Recommendation for Autonomous workflows
Autonomous Agents
Our top recommendation for Autonomous Agents, based on the public evidence we track, is OpenAI: GPT-5.6 Sol.[1][2][3][4][5] Use as the lead implementation model in SPARQ agent protocols where it owns autonomous execution, repair loops, and subagent spawning for investigation and testing tasks. Watch out: Watch for mid-turn steer hijacking where a short interrupt can cause the agent to complete only the latest mini-task and omit the main task from the final response. Anthropic: Claude Fable 5 is the next-ranked alternative. Assign to review and soundness lanes where SPARQ protocols explicitly designate Fable 5 for plan-review, code-review, advise, and security tasks.
About this recommendation
- Updated
- Sep 4, 2026
- Evidence through
- Sep 4, 2026
- Sources
- 26
- Revision
- v59
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
100%
intended feed weight
Largest provider share
4 of 8
Anthropic
Sources evaluated
The task sets these weights before any model is scored.
| Evaluation feed | Weight | Winner result | Field measured |
|---|---|---|---|
| Terminal-Bench 2.1 | 28% | #13 | 13/21 |
| SWE-rebench | 24% | #4 | 16/21 |
| Frontier-Bench | 18% | #4 | 11/21 |
| LiveBench Agentic Coding | 12% | #17 | 18/21 |
| LMArena Agent | 8% | #3 | 15/21 |
| OpenRouter usage | 5% | 97/100 | 21/21 |
| OSWorld Verified | 5% | #2 | 5/21 |
Provider concentration
Each exact model is scored separately; provider identity is not a ranking input.
- Anthropic4 models
- OpenAI2 models
- deepseek1 model
- Z.ai1 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 | 84 | 100% | 21 threads · 13 families · 7 cautions | #2 OSWorld Verified · #3 LMArena Agent |
| 02 | Claude Fable 5Anthropic | 82 | 99% | 6 threads · 4 families · 0 cautions | #1 SWE-rebench · #2 LMArena Agent |
| 03 | Claude Opus 4.8Anthropic | 76 | 100% | 7 threads · 6 families · 4 cautions | #1 Terminal-Bench 2.1 · #3 OSWorld Verified |
| 04 | GLM 5.2Z.ai | 67 | 87% | 8 threads · 6 families · 5 cautions | #8 LMArena Agent · #9 SWE-rebench |
| 05 | GPT-5.6 LunaOpenAI | 65 | 99% | 10 threads · 9 families · 5 cautions | #4 Terminal-Bench 2.1 · #11 Frontier-Bench |
| 06 | Claude Opus 4.6Anthropic | 65 | 81% | 4 threads · 4 families · 0 cautions | #13 LMArena Agent · #16 SWE-rebench |
| 07 | Claude Sonnet 4.6Anthropic | 62 | 82% | 1 threads · 1 families · 1 cautions | #5 OSWorld Verified · #13 SWE-rebench |
| 08 | DeepSeek V4 Flash Vision Expdeepseek | 52 | 24% | 2 threads · 2 families · 0 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 is explicitly designated as the implementation lane for autonomous agent work in production SPARQ protocols, where it handles long-running implementation, review fixes, and subagent delegation despite exhibiting context compaction and steerability issues.
Best when: Use as the lead implementation model in SPARQ agent protocols where it owns autonomous execution, repair loops, and subagent spawning for investigation and testing tasks.
Tips
- Use as the lead implementation model in SPARQ agent protocols where it owns autonomous execution, repair loops, and subagent spawning for investigation and testing tasks.
- Deploy for long-horizon workflows with explicit safety guardrails (AGENTS.md, execute-plan, review-branch, workspace-write sandbox) rather than fine-grained step control.
Watch out for
- Watch for mid-turn steer hijacking where a short interrupt can cause the agent to complete only the latest mini-task and omit the main task from the final response.
- Expect automatic context compaction to break execution reliability; pending actions and shell sessions may be recorded but not reliably resumed after compaction triggers.
