Recommendation for Agents
AI Agents
Our top recommendation for AI Agents, based on the public evidence we track, is OpenAI: GPT-5.6 Sol.[1][2][3] Use as the primary implementation agent when your workflow separates planning from execution, with Claude Opus 5 handling review and soundness checks. Watch out: Watch for context compaction failures where pending actions and shell sessions are recorded but execution does not reliably resume. Anthropic: Claude Fable 5 is the next-ranked alternative. Assign to review and advisory roles where soundness checking matters more than raw implementation speed.
About this recommendation
- Updated
- Sep 4, 2026
- Evidence through
- Sep 4, 2026
- Sources
- 23
- Revision
- v65
Decision audit
Why this result
Inspect the inputs and the computed order behind the recommendation.
Models screened
21
live candidates
Evaluation feeds
7
task-weighted
Winner coverage
76%
intended feed weight
Largest provider share
4 of 9
Anthropic
Sources evaluated
The task sets these weights before any model is scored.
| Evaluation feed | Weight | Winner result | Field measured |
|---|---|---|---|
| Berkeley Function Calling | 30% | not measured | 3/21 |
| Terminal-Bench 2.1 | 18% | #13 | 12/21 |
| SWE-rebench | 15% | #4 | 16/21 |
| LMArena Agent | 12% | #3 | 15/21 |
| LiveBench Agentic Coding | 10% | #17 | 18/21 |
| OSWorld Verified | 10% | #2 | 5/21 |
| OpenRouter usage | 5% | 97/100 | 21/21 |
Provider concentration
Each exact model is scored separately; provider identity is not a ranking input.
- Anthropic4 models
- OpenAI2 models
- Z.ai2 models
- deepseek1 model
Decision table
Every published model is shown in computed order. Practitioner sources are distinct community threads, not the citations repeated in the prose below.
| Rank | Model | Relative score | Coverage | Practitioner evidence | Strongest measured reason |
|---|---|---|---|---|---|
| 01 | GPT-5.6 SolOpenAI | 75 | 76% | 20 threads · 12 families · 5 cautions | #2 OSWorld Verified · #3 LMArena Agent |
| 02 | Claude Fable 5Anthropic | 73 | 74% | 9 threads · 4 families · 1 cautions | #1 SWE-rebench · #2 LMArena Agent |
| 03 | GLM 5.2Z.ai | 70 | 67% | 35 threads · 12 families · 18 cautions | #8 LMArena Agent · #9 SWE-rebench |
| 04 | Claude Sonnet 4.5Anthropic | 62 | 65% | 1 threads · 1 families · 0 cautions | #2 Berkeley Function Calling · #20 SWE-rebench |
| 05 | GPT-5.6 LunaOpenAI | 60 | 74% | 11 threads · 9 families · 5 cautions | #4 Terminal-Bench 2.1 · #16 LMArena Agent |
| 06 | Claude Opus 4.6Anthropic | 59 | 67% | 4 threads · 3 families · 1 cautions | #13 LMArena Agent · #16 SWE-rebench |
| 07 | Claude Sonnet 4.6Anthropic | 56 | 68% | 4 threads · 3 families · 2 cautions | #5 OSWorld Verified · #13 SWE-rebench |
| 08 | GLM 4.6Z.ai | 52 | 65% | no linked practitioner threads | #3 Berkeley Function Calling · #44 SWE-rebench |
| 09 | DeepSeek V4 Flash Vision Expdeepseek | 52 | 22% | 6 threads · 6 families · 2 cautions | #3 LiveBench Agentic Coding |
Relative score combines normalized benchmark quality and signal coverage; independent practitioner evidence and freshness are bounded tie-breakers. It is an ordering score, not an absolute quality percentage. The writing model receives this order and cannot change it.
GPT-5.6 Sol ranks #3 on LMArena's agentic arena for tool-use and multi-step tasks, and is explicitly adopted as the lead implementation model in production agent protocols like SPARQ, though it shows weaker episodic learning gains compared to some alternatives.
