Recommendation for Agentic coding
Agentic Coding
Our top recommendation for Agentic Coding, based on the public evidence we track, is Anthropic: Claude Fable 5.[1][2][3] Use for full-stack agent builds where human preference matters, as it ranks #5 on Design Arena's blind evaluation of agent-built results. Watch out: Watch for silent fallback to claude-opus-5 in Claude Code when fable-5 is pinned, which occurs repeatedly without recovery in version 2.1.220. OpenAI: GPT-5.6 Sol is the next-ranked alternative. Use as a stable fallback when newer Sol variants degrade, as practitioners report switching back from Sol 6 due to quality regression.
About this recommendation
- Updated
- Sep 25, 2026
- Evidence through
- Sep 25, 2026
- Sources
- 25
- Revision
- v84
Decision audit
Why this result
Inspect the inputs and the computed order behind the recommendation.
Models screened
20
live candidates
Evaluation feeds
8
task-weighted
Winner coverage
100%
intended feed weight
Largest provider share
3 of 7
Anthropic
Sources evaluated
The task sets these weights before any model is scored.
| Evaluation feed | Weight | Winner result | Field measured |
|---|---|---|---|
| Terminal-Bench 2.1 | 22% | #12 | 13/20 |
| SWE-rebench | 18% | #1 | 11/20 |
| Frontier-Bench | 14% | #7 | 11/20 |
| LiveBench Agentic Coding | 14% | #7 | 19/20 |
| Design Arena Full-stack | 10% | #6 | 13/20 |
| LMArena Agent | 9% | #2 | 14/20 |
| Design Arena Web-apps | 8% | #8 | 14/20 |
| OpenRouter usage | 5% | 79/100 | 20/20 |
Provider concentration
Each exact model is scored separately; provider identity is not a ranking input.
- Anthropic3 models
- OpenAI2 models
- xiaomi1 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 | Claude Fable 5Anthropic | 75 | 100% | 5 threads · 5 families · 1 cautions | #1 SWE-rebench · #2 LMArena Agent |
| 02 | GPT-5.6 SolOpenAI | 74 | 84% | 28 threads · 19 families · 10 cautions | #4 LMArena Agent · #4 SWE-rebench |
| 03 | GLM 5.2Z.ai | 66 | 90% | 21 threads · 9 families · 12 cautions | #9 SWE-rebench · #11 Design Arena Full-stack |
| 04 | GPT-5.6 LunaOpenAI | 62 | 84% | 20 threads · 16 families · 7 cautions | #4 Terminal-Bench 2.1 · #13 Frontier-Bench |
| 05 | Claude Opus 4.6Anthropic | 60 | 85% | 4 threads · 2 families · 1 cautions | #10 LMArena Agent · #13 Design Arena Full-stack |
| 06 | Claude Opus 5.5Anthropic | 51 | 26% | 4 threads · 4 families · 1 cautions | #2 LiveBench Agentic Coding |
| 07 | MiMo-V2.6-Proxiaomi | 46 | 16% | 3 threads · 3 families · 3 cautions | #3 Design Arena Web-apps |
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.
Claude Fable 5 ranks #2 on LMArena's agentic arena for tool-use and multi-step tasks, though it suffers from silent fallback issues in Claude Code and availability gaps in some agent platforms.
Best when: Use for full-stack agent builds where human preference matters, as it ranks #5 on Design Arena's blind evaluation of agent-built results.
Tips
- Use for full-stack agent builds where human preference matters, as it ranks #5 on Design Arena's blind evaluation of agent-built results.
- Deploy for one-shot game and application generation from natural language premises, demonstrated in building a complete Raccoon Heist game.
Watch out for
- Watch for silent fallback to claude-opus-5 in Claude Code when fable-5 is pinned, which occurs repeatedly without recovery in version 2.1.220.
- Verify availability in your specific agent platform, as Cursor's nightly scan shows composer-2 and other models are rejected while fable-5 variants may have inconsistent registration.
