Model / Product | Developer access path |
|---|---|---|
| 10/21 | Nemotron 3.5 Super VL, Bolt 74B | NVIDIA NIM, Hugging Face open weights |
| 10/23 | Kimi K3.1 Pro, Pro Highspeed | Moonshot API, Kimi Code (open weights pending) |
| 10/23 | Nemotron 3.5 Bolt E4B 4GB | Hugging Face, local runtimes |
| 10/23 | Coding Agent Index v2.0 | Artificial Analysis leaderboard |

## Common misconceptions

Lower raw numerical scores on Coding Agent Index v2.0 compared to v1.0 (dropping from late 60s into the 50s and low 60s) reflect benchmark toughening via FrontierCode v1.1, not model regressions. For Kimi K3.1 Pro, headline token pricing matches Sonnet ($2/$10), but verbose generation patterns yield a $2.11 task cost that developers must factor into operational budgets. Finally, Bolt E4B 4GB is engineered for local edge execution on memory-constrained hardware, not to challenge multi-hundred-billion cloud frontiers.

## What to watch next

- How Kimi K3.1 Pro performs when evaluated directly within standard Coding Agent Index v2.0 agent harnesses
- Release timelines for Kimi K3.1 Pro's open weights and subsequent local deployment recipes
- Model utilization shifts ahead of Nov. 30, when Gemini 4 Argon's introductory API pricing ($2/$10) expires
- Real-world agent tool workflows building on 4GB-footprint Nemotron Bolt models on developer machines

## Sources

- [Artificial Analysis Intelligence Index](https://artificialanalysis.ai/evaluations/artificial-analysis-intelligence-index) (Artificial Analysis)
- [Coding Agent Benchmarks](https://artificialanalysis.ai/agents/coding-agents) (Artificial Analysis)
- [LLM Leaderboard](https://artificialanalysis.ai/leaderboards/models) (Artificial Analysis)
- [Moonshot AI Kimi K3.1 Pro Announcement](https://kimi.ai/blog/k3-1-pro) (Moonshot AI)
- [Nemotron 3.5 Bolt E4B Announcement](https://blogs.nvidia.com/blog/nemotron-3-5-bolt-e4b/) (NVIDIA)
- [OpenAI Decisions API GA](https://openai.com/index/decisions-api-ga/) (OpenAI)
- [Introducing Claude Opus 5.5](https://www.anthropic.com/claude-opus-5-5) (Anthropic)
- [Introducing Claude Sonnet 5.5](https://www.anthropic.com/claude-sonnet-5-5) (Anthropic)
- [Introducing GPT-6.1 Sol](https://openai.com/index/introducing-gpt-6-1-sol/) (OpenAI)