Best when: Use as the primary implementation agent when your workflow separates planning from execution, with Claude Opus 5 handling review and soundness checks.
Tips
- Use as the primary implementation agent when your workflow separates planning from execution, with Claude Opus 5 handling review and soundness checks.
- Deploy for long-running, multi-step server operations with local shell/MCP tools where function calling reliability matters.
Watch out for
- Watch for context compaction failures where pending actions and shell sessions are recorded but execution does not reliably resume.
- Expect weaker learning across episodic attempts compared to DeepSeek V4 Flash or GLM-5.2 in self-improvement protocols.
Claude Fable 5 ranks #2 on LMArena's agentic arena and is explicitly designated for plan-review, code-review, advise, and security roles in multi-model orchestration, though it trails GPT-5.6 Sol on Agents' Last Exam by 13+ points.
Best when: Assign to review and advisory roles where soundness checking matters more than raw implementation speed.
Tips
- Assign to review and advisory roles where soundness checking matters more than raw implementation speed.
- Use as an orchestrator in multi-model pipelines with GPT-5.6-Sol and Claude Opus 4.8 as prover/builder agents.
Watch out for
- Budget for substantially higher cost than GPT-5.6 Sol at comparable or lower performance on long-horizon professional workflows.
- Handle temporary unavailability errors that can interrupt agent loops mid-execution.
GLM-5.2 ranks #9 on LMArena's agentic arena and shows strong episodic learning gains.
Best when: Use for GUI automation agent loops with local tool integration where lightweight architecture matters.
Tips
- Use for GUI automation agent loops with local tool integration where lightweight architecture matters.
Watch out for
- Guard against silent tool-call loss when SSE chunks fail JSON parsing, causing the agent to stop mid-action with half-finished output.
- Watch for thinking/action mismatches where the model plans the correct command in thinking blocks but emits a broken tool call.
Claude Sonnet 4.5 scores 73.24% on BFCL (#2 of 32) with strong tool-calling performance across single-turn, multi-turn, web-search, memory, and hallucination tests, though it lacks specific agentic arena ranking data.
Best when: Use for format-sensitive tool calling where BFCL-measured reliability across diverse tool-use patterns matters.
Tips
- Use for format-sensitive tool calling where BFCL-measured reliability across diverse tool-use patterns matters.
GPT-5.6 Luna ranks #18 on LMArena's agentic arena with a low score of 2.1, and is explicitly deprecated in favor of Luna High in some production configurations, with observed issues in ordinal turn-tracking and child agent spawning.
Best when: Consider only if explicitly migrating to GPT-5.6 Luna High with fixed reasoning effort and verbosity settings.
Tips
- Consider only if explicitly migrating to GPT-5.6 Luna High with fixed reasoning effort and verbosity settings.
Watch out for
- Avoid for multi-turn coordination tasks requiring reliable ordinal turn-tracking, which breaks compared to prior defaults.
- Do not use as a child agent in V2 multi-agent runtimes due to runtime/catalog version mismatches that cause spawning failures.
Claude Opus 4.6 ranks #14 on LMArena's agentic arena and is available through heavily subsidized channels like GitHub Copilot, though users report burning through monthly quotas rapidly with high reasoning settings.
Best when: Access through subsidized subscriptions for experimental agent harnesses where API costs would otherwise be prohibitive.
Tips
- Access through subsidized subscriptions for experimental agent harnesses where API costs would otherwise be prohibitive.
Watch out for
- Budget carefully for always-on agent deployments where high reasoning settings can exhaust monthly quotas in a single day.
Claude Sonnet 4.6 ranks #22 on LMArena's agentic arena with a score of 1, achieves only 9.3% binary accuracy on OSWorld Verified, and costs $26.78 per win in game benchmarks compared to $0.97 for Grok 4.1 Fast.
Best when: Use for directed agent flows with explicit test validation where the model works as far as given direction.