GPT-5.6 Sol serves as a workhorse for Codex CLI agent workflows, though transport errors and safety false positives plague OpenAI's Codex OAuth integration.
Best when: Use as a stable fallback when newer Sol variants degrade, as practitioners report switching back from Sol 6 due to quality regression.
Tips
- Use as a stable fallback when newer Sol variants degrade, as practitioners report switching back from Sol 6 due to quality regression.
Watch out for
- Expect transport errors and fetch failures in long Codex OAuth sessions, with 19 of 37 transport errors in one scan attributed to gpt-5.6-sol and gpt-5.6-terra combined.
- Watch for safety-check false positives on legitimate systems research involving fuzzers, sanitizers, and coverage tools, which can block authorized development work.
- Verify context window handling when using custom proxies, as Codex OAuth models may resolve to direct-API catalog windows (1,050,000 vs 272,000 tokens) causing compaction failures.
GLM 5.2 offers competitive open-weight agentic coding with Design Arena and LMArena rankings near the top 10, though it exhibits planning-to-execution gaps and higher verbosity than Sonnet alternatives.
Best when: Use for CI/CD triggered agent workflows where session persistence and PR generation are required, as demonstrated in extended production use.
Tips
- Use for CI/CD triggered agent workflows where session persistence and PR generation are required, as demonstrated in extended production use.
Watch out for
- Watch for execution gaps where the model plans correct commands in thinking blocks but emits truncated or malformed tool calls, such as dropping leading path segments from bash commands.
- Expect ~20% more verbose code submissions and higher reasoning token usage compared to Sonnet 5, which narrows the cost advantage in agent workflows.
- Note that Frontier-Bench resolution rates remain modest at 4.59% when paired with Claude Code at max effort, suggesting limitations on complex agent tasks.
GPT-5.6 Luna functions as a faster, lighter variant in the 5.6 family for Codex agent workflows, though it shares the transport instability and provider registration gaps affecting all OpenAI Codex models.
Best when: Use for rapid requirement extraction and structured output tasks where Pydantic JSON schema validation is required, as demonstrated in two-stage analysis pipelines.
Tips
- Use for rapid requirement extraction and structured output tasks where Pydantic JSON schema validation is required, as demonstrated in two-stage analysis pipelines.
Watch out for
- Expect HTTP 400 and 500 errors on first turns when using OpenAI Responses protocol through certain providers, with Luna specifically flagged in provider compatibility issues.
- Verify model registration in Cursor and other agent platforms, as Luna is among the newer model families missing from hardcoded provider lists.
Claude Opus 4.6 remains a reference point for reliable long-horizon agent execution with strong reasoning over complex codebases, though it ranks lower on preference-based benchmarks than newer alternatives.
Best when: Use for extended autonomous sessions requiring hours of stable execution without interruption, as practitioners report it 'does not bug, does not stop and works for hours'.
Tips
- Use for extended autonomous sessions requiring hours of stable execution without interruption, as practitioners report it 'does not bug, does not stop and works for hours'.
- Deploy for complex reasoning over multiprocess code and security-sensitive analysis, demonstrated in Mozilla's fuzzing harness for sandbox escape detection.
Watch out for
- Note lower blind preference rankings (#18 on Design Arena web-app category) compared to newer Fable and Opus variants, suggesting human evaluators favor alternatives for visible outputs.
- Watch for environment assumption failures when remote execution contexts lack expected tooling, such as missing Python environments for parquet schema inspection.
Claude Opus 5.5 delivers top-tier LiveBench agentic coding scores at reduced cost, though it faces availability restrictions in some Claude Code versions and Vercel rate limiting in production deployments.
Best when: Use for large-scale code audits and security-critical reviews where it uniquely catches issues missed by cheaper API models, as demonstrated in a 3,600 line payment PR analysis.
Tips
- Use for large-scale code audits and security-critical reviews where it uniquely catches issues missed by cheaper API models, as demonstrated in a 3,600 line payment PR analysis.
- Deploy for cost-sensitive agent workflows requiring Opus-level capability, with 40% lower typical workload costs and 20% reduced per-token pricing versus Opus 5.