- Account for ungovernable effort escalation where the model routinely selects max reasoning even for simple tasks, overriding explicit medium or high configurations.
Claude Fable 5 ranks second on LMArena's agentic arena and is explicitly assigned to plan-review, code-review, advise, and security lanes in production agent protocols, though it trails GPT-5.6 Sol on Agents' Last Exam by 13+ points.
Best when: Assign to review and soundness lanes where SPARQ protocols explicitly designate Fable 5 for plan-review, code-review, advise, and security tasks.
Tips
- Assign to review and soundness lanes where SPARQ protocols explicitly designate Fable 5 for plan-review, code-review, advise, and security tasks.
- Use for analysis and solution design with cheaper models handling execution, then have Fable monitor, validate, and create adjustment tasks.
- Deploy for web-app agent builds where it ranks #3 on Design Arena (Elo 1288) for blind human preference of agent-built results.
Watch out for
- Verify workflow subagent model inheritance; switching models mid-session with /model does not propagate to spawned workflow agents which retain the original session model.
- Expect lower raw task-completion scores than GPT-5.6 Sol; at medium reasoning Sol beats Fable 5 by 11.4 points at roughly one-quarter the cost on Agents' Last Exam.
Claude Opus 4.8 serves as a fallback in production routing chains and has been evaluated for temperature handling in orchestration proxies, with limited direct evidence for autonomous agent performance.
Best when: Use as a fallback model in routing chains where production systems fall back through Opus 4.8 after GLM-5.2.
Tips
- Use as a fallback model in routing chains where production systems fall back through Opus 4.8 after GLM-5.2.
GLM 5.2 is an open-weight model that ranks mid-tier on agentic benchmarks and serves as a fallback option in production routing chains, though it exhibits silent tool-call loss, timeout stalls, and lower Frontier-Bench performance.
Best when: Deploy as a fallback in model routing chains where production systems chain through GLM-5.2 before Claude Opus 4.8 and Claude Fable 5.
Tips
- Deploy as a fallback in model routing chains where production systems chain through GLM-5.2 before Claude Opus 4.8 and Claude Fable 5.
- Use for episodic learning protocols where GLM-5.2 showed +2.9 improvement from attempt 1 to attempt 3 in visible learning metrics.
Watch out for
- Watch for silent tool-call loss when SSE chunks fail JSON parsing; tool calls may disappear mid-action and the turn ends as if the model chose to stop.
- Monitor for complete execution stalls where the model enters a sleeping state with zero output for hours despite normal container and heartbeat metrics, eventually hitting global timeout.
- Expect timeout failures on review tasks; opencode agents with GLM-5.2 may fill 900-second budgets and exit with code 124 without producing review results.
GPT-5.6 Luna serves as a delegated subagent for implementation, investigation, and testing under GPT-5.6 Sol leadership in production Codex workflows, though it ranks significantly lower on agentic benchmarks and suffers from runtime/catalog version mismatches.
Best when: Spawn as a subagent from GPT-5.6 Sol for implementation, investigation, and testing tasks where Luna runs at xhigh reasoning effort.
Tips
- Spawn as a subagent from GPT-5.6 Sol for implementation, investigation, and testing tasks where Luna runs at xhigh reasoning effort.
- Use for volume model routing in Symphony workflows for mechanical, code, and test capabilities where cost efficiency matters more than architecture decisions.
Watch out for
- Check runtime/catalog consistency before spawning; Luna may be rejected as a child agent due to static catalog version (v1) mismatches against effective V2 runtime capabilities.
- Account for significantly lower benchmark standing; Luna ranks #18 of 45 on LMArena's agentic arena with a score of 2.1 versus Sol's 9.8.
- Verify spawn_agent actually creates child sessions; calls with fork_turns="none" may silently fail to create any child session.