Tips
- Use for directed agent flows with explicit test validation where the model works as far as given direction.
Watch out for
- Expect high cost per successful task completion compared to newer alternatives in competitive benchmarks.
GLM 4.6 scores 72.38% on BFCL (#3 of 32) with competitive tool-calling latency, though it lacks specific agentic arena ranking or long-horizon task evidence.
Best when: Use as an open-weight alternative for format-sensitive tool calling where BFCL-measured performance is acceptable.
Tips
- Use as an open-weight alternative for format-sensitive tool calling where BFCL-measured performance is acceptable.
DeepSeek V4 Flash Vision Exp suffers from intermittent SSE stream freezes mid-generation without closing connections, vision modality failures where images are silently dropped or downgraded to path text, and Chat-Completions-only auxiliary restrictions.
Best when: Consider only after reviewing the cited caution.
Watch out for
- Avoid for vision-dependent agent tasks where images may be silently dropped, downgraded to filenames, or never encoded into user messages.
- Handle indefinite stream hangs where reasoning_content stops mid-generation with no terminal chunk or connection close.
Frequently asked
- What is the top-ranked model for AI Agents?
- OpenAI: GPT-5.6 Sol ranks first in the current evidence-weighted comparison. Use as the primary implementation agent when your workflow separates planning from execution, with Claude Opus 5 handling review and soundness checks.[1]
- What should I watch out for with OpenAI: GPT-5.6 Sol?
- Watch for context compaction failures where pending actions and shell sessions are recorded but execution does not reliably resume.[2]
- What is an alternative to OpenAI: GPT-5.6 Sol?
- Anthropic: Claude Fable 5 is the next-ranked option. Assign to review and advisory roles where soundness checking matters more than raw implementation speed.[3]
Sources
- 1
“> 🤖 SPARQ agent\n\nAdopt the maintainer's new default model protocol in Sparq's target routing catalog:\n\n- GPT-5.6 Sol leads implementation work.\n- Claude Opus 5 owns review and soundness.\n- Preserve the restricted security/soundness posture and explicit exhaustion behavior.\n- Keep target PLAN routing and registry CLAIM routing in exact agreement.\n\nUpdate the routing table and its regression tests; do not modify protected agent briefs.”
jeswr · GitHub · Sep 2, 2026 - 2
“### 📦 Environment - LobeHub Desktop: 2.2.13 (stable) - Deployment: self-hosted `lobehub/lobehub:latest` - OS: macOS 26.5.1, arm64 - Model: GPT-5.6-Sol with function calling - Task type: long-running, multi-step server operation using local shell/MCP tools ### 🐛 Description When the conversation reaches automatic context compaction while an agent task is still unfinished, the compacted context correctly records pending actions and even an outstanding shell session, but execution does not relia…”
brlanweb · GitHub · Aug 6, 2026 - 3
“## Summary A five-model advisor review of the newly merged model-tier taxonomy (`c266410`, agent-harness#309) converged on an author/reviewer split, but rejected making the heavy tier the default implementation tier. Recommended upstream policy: - roadmap/plan/supervise: Claude Opus 5 (`heavy`) - execute/repair: Claude Sonnet 5 (`regular`), with a recorded escalation to Opus 5 after repeated verification failure - plan-review/code-review/advise/security: Claude Fable 5 (`ultra`) - worker/econom…”
ViperJuice · GitHub · Jul 26, 2026 - 4