Watch out for
- Verify Claude Code version compatibility before deployment, as version 2.1.278 rejects Opus 5.5 with HTTP 400 requiring 2.1.280 or newer.
- Watch for infrastructure rate limiting when using API routes, with Vercel 429 errors blocking execution in comparative evaluations.
MiMo-V2.6-Pro serves as a lower-tier implementation model in tiered agent workflows, though evidence shows timeout failures on simple tasks and misses critical issues caught by Opus models.
Best when: Use as a mechanical implementation tier when a detailed plan from a frontier model is already available, preserving expensive reasoning models for planning work.
Tips
- Use as a mechanical implementation tier when a detailed plan from a frontier model is already available, preserving expensive reasoning models for planning work.
Watch out for
- Expect timeout failures on basic one-file write tasks, with three consecutive 180-second wall timeouts observed in quack runs against simple card implementations.
- Watch for critical issue misses in security-sensitive code review, as it failed to catch a payment flow bug that Opus 5.5 identified in a 3,600 line PR.
Frequently asked
- What is the top-ranked model for Agentic Coding?
- Anthropic: Claude Fable 5 ranks first in the current evidence-weighted comparison. Use for full-stack agent builds where human preference matters, as it ranks #5 on Design Arena's blind evaluation of agent-built results.[1]
- What should I watch out for with Anthropic: Claude Fable 5?
- Watch for silent fallback to claude-opus-5 in Claude Code when fable-5 is pinned, which occurs repeatedly without recovery in version 2.1.220.[2]
- What is an alternative to Anthropic: Claude Fable 5?
- OpenAI: GPT-5.6 Sol is the next-ranked option. Use as a stable fallback when newer Sol variants degrade, as practitioners report switching back from Sol 6 due to quality regression.[3]
Sources
- 1
“Ranks #5 of 41 on Design Arena's full-stack agent category (Elo 1283), based on blind human preference between agent-built results.”
Design Arena full-stack agents · Benchmark · Sep 24, 2026 - 2
“**Bug Description** # Bug report: `claude-fable-5` silently falls back to `claude-opus-5` (13 occurrences, one-way, never recovers) ## Environment | Item | Value | |---|---| | Claude Code | 2.1.220 | | OS | Windows 11 Pro, Build 26200 (x64) | | Node | v24.14.1 | | `settings.json` model | `claude-fable-5[1m]` | | `effortLevel` | `high` | | Concurrent sessions | 8 | ## Summary With `claude-fable-5[1m]` pinned in `settings.json`, sessions are repeatedly switched to `claude-opus-5` by a server-emit…”
XrentX · GitHub · Aug 2, 2026 - 3
“Update: Sol 6 is a heaping pile of garbage. Just epic levels of slop. And r codex etc is full of people noticing the same. I've switched back to 5.6 Sol. What they're selling as Sol 6 is really what would have been Terra before, and it's awful.”