Claude Opus 4.6 is noted for reliable long-duration autonomous operation without bugs or stops, and is available through subsidized channels, though it ranks mid-tier on agentic benchmarks.
Best when: Deploy for extended autonomous sessions where the model has demonstrated hours of continuous operation without stopping or requiring intervention.
Tips
- Deploy for extended autonomous sessions where the model has demonstrated hours of continuous operation without stopping or requiring intervention.
- Access through subsidized subscriptions like GitHub CoPilot at $39/month for API agentic harness work while the pricing lasts.
Watch out for
- Expect lower agentic benchmark performance; Opus 4.6 ranks #14 of 45 on LMArena's agentic arena with a score of 5.2, below GPT-5.6 Sol and Claude Fable 5.
Claude Sonnet 4.6 ranks low on LMArena's agentic arena and has limited evidence for autonomous agent use, with one benchmark showing poor computer-use performance.
Best when: Consider for full-stack agent builds where it ranks #12 on Design Arena (Elo 1238) for blind human preference.
Tips
- Consider for full-stack agent builds where it ranks #12 on Design Arena (Elo 1238) for blind human preference.
Watch out for
- Expect very low agentic benchmark scores; Sonnet 4.6 ranks #22 of 45 on LMArena's agentic arena with a score of 1.
- Account for poor computer-use performance; achieves only 9.3% binary accuracy on OSWorld Verified with 500-step budget.
DeepSeek: DeepSeek V4 Flash Vision Exp scores 65.1% on LiveBench Agentic Coding (#3 of 51), covering executable JavaScript, TypeScript, and Python tasks.
Best when: Consider only after reviewing the cited caution.
Frequently asked
- What is the top-ranked model for Autonomous Agents?
- OpenAI: GPT-5.6 Sol ranks first in the current evidence-weighted comparison. Use as the lead implementation model in SPARQ agent protocols where it owns autonomous execution, repair loops, and subagent spawning for investigation and testing tasks.[1][2][3]
- What should I watch out for with OpenAI: GPT-5.6 Sol?
- Watch for mid-turn steer hijacking where a short interrupt can cause the agent to complete only the latest mini-task and omit the main task from the final response.[4]
- What is an alternative to OpenAI: GPT-5.6 Sol?
- Anthropic: Claude Fable 5 is the next-ranked option. Assign to review and soundness lanes where SPARQ protocols explicitly designate Fable 5 for plan-review, code-review, advise, and security tasks.[5]
Sources
- 1
“> 🤖 SPARQ agent\n\nThe current registry routing contract still assigns general implementation to Opus 5 and computes review by inverting the implementer's provider. The maintainer's new default protocol is:\n\n- GPT-5.6 Sol is the implementation lane, including autonomous review fixes.\n- Claude Opus 5 is the review and soundness lane.\n- Capacity handling must remain explicit and fail closed; retired model aliases must stay unreachable.\n- PLAN and CLAIM routing must continue to agree exactly…”
jeswr · GitHub · Sep 2, 2026 - 2
“> 🤖 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 - 3
“## 背景 現在のCodex開発環境では、`GPT-5.6 Sol high` をLeadとして、必要に応じて `GPT-5.6 Luna xhigh` のsubagentへ実装・調査・テスト等を委譲している。 また、以下のような安全策・ワークフローをdotfilesで管理している。 * `AGENTS.md` による行動規則 * `execute-plan` による実装フロー * `review-branch` による独立レビュー * Git/GitHub操作を制限するhook * `workspace-write` Codex Sandbox * commit / push / Draft PRまでの自動化 * protected branchへの直接操作制限 これらによって安全性・再現性は高くなっている一方、近年のフロンティアモデルの能力向上を考えると、**モデルの作業方法まで細かく規定しすぎている可能性がある**。 今後は、 > **Codexの行動そのものを細かく制御するのではなく、安全に自由行動できる環境を制御する** という設計思想へ移行したい。 OpenAI / An…”