“## Observation from main-1 The episodic protocol (3 attempts, context wiped between them, persistent self-written notebook) was expected to produce visible learning. Result is mixed — mean final height attempt 1 → 3: | model | a1 → a3 | Δ | |---|---|---| | deepseek-v4-flash | 2.08 → 5.80 | **+3.7** | | glm-5.2 | 2.98 → 5.91 | **+2.9** | | k3 | 1.64 → 4.25 | **+2.6** | | gpt-5.6-sol | 3.17 → 4.76 | +1.6 | | claude-opus-5 | 6.50 → 8.10 | +1.6 (only monotonic riser) | | gpt-5.5 | 5.59 → 6.90 | +1.…”
eanderson4 · GitHub · Aug 6, 2026 - 5
“### Submission URL https://github.com/Solarys431/lean-eval-platonic-classification/tree/0742f2ff516619b3afebfeebd0f1473fe0b5ef57 ### Model UNICO/NOUS: Claude (Fable 5 + Opus 4.8) + GPT-5.6-Sol ### How this solution was produced (optional) Produced by UNICO/NOUS, a multi-model pipeline: Claude Fable 5 as the orchestrator in the main loop, with GPT-5.6-Sol (via Codex CLI) and Claude Opus 4.8 as prover/builder agents. A Python numerical scout computed exact golden-ratio data (flags, exposing funct…”
Solarys431 · GitHub · Jul 21, 2026 - 6
“"On Agents’ Last Exam (opens in a new window), an evaluation of long-running professional workflows across 55 fields, GPT‑5.6 Sol sets a new high of 53.6, eclipsing Claude Fable 5 (adaptive reasoning) by 13.1 points. Even at medium reasoning, it beats Fable 5 by 11.4 points at roughly one-quarter the estimated cost. That efficiency extends to smaller models, which are essential to making intelligence more abundant and affordable: GPT‑5.6 Terra and GPT‑5.6 Luna outperform Fable 5 at around one-s…”
saberience · Hacker News · Jul 9, 2026 - 7
“I keep getting this error mid agent loop: "Error: claude-fable-5 is temporarily unavailable" Planning went well, started working on the code, reading the code - all went fine But when it started writing the code or executing the bash, sarted tetting lots of these errors”
throwaw12 · Hacker News · Jul 1, 2026 - 8
“## 项目地址 https://github.com/xcq20100224/Nolan ## 一句话介绍 市面上的 AI 助手只会给建议，Nolan 直接动手：对它说「打开网易云音乐，播放我喜欢列表里的第一首歌」，它会真的移动鼠标去完成。 ## 主要特点 - **真·动手**：语音指令直接驱动 GUI 自动化——打开软件、点击、播放音乐、写文件、定时叫醒，不是输出聊天建议 - **本地优先**：语音识别用本地 faster-whisper，隐私不出本机；语音合成 GLM-TTS 带离线兜底 - **架构轻**：GLM-5.2 Agent 循环 + 14 个本地工具；React 前端 + Python 标准库后端，无 Electron - **一键安装**：install.bat 自动探测 Python 环境，配一个免费智谱 API key 即可运行 ## 演示 真机演示 GIF（自动打开网易云放歌全过程）： https://raw.githubusercontent.com/xcq20100224/Nolan/main/docs/demo.gif README 里还附了一次 Rea…”
xcq20100224 · GitHub · Jul 31, 2026 - 9
“## Problem When a provider's SSE stream contains a chunk that fails JSON parsing **and that chunk carries tool-call deltas**, the tool calls are silently lost and the agent turn ends as if the model had chosen to stop — mid-action, with a half-finished sentence. Live incident (2026-08-21 02:30, PR #85 review): The regenerated glm-5.2 completion contained **text only, no tool call** → agent loop saw "no tool calls" → turn ended normally. The already-decided `gh pr review --approve` never execute…”
xiaoer-bot · GitHub · Aug 21, 2026 - 10
“## Environment - **Model:** glm-5.2 (GLM Coding Plan, Anthropic-compatible endpoint) - **Harness:** Claude Code CLI v2.1.215 (`ANTHROPIC_BASE_URL` pointed at Z.ai) - **Context:** agentic coding session, Japanese-language context, ~450 assistant turns at the time of failure - **Date observed:** 2026-07-21 (JST) ## Summary During error recovery, GLM-5.2 entered a loop where it **planned the correct Bash command verbatim inside `thinking`, then emitted a tool call with the leading `cd && ` segment…”
carrotRakko · GitHub · Jul 21, 2026 - 11
“Scores 73.24% overall on BFCL (#2 of 32), measuring single-turn, multi-turn, web-search, memory, hallucination, and format-sensitive tool calling; mean latency was 4.31s.”