cmrdporcupine · Hacker News · Sep 23, 2026 - 4
“<p><strong><a href="https://simonw.github.io/raccoon-heist-codex/">Moonlight & Mayhem (Raccoon Heist by Codex + GPT-5.6 Sol Ultra)</a></strong></p> On Wednesday I wrote about <a href="https://simonwillison.net/2026/Aug/5/raccoon-heist/">One-shotting a Raccoon Heist game using Claude Fable 5</a>, where I had Claude Fable 5 build a full working game from a premise I generated with GPT-3 and DALL-E <a href="https://twitter.com/simonw/status/1555626060384911360">four years ago</a>.</p> <p>I dec…”
Engineering publication · Aug 7, 2026 - 5
“# Nightly Bug Scan (2026-09-13) Failed to start Cursor agent. - Message: Cannot use this model: composer-2. Available models: default, grok-4.6, composer-2.5, claude-opus-5, claude-opus-4-8, gpt-5.6-sol, gpt-5.5, claude-fable-5-1, claude-fable-5, grok-4.5, gemini-3.8-flash, gemini-3.7-flash, muse-spark-1.3, gpt-5.6-terra, claude-sonnet-5, claude-sonnet-4-6, gpt-5.3-codex, claude-opus-4-7, gpt-5.4, claude-opus-4-6, claude-opus-4-5, gpt-5.2, gpt-5.6-luna, gemini-3.6-flash, gemini-3.1-pro, gpt-5.4…”
github-actions[bot] · GitHub · Sep 13, 2026 - 6
“## 背景 Pi Web Desktop 中 openai-codex（ChatGPT 订阅 OAuth）会话频繁出现「输出一半停止 / fetch failed / terminated」，其他模型（aliyun/deepseek）偶发但次数少。 ## 实锤数据（2026-08-01，scripts/session-stops.mjs 修复后全量扫描 185 会话） - transport-error 共 37 条：openai-codex/GPT 19（gpt-5.6-sol×11 + gpt-5.6-terra×8）、aliyun 9、pi-router 8、deepseek 1。 - 最大异常会话：311 消息 / ~2.2M tokens，会话开头连续 6 次 fetch failed/terminated。 - compaction 仅 1 次且在非中断点，**不是主因**；主因是 transport error。 ## 工具修复（commit bfe020d） session-stops.mjs 原判定只在独立 type=error 事件时识别 transport-er…”
dust617 · GitHub · Aug 1, 2026 - 7
“### What version of Codex CLI is running? 0.145.0 ### What subscription do you have? Pro ### Which model were you using? gpt-5.6-sol ### What platform is your computer? _No response_ ### What terminal emulator and version are you using (if applicable)? _No response_ ### Codex doctor report ### What issue are you seeing? While implementing standard C software quality tests for a private PostgreSQL Bot project, legitimate work involving libFuzzer, sanitizers, code coverage, parser robustness test…”
JustCoding247 · GitHub · Jul 24, 2026 - 8
“### Bug Description When a Codex OAuth model (`gpt-6-astra`, `gpt-5.6-sol`, `gpt-5.6-terra`, `gpt-5.6-luna`) is served through a custom `api_mode: codex_responses` provider (CLIProxyAPI, any local proxy), `get_model_context_length()` resolves the **direct-API catalog window** instead of the **Codex OAuth window**. For `gpt-6-astra` that is `1,050,000` instead of `272,000`. The compressor then sets its threshold at 50% of the wrong window, so conversations grow to ~525K tokens before compaction.…”
0xble · GitHub · Sep 19, 2026 - 9
“If you haven't yet you should give a chance to https: pi.dev I've been using it exclusively (and extending it, see https: a.l3x.in ai) for months with mainly GLM-4.7 then 5.1 and now 5.2 and I could hardly be any happier. I'm still working on a "Github Forgejo first" based workflow but also quite happy with it already, basically most of my sessions run as a ci cd job (triggered by " pi" comments) and generate PRs or push commits to PRs, see https: github.com shaftoe pi-coding-agent-action”
alexfortin · Hacker News · Jul 2, 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
“In our coding evaluations, we found Sonnet 5 is more capable than Sonnet 4.6 (which was an underrated model itself), but is now faster and slightly cheaper. Sonnet 5's performance is comparable to GLM 5.2 in both one-shot coding and agentic ability. However, it's about ~20% less verbose than GLM 5.2 in average code submission sizes, and uses fewer reasoning tokens, which reduces the cost gap and suggests it writes cleaner code. In practice, Sonnet 5 ends up being 40% more expensive and ~2x fast…”
gertlabs · Hacker News · Jul 1, 2026 - 12
“using Claude Code at max effort, resolves 4.59% ± 0.97 of Frontier-Bench tasks (#16 of 16 public submissions), measuring a submitted model-and-agent configuration on diverse, difficult agent work.”