Daiki48 · GitHub · Aug 19, 2026 - 4
“## Description With GPT-5.6 Sol at high reasoning, a short mid-turn steer sent during an important task can occasionally cause Hermes to complete only the latest mini-task or added instruction. The final response then omits the main task instead of merging both. ## Steps to reproduce 1. Start a substantial task. 2. While it is running, send a short steer with a small extra task or detail. 3. In some runs, Hermes returns only the steer result. ## Expected behavior Hermes should preserve the main…”
YdocYNj · GitHub · Jul 31, 2026 - 5
“## 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 - 6
“### 📦 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 - 7
“### Description Since 17.1.2, the `task` tool accepts `effort: "lo" | "med" | "hi"`. The tool prompt tells the calling model to choose one based on task complexity, so models routinely send `"hi"` even for relatively simple tasks. `"hi"` maps to the highest level supported by the model and takes precedence over `task.agentModelOverrides`. With `gpt-5.6-sol`, this means `max`, even when I explicitly configure the subagent as `medium` or `high`. There is no setting to cap or disable this override…”
mixmav · GitHub · Jul 25, 2026 - 8
“I have started testing roles separation: -- Stronger model from claude subscription is responsible for analysis solution design task creation. -- Cheaper model (qwen3.6, local deployment on 36GB GPU with 250K context size) is responsible for task execution Doing this for less then 24 hours. Fable can monitor and validate execution and create smaller tasks when something has to be adjusted after qwen execution. It is slower ofc, but this can work non stop without hitting the limits even when usi…”
sermakarevich · Hacker News · Jun 12, 2026 - 9
“Ranks #3 of 37 on Design Arena's web-app agent category (Elo 1288), based on blind human preference between agent-built results.”
Design Arena web-app agents · Benchmark · Sep 4, 2026 - 10
“## Problem Workflow subagents run on the model the session **started with**, not the model the session is **currently on**. Switching models mid-session with `/model` never reaches workflow agent spawning. Evidence from a live fleet session: 1. Started a pi session with model `aigw/round-robin/qwen3.8-max`. 2. Switched to `dongli/claude-fable-5` via `/model` (~1h later). 3. Ran the built-in `code-review` workflow. All 7 agents spawned with `aigw/round-robin/qwen3.8-max` (recorded in the run fil…”
Explorer1092 · GitHub · Aug 11, 2026 - 11
“"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 - 12
“An operator switched the model to GLM-5.2 in the web console, asked the agent what it was running, and was told: > I'm running **GPT-5.6 Sol** (`openai-codex:gpt-5.6-sol`) as the primary model, configured at `effort: high`. > If that model is unavailable I fall back through a chain: Gemini 3.1 Pro → Gemini 3.5 Flash → Kimi K2.7 Code → GLM-5.2 → Claude Opus 4.8 → Claude Fable 5 → and a few others, ending with GPT-5.6 Terra. The override had worked. The run record for that exact turn: GLM-5.2 ser…”
robertsreberski · GitHub · Jul 25, 2026 - 13
“## 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 - 14
“## 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 - 15
“问题现象 chrome 项目 hub_reason 任务 5f7188f7（08-18 03:28 下发）进入 running 后 1.5 小时零产出，但一切监控指标看起来正常： - 容器存活、任务心跳正常（26 秒内） - pi 进程存在（PID 75，状态 sleeping，11 线程） - job 不会失败、不报错——只能等 7200s 全局 timeout 被 Reaper 判死 证据链 pi 会话日志（/workspace/.deepsonar-home/.pi/agent/*.jsonl，共 5 行后无新记录）： 03:29:23.994 model_change → provider=, modelId=GLM-5.2, thinking=high 03:29:24.009 收到任务 prompt（Chrome 审计任务书） 03:29:24.199 模型返回：content=[]（空内容）usage={input:0, output:0} ← 之后 1.5h 无任何日志 进程实况（/proc/75/net/tcp）：只有本地 rpc 监听端口，无任何对外 TCP 连接…”