Berkeley Function-Calling Leaderboard · Benchmark · Sep 4, 2026 - 12
“## 概要 本リポジトリで行うすべての OpenAI Responses API 呼び出しを、次の固定構成へ全面移行する。 `gpt-5-mini`、GPT-5.4 系、GPT-5.5 系、GPT-5.6 Sol、GPT-5.6 Terraを含む旧構成・代替モデル・フォールバックモデルは今後使用しない。 モデルの自動ルーティング、用途別モデル切り替え、失敗時の上位モデルへのエスカレーションも導入しない。通常計画、低確信度時の確認、障壁通知、失敗後の再計画を含め、LLMを呼ぶ全経路を GPT-5.6 Luna High に統一する。 --- ## 決定事項 - 採用モデルIDは `gpt-5.6-luna` とする。 - 推論強度は常に `high` とする。 - `reasoning.effort` を省略してAPI既定値へ委ねない。 - `gpt-5.6` エイリアスは Sol を指すため使用しない。 - `text.verbosity` は `low` を標準とする。構造化出力に必要なフィールドはJSON Schemaで担保する。 - `reasoning.mode: pro`、…”
stillshore-chirp · GitHub · Jul 31, 2026 - 13
“## Background While implementing #14 (renaming the Gemini-branded OpenRouter node), changing `OPENROUTER_MODEL`'s default from `google/gemini-3.5-flash-lite` to `openai/gpt-5.6-luna` broke 3 existing real end-to-end debate-coordinator tests. Investigation (running the real debate flow manually, see PR history for tickets #15/#16) found two distinct behavior differences from the prior default: 1. **Ordinal turn-tracking is unreliable.** Given a topic instructing "Step 1 (now)... Step 2 (on your…”
Weilun2022 · GitHub · Jul 28, 2026 - 14
“### What issue are you seeing? `features.multi_agent_v2 = true` promotes a fresh GPT-5.6 Luna session to the V2 multi-agent runtime, but a V2 Sol parent still rejects Luna as a child because `spawn_agent` filters against Luna's static catalog value (`multi_agent_version = "v1"`) rather than its effective runtime. This is a narrower runtime/catalog consistency issue related to #34700. The existing issue demonstrates the user-visible rejection; this report adds A/B session metadata showing that L…”
zhanglg12 · GitHub · Jul 31, 2026 - 15
“Right now - there's some heavily subsidized subscriptions that are more or less cheating. For instance, Github CoPilot at $39 month gives you claude opus 4.6. They're going to close that off, but right now it's like a freebie for those doing API agentic harnesses. That said, if you are doing always on agents and you spend $3k-$4k on a GB10 or, $5+ k on Apple Silicon as your sunk cost, you will probably come out ahead. I've got 5 agents running a purely experimental social experiment. AThey oper…”
ytjohn · Hacker News · May 31, 2026 - 16
“Yeah, how many agents can you people even run at once and how much does it cost you? In company we used the monthly token quota and nowadays it's basically unusable with claude opus 4.6 on high reasoning. You can basically burn through 100% usage through a single day. How does it even scale for you with N agents and which magical plans or models do you use, where tools like this are even viable?”