Frontier-Bench · Benchmark · Jul 19, 2026 - 13
“## 목적 및 담당 범위 자연어 서비스 요구사항을 입력받아 Requirements와 Gaps로 구조화하고, UI에서 확인할 수 있는 초기 분석 모듈을 구현한다. - 담당: 권지연 - 범위: Input UI → Requirement Extraction / Structuring → Gap Analysis → Handoff - Clarification 이후 기능은 이번 구현 범위에서 제외한다. ## 구현 내용 - [x] 자연어 입력 UI와 FastAPI 연결 - [x] Mock / SnowChat Provider 선택 구조 및 환경 변수 설정 - [x] Requirement Extraction과 Gap Analysis의 호출·Prompt·Schema 분리 - [x] 모델 목록 확인 후 두 단계 모두 `gpt-5.6-luna` 사용 - [x] Pydantic JSON Schema 기반 strict Structured Output 및 반환값 검증 - [x] Requirement–Gap 참조…”
Jiyeon125 · GitHub · Sep 24, 2026 - 14
“### What happened? Every model on the `opencode-go-responses` provider (`grok-4.5`, `muse-spark-1.2-contributor`, `gpt-5.6-luna`) fails with HTTP 400 on the first turn from the Codex desktop app. Models on `opencode-go` (chat completions) are unaffected, which is what made this hard to see. `buildRoutedRequest` keeps `type: "namespace"` and `type: "custom"` tools whenever `provider.protocol === "openai-responses"`, and `needsZenFreeToolCompatibility` is scoped to `opencode-free` / `opencode-fre…”
pablorodriguesudesc · GitHub · Aug 26, 2026 - 15
“This issue might be a duplicate of several existing open issues — it appears to aggregate multiple pre-existing per-model failures into one report. Please check the following before expecting a fresh investigation: - #46124: `gpt-5.6-luna` HTTP 500 via `/zen/go/v1` - #39873: `mimo-v2-pro` and `mimo-v2-omni` always fail with 'Upstream request failed' - #39874: `hy3-preview` 404 'No allowed providers' - #40886: `grok-4.5` HTTP 503 'Endpoint is unavailable' - #47049: `muse-spark-1.3-contributor` 4…”
github-actions[bot] · GitHub · Sep 5, 2026 - 16
“## Problem The `cursor-agent` provider registers a hardcoded list of 6 models that has not been updated since the provider landed (#1112). Against cursor-agent CLI `2026.08.11-e8db854`, **5 of the 6 registered IDs no longer exist** in the CLI, and every model family shipped since (gpt-5.6-sol/terra/luna, claude-sonnet-5, claude-opus-5, gpt-5.3-codex, kimi-k3, glm-5.2, gemini-3.x, cursor-grok-4.6) is missing from GSD model selection. ## Root Cause `extensions/cursor-cli/models.js:8-62` — `CURSOR…”
rager306 · GitHub · Aug 20, 2026 - 17
“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 - 18
“From Mozilla post [1]: "...After fixing the initial set of issues that Anthropic sent to us in February, we built our own harness atop our existing fuzzing infrastructure. We began with small-scale experiments prompting the harness to look for sandbox escapes with Claude Opus 4.6. Even with this model, we identified an impressive amount of previously-unknown vulnerabilities which required complex reasoning over multiprocess browser engine code..." So yeah, Anthropic and Mozilla likely compare "…”
ZrArm · Hacker News · May 22, 2026 - 19
“Ranks #18 of 40 on Design Arena's web-app agent category (Elo 1210), based on blind human preference between agent-built results.”
