roobriefIT news for developers
Search

By AI

AI Coding Models: Claude Opus 5.5 Leads; New Models Arrive

Claude Opus 5.5 leads overall intelligence, while Sonnet 5.5 and Muse Wonder Full are neck and neck on the coding-agent board. There is no official value ranking, so on cost per task GPT-6.1 Sol stands out in the high-score band.

This is a weekly brief for picking models for coding and agent workflows. Rankings come from Artificial Analysis (AA) boards, with numbers from a snapshot around Oct. 18, 2026.

The Intelligence Index is a composite score across math, science, coding, and reasoning. The Coding Agent Index covers repository work, terminal use, and technical Q&A; coding ranks in this piece use a BenchLM mirror captured Oct. 15. “Hot models” is an editorial roundup of last week’s launches and default changes, not an official ranking.

Intelligence Index top 3

AA Intelligence Index v4.3.2 ranks the same model once per setting. The table below keeps only each model’s best setting, and second place is a tie 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

AA’s table leaves the cost per task blank for Muse Wonder Full. Dropping Opus 5.5 to lower settings gives 56 at $3.46 (xhigh) and 54 at $1.82 (high). Opus 5.5 shipped Sept. 22 and Sonnet 5.5 on Sept. 28; both have a 1 million token context. A token is a small piece of text, roughly part of a word, that AI models read and write. API prices are $4 input and $20 output per million tokens for Opus, and $2 input and $10 output for Sonnet.

The coding board looks different. The Coding Agent Index scores a model plus agent harness plus effort setting. Claude Code with Sonnet 5.5 (max) is at 68.4%, Muse Code with Muse Wonder Full (max) at 68.0%, and Claude Code with Opus 5.5 (max) at 66.0%. Gemini 4 Argon scores 63.8% in Antigravity CLI, and GPT-6.1 Sol (xhigh) in Codex scores 62.9%. The week’s takeaway is that overall intelligence and coding-agent leadership are not the same.

Value top 3

AA does not publish a “value top 3.” The table below is an editorial reading of Intelligence Index score versus cost per task. Sol’s score changes with setting, so max represents the high-score band and medium the cheaper sibling in the same family.

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 MiMo-V2.6-Pro (Xiaomi) 46 $0.13

In the high-score band (roughly 48 and up), Sol is among the cheapest options at a given score. At xhigh it scores 51 for $0.39 per task. OpenAI describes Sol as about one-fifth of Astra’s standard input and output prices ($2 input, $10 output, $0.10 cached input per million tokens). On the coding-agent board, Codex with Sol (medium) hits 61.4% at about $0.70 per task and (xhigh) 62.9% at about $1.04—far below Sonnet 5.5 max ($14) or Opus 5.5 max ($13).

Other options in a similar band include Claude Haiku 5.5 (max, intelligence 43, $0.21 per task) and open-weight GLM-5.3-Flash (intelligence 42, $0.28). Open-weight means releasing model weights so others can download, run, or adapt them. DeepSeek’s Oct. 17 claim of 77% on DeepSWE and about $0.20 per task for Flash-1017 is from its own harness, so it is not mixed into the AA table above.

Hot models last week, and what shipped new

Four threads stood out for coding roughly Oct. 12–18:

  1. Gemini 4 Argon — generally available Oct. 13. API ID gemini-4-argon. Introductory price is $2 input and $10 output per million tokens; after Nov. 30 that becomes $4 and $20. It is available in Google Antigravity, AI Studio, and the paid Gemini API, and already sits high on the Coding Agent board.
  2. Muse Wonder / Muse Wonder Full — Oct. 12. Meta’s frontier models for coding and long-running agents. Wonder is $4 input and $20 output on the API and Muse Code. Full is only on Muse Code Power and Maximum subscriptions—not sold via API. On the Coding Agent board, Full sits just under Sonnet.
  3. GPT-6.1 Sol — the model shipped Sept. 29, but after GPT-5.5 left ChatGPT-signed Codex on Oct. 14, paid plans were steered to Sol. Not a new release this week; a hot default-path change.
  4. DeepSeek-V4.1-Flash-1017 — Oct. 17. Still called deepseek-flash on the API at the same price, with open weights. Tuned for coding agents, and Search & Fetch APIs opened the same day (server-tool use free until Oct. 31, 16:00 UTC).

New releases, including local and open-weight options, by date:

Date Model How developers reach it
Oct. 12 Muse Wonder, Muse Wonder Full Meta Model API / Muse Code subscription
Oct. 13 Gemini 4 Argon (GA) Gemini API, Antigravity, AI Studio
Oct. 15 Nemotron 3.5 Super VL, Bolt 74B-A9B 24GB Open weights, NVIDIA NIM, local (RTX/DGX Spark, etc.). Bolt E4B due Oct. 23
Oct. 17 DeepSeek-V4.1-Flash-1017 deepseek-flash API, Hugging Face / ModelScope

Nemotron is less about cloud frontier competition and more about running agents on your own infra or locally with open weights. Super VL is a 120B (12B active) multimodal model; Bolt is sized for 24GB so you can do web, coding, and debug work on a local box.

Easy to misread

Do not read the Intelligence Index rank numbers as a count of models. AA lists the same model once per setting, and the table here keeps only each model’s best. Coding Agent scores are not bare model names—they include harnesses such as Claude Code, Codex, Muse Code, and Antigravity plus effort settings—and the value table is an editorial cost reading, not an AA official top list. Argon’s $2/$10 rates are introductory, and DeepSeek’s benchmark numbers are its own, so treat them separately from the AA tables.

What to watch next

  • How AA’s Coding Agent Index updates for DeepSeek-1017, Argon, and Wonder combinations
  • How API bills shift around Argon’s introductory price end on Nov. 30
  • Whether Muse Wonder Full ever ships on the API or stays Muse Code–only
  • Nemotron Bolt E4B (due Oct. 23) and how local agent harnesses feel in practice

Sources

Source: artificialanalysis.ai

← Back