Tooster · Hacker News · May 22, 2026 - 17
“> all the flows where a model has the initiative are strictly biased towards unwarranted stops Can you elaborate on what you think causes such a bias? My experience is that Qwen3.6, Claude Sonnet 4.6 and Opus 4.6 4.7 will work as far as they can given direction and a way to test their work. My so-far limited experience with Opus 4.8 is that it does stop somewhat earlier for feedback, but in places where I am glad it is checking assumptions or where I agree with it identifying a change in scope…”
entrope · Hacker News · May 31, 2026 - 18
“L icon Grok 4.1 Fast won 13 of 30 games at $0.97 per win The next-best winner was A icon Claude Sonnet 4.6 with 5 wins, at $26.78 per win. That’s a 27x difference. The model that isn’t on most top-model lists beat the model that is, on the thing a routing customer actually cares about. The model with the most kills did not win H icon GPT 5.4 killed 38 agents across 30 games. More than anyone else. It came in second on the leaderboard with 2 wins. If grok-4.1-fast was the top-winning model, and…”
trb · Hacker News · Jun 17, 2026 - 19
“Scores 72.38% overall on BFCL (#3 of 32), measuring single-turn, multi-turn, web-search, memory, hallucination, and format-sensitive tool calling; mean latency was 4.34s.”
Berkeley Function-Calling Leaderboard · Benchmark · Sep 4, 2026 - 20
“在飞书里发一张图片给 bot,视觉模型**完全看不到它**。三个独立缺陷叠在一起,把所有路径都堵死了; 最后 agent 只能靠文件名编造图片内容(实测它凭文件名编出了一段"美军黑鹰直升机照片"的描述)。 环境:v0.17.2,`adapter: sdk`,模型 `deepseek-v4-flash-vision-exp`(provider 自带 catalog 里 声明了 `inputModalities: ["text","image"]`)。 ## 1. SDK 适配器把图片"降级"成一行路径文本,从不作为图像内容发给模型 `dist/plugin.js:1190` → `buildPromptText(prompt, options.images)`, 以及 CLI 路径 `dist/plugin.js:919-922`: 图片变成了 prompt 里的**一串本机路径**。模型收不到任何像素,只能指望 agent 自己回头去 `read_image` 那个路径 —— 而下面两条正好让它读不了。 对比同一份代码里的 ACP 路径(`dist/plugin.js:673-676…”
zhuguangjun2002 · GitHub · Aug 23, 2026 - 21
“### Version line v2 — Go rewrite (1.x), main-v2 (active development) ### Exact version v1.31.3 ### What happened? 本 issue 是 #9288 的后续，针对其修复（PR #9299，方向 1+4）未覆盖的 **B 类**症状：会话中**首次**贴图时，客户端并未把图片编码进 user 消息，导致模型输入里**没有图像**（与 A 类"通道可用但 agent 绕路 OCR/MCP"不同）。 在运行 `deepseek/deepseek-v4-flash-vision-exp` 的会话中，首次贴图产生的 user 消息没有 `images` 数组： - `…\sessions\ .ckpt\turn-0.json` → `"files":[]`、`"expiredFilePayload":true`（`time=2026-08-24T08:46:23+08:00`） - `…\sessions\ .jsonl` 中该 user 消息字段为 `role, content, r…”
233-bot · GitHub · Aug 24, 2026 - 22
“## Summary DeepSeek endpoints intermittently stop sending SSE chunks mid-generation — typically mid-`reasoning_content` (thinking) — without closing the connection, without a terminal `finish_reason` chunk, and without `[DONE]`. The stream just goes silent. Observed on both the official `api.deepseek.com` and the `opencode-go` proxy, across multiple DeepSeek models (`deepseek-v4-flash`, `deepseek-v4-flash-vision-exp`). Other providers on the same transport (e.g. kimi-code/k3) never exhibit this…”
iacore · GitHub · Aug 25, 2026 - 23
“## Summary The DeepSeek OpenAI-compatible `/chat/completions` SSE endpoint (both `api.deepseek.com` and the `opencode-go` proxy) **intermittently stops sending chunks mid-generation without closing the connection** — typically mid-`reasoning_content` (thinking) — with no terminal `finish_reason` chunk and no `[DONE]`. The stream simply goes silent and stays open indefinitely. Observed across multiple models: - `deepseek-v4-flash` - `deepseek-v4-flash-vision-exp` This does **not** happen with ot…”
iacore · GitHub · Aug 26, 2026
Rankings synthesized from community evidence and open benchmarks. See methodology. Not driven by vendor marketing.