Design Arena web-app agents · Benchmark · Sep 24, 2026 - 20
“I was trying to use Opus 4.6 in Claude Code to add some functionality to python code intended to run on a cluster and it didn't have any python environment in its remote environment. It needed to look at the schema of a parquet file to make sure it did things right and couldn't figure out how to do so with code because for god knows what reason there is no python environment in the dev environment for code intended to be run on a compute cluster in Python. Eventually it decided to just examine…”
ls612 · Hacker News · Jun 6, 2026 - 21
“## 目的 Claude のサブスクリプション（Opus 5.5）を、pi-web-idd の lane の主力にします。#30 で導入する Claude Code の非対話 adapter を土台に、Issue lane と review lane を実際に Claude のサブスクリプションで運用するための残りの作業を整えます。 ## 背景 - 2026-09-24 の比較では、3,600 行の決済 PR（Meltly #5638）の must（取得の失敗中でも決済に進め、画面に出ていない商品が支払いに含まれる）を見つけられたのは、Claude Code 上の Opus だけでした。API の安い model（glm-5.3-flash・v4.1-flash・glm-5.3 high・mimo-v2.6-pro）は見落とし、API の Opus 5.5 は Vercel の 429 により実行できませんでした。 - 仕事量は、1 か月で Issue 67 本（約 126 ポイント）、他人の PR のレビュー 67 本、自分の PR 約 69 本とその指摘への対応です。 - この量…”
penne-0505 · GitHub · Sep 24, 2026 - 22
“# Hacker News Claude Opus 5.5 # Claude Opus 5.5 发布摘要 Anthropic 推出了 Claude 5.5 系列的首个模型——**Claude Opus 5.5**。该模型在多数工作任务上的表现可与 Claude Fable 5.1 媲美，但运行成本比 Opus 5 降低了 40%。 ### 核心改进 #### 1. 性能与效率 * **综合能力**：Opus 5.5 在智能编程（Agentic coding）、计算机使用（Computer use）和知识工作领域处于领先地位。它在处理大规模代码库迁移、软件效率优化及复杂逻辑任务（如从单一提示词构建游戏）方面表现出色。 * **编程与知识工作**：在代码审计和大规模迁移任务中效率极高。在知识工作评估（如 GDPval-AA v2.1）中，其表现优于 Fable 5.1 和 Opus 5。 * **成本降低**：相比 Opus 5，典型工作负载的成本降低了 40%。 * **输入 Token**：$4 / 百万（降低 20%） * **输出 Token**：$20 / 百万（降低 20%）…”
github-actions[bot] · GitHub · Sep 24, 2026 - 23
“## Nález (songplayer session, dev1, 2026-09-23 ~08:40 UTC) Managed env `CLAUDE_CODE_SUBAGENT_MODEL` na dev1 = `claude-opus-5-5` — t. j. Opus 5 rodina, ktorú `model-awareness.md` označuje za BANNED (`airuleset.BANNED_MODELS`, hook `block-banned-model.sh` odmieta ju na Agent/Workflow) a ktorú aktuálny Claude Code (2.1.278 na dev1) nepodporuje: model-less `general-purpose` review dispatch z autopilot-worker lane skončil HTTP 400 „Claude Code 2.1.278 does not support this model; version 2.1.280 or…”
zbynekdrlik · GitHub · Sep 23, 2026 - 24
“# Frontier models plan, lower-tier models implement — instruction sheet + global wiring ## What this is Albert's frontier models — GPT-Sol-6 (medium) in Codex, GLM 5.3 MAX in ZCode, Mimo v2.6 pro in MiMoCode — currently do mechanical implementation themselves even when a fully-specced plan already exists. That wastes the most expensive tier on work a lower tier (GPT-6 Luna, GLM 5.3 Flash, Mimo v2.6 flash) can do from a plan that passes the implementation-plan standard (zero questions, per-step…”
u2giants · GitHub · Sep 24, 2026 - 25
“STREAM: swarm: cards The pro spread (`nova-sprint routes --tier pro --label `, #3641; provider = fnv32a(label) mod N) gives OpenRouter's pro slot to `ormimo26pro` = openrouter/xiaomi/mimo-v2.6-pro, state held ("too few to say (U=3)"). Evidence, quack runs of 2026-09-25: it ran three one-line cards (write one file, commit) and all three ran out the 180 s wall with no RESULT: - quack-0925b s00-0302-quack-hulk-pro (04:00Z; repo not staged, every model had to clone, but deepseek/qwen/glm still fini…”
rowan-claude · GitHub · Sep 25, 2026
Rankings synthesized from community evidence and open benchmarks. See methodology. Not driven by vendor marketing.