SummerSec · GitHub · Aug 18, 2026 - 16
“## 증상 `impact-review` 의 opencode 에이전트 레인(`regression`, `architecture`)이 **둘 다 예산 900초를 꽉 채우고 exit 124(타임아웃)로 죽는다.** 결과적으로 AI 리뷰 결과는 항상 0건이고, 결정적(deterministic) 검사만 남는다. opencode 가 `build · glm-5.2` 상태줄까지 출력한 뒤 그대로 멈춘다. **API 에러가 아니라 응답 없는 정지**다. `[timeout]` 은 우리 쪽 `timeout` 래퍼가 찍은 것이다. ## 언제부터인가 **2026-08-09 06:35Z ~ 17:15Z 사이**에 깨졌다. 31분은 리뷰 시간이 아니라 `900초 × 2레인` 을 태운 시간이다. ## 저장소 안에서는 원인이 아니다 - `[email protected]` — 워크플로에 핀 고정, 변경 없음 - `.github/impact-review/opencode.json` — 2026-06-15(#48) 이…”
yeongjunyoo · GitHub · Aug 10, 2026 - 17
“The live Symphony workflow currently hardcodes `/home/timwhite/.local/bin/codex-rotate --config model="gpt-5.6-luna" app-server` for every issue. The checked-in Gem model registry supports mechanical, code, tests, review, semantic, architecture, and root-cause capabilities plus Luna/Terra/Sol routes, but Symphony does not consume any per-issue classification or routing receipt. This sends control-plane and architecture work such as the fleet-gate freshness repair through the volume model and pr…”
itstimwhite · GitHub · Aug 13, 2026 - 18
“### 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 - 19
“Ranks #18 of 45 on LMArena's agentic arena (score 2.1), measuring tool-use and multi-step task performance from human preference.”
LMArena agentic arena · Benchmark · Aug 31, 2026 - 20
“Additional reproduction on macOS with a newer build: - Date: 2026-07-24 KST - Codex Desktop build: `26.721.31836` - Codex CLI: `0.146.0-alpha.3.1` - Parent: `gpt-5.6-sol` / medium - Requested child: `gpt-5.6-luna` / high - Call shape: `spawn_agent(model="gpt-5.6-luna", reasoning_effort="high", fork_turns="none")` Actual result remains: No child session is created. This reproduced in a fresh top-level Codex Desktop task. The local raw model catalog exposes Luna, while the current native `spawn_a…”
CHKKOREA · GitHub · Jul 24, 2026 - 21
“For the first time a model does not bug, does not stop and works for hours. I haven't tested on big development projects yet but for now he surprises me, he managed to do things without even asking him he reminds me a lot Claude Opus 4.6 who was one of the best models for me, he seemed to read minds and I feel the same. There I am in the process of making him do a very big application from A to Z to see how he behaves, then I would understand that he can not do it but it will allow me to see ho…”
jbourny · Hugging Face · Aug 15, 2026 - 22
“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 - 23
“Ranks #14 of 45 on LMArena's agentic arena (score 5.2), measuring tool-use and multi-step task performance from human preference.”
LMArena agentic arena · Benchmark · Aug 31, 2026 - 24
“Ranks #12 of 40 on Design Arena's full-stack agent category (Elo 1238), based on blind human preference between agent-built results.”
Design Arena full-stack agents · Benchmark · Sep 4, 2026 - 25
“Ranks #22 of 45 on LMArena's agentic arena (score 1), measuring tool-use and multi-step task performance from human preference.”
LMArena agentic arena · Benchmark · Aug 31, 2026 - 26
“Achieves 9.3% binary accuracy on OSWorld Verified (#6 of 8) with a 500-step budget, measuring computer use in real desktop environments.”
OSWorld Verified · Benchmark · Jun 24, 2026
Rankings synthesized from community evidence and open benchmarks. See methodology. Not driven by vendor marketing.