Recommendation for Coding
Coding
Our top recommendation for Coding, based on the public evidence we track, is Anthropic: Claude Fable 5.[1][2][3][4][5] Use for one-shot game builds and creative coding projects where you describe the concept in plain English and need working code delivered. Watch out: Watch for silent model fallback where pinned Fable 5 sessions get switched to Opus 5 server-side without notification. Z.ai: GLM 5.2 is the next-ranked alternative. Use for high-context coding tasks with 1M token windows where you need to process large codebases without the cost of proprietary models.
About this recommendation
- Updated
- Sep 4, 2026
- Evidence through
- Sep 4, 2026
- Sources
- 27
- Revision
- v66
Decision audit
Why this result
Inspect the inputs and the computed order behind the recommendation.
Models screened
22
live candidates
Evaluation feeds
8
task-weighted
Winner coverage
93%
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 |
|---|---|---|---|
| LMArena WebDev | 18% | #6 | 16/22 |
| LiveBench Coding | 16% | #2 | 16/22 |
| SWE-rebench | 16% | #1 | 14/22 |
| Design Arena Coding | 14% | #5 | 17/22 |
| Aider Polyglot | 10% | not measured | 4/22 |
| OpenRouter usage | 10% | 87/100 | 22/22 |
| Terminal-Bench 2.1 | 10% | #11 | 10/22 |
| SWE-bench Verified | 6% | not measured | 5/22 |
Provider concentration
Each exact model is scored separately; provider identity is not a ranking input.
- Anthropic4 models
- deepseek1 model
- Google1 model
- OpenAI1 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 | 81 | 93% | 12 threads · 4 families · 5 cautions | #1 SWE-rebench · #2 LiveBench Coding |
| 02 | GLM 5.2Z.ai | 80 | 90% | 34 threads · 5 families · 15 cautions | #7 Design Arena Coding · #9 SWE-rebench |
| 03 | Claude Opus 4.8Anthropic | 76 | 93% | 19 threads · 4 families · 8 cautions | #1 Terminal-Bench 2.1 · #7 LiveBench Coding |
| 04 | Claude Sonnet 5Anthropic | 72 | 93% | 7 threads · 4 families · 4 cautions | #5 Terminal-Bench 2.1 · #10 SWE-rebench |
| 05 | Claude Opus 4.5Anthropic | 64 | 93% | 8 threads · 2 families · 4 cautions | #1 SWE-bench Verified · #12 LiveBench Coding |
| 06 | DeepSeek V4 Flash 0423deepseek | 63 | 90% | 14 threads · 9 families · 3 cautions | #13 LMArena WebDev · #40 Design Arena Coding |
| 07 | o3 ProOpenAI | 17 | 20% | no linked practitioner threads | #1 Aider Polyglot |
| 08 | Gemini 2.5 Pro Preview 06-05Google | 16 | 23% | no linked practitioner threads | #2 Aider Polyglot · #6 SWE-bench Verified |
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.
Ranks #6 on LMArena WebDev and #2 on LiveBench Coding, with demonstrated ability to build complete games from natural language prompts and write working WebAssembly interpreters.
Best when: Use for one-shot game builds and creative coding projects where you describe the concept in plain English and need working code delivered.
Tips
- Use for one-shot game builds and creative coding projects where you describe the concept in plain English and need working code delivered.
- Deploy for low-level systems programming including WebAssembly with GC extensions, where it produced correct assembly on first attempt without debugging.
- Leverage its strong benchmark position for competitive coding scenarios where LiveBench and LMArena rankings predict human preference.
Watch out for
- Watch for silent model fallback where pinned Fable 5 sessions get switched to Opus 5 server-side without notification.
- Budget for higher token costs due to a tokenizer that uses roughly 1.6x-2x more tokens than GPT on equivalent codebases.
Ranks #7 on Design Arena and #12 on LMArena WebDev with LiveBench score of 79.65%, offering million-token context windows and competitive open-weight performance at significantly lower cost than proprietary alternatives.
Best when: Use for high-context coding tasks with 1M token windows where you need to process large codebases without the cost of proprietary models.
Tips
- Use for high-context coding tasks with 1M token windows where you need to process large codebases without the cost of proprietary models.
- Deploy as a cost-effective implementation model after planning with stronger models, where it can execute from context documents at lower token cost.
Watch out for
- Watch for silent truncation at 4096 tokens on some provider configurations where max_output_tokens is not properly passed through.
- Verify availability in your specific client, as some distributions lack the GLM Coding Plan preset despite local builds including it.
Ranks #27 on Design Arena coding with LiveBench score of 81.83%, showing solid debugging capabilities for compiled language issues but trailing newer models.
Best when: Use for debugging compiled language runtime issues, where it successfully identified and fixed Minecraft mod bugs from crash reports and jar files.
Tips
- Use for debugging compiled language runtime issues, where it successfully identified and fixed Minecraft mod bugs from crash reports and jar files.
- Deploy for Ruby static analysis tasks, where it found genuine compile/runtime issues with documented classes and methods.
Watch out for
- Avoid for French text alignment tasks with accentuated characters, where UTF-8 multibyte handling produces incorrect layouts.
- Expect higher token consumption than GPT equivalents due to less efficient tokenizer encoding.
Positioned by users as a workhorse model for implementation after planning with stronger models, with GPT 5.3-codex noted as a more cost-effective alternative at 90% of the capability.
Best when: Use for implementation tasks where you have already established the architecture with stronger planning models.
Tips
- Use for implementation tasks where you have already established the architecture with stronger planning models.
Watch out for
- Evaluate GPT 5.3-codex as a cheaper alternative, which users report offers 90% of the capability at roughly 20% of the cost.
Ranks #30 on Design Arena coding with LiveBench score of 79.65%, noted by users as a peak usefulness point before later versions introduced thinking delays and dead-end loops.
Best when: Use for interactive precision coding where you want direct control without extended thinking delays that obscure the reasoning process.
Tips
- Use for interactive precision coding where you want direct control without extended thinking delays that obscure the reasoning process.
- Deploy for small to medium Python and JavaScript codebases under 500k lines where functional breakage is rare and self-correction works.
Watch out for
- Verify outputs daily as users report verbose code with missed edge cases and misunderstood requirements requiring constant correction.
- Do not use for GPU programming where multiple users report consistent failure across Sonnet, Opus 4.5 and 4.6 variants.
Ranks #13 on LMArena WebDev coding with strong self-hosting potential, though reasoning parser configurations can cause silent verification failures.
Best when: Use for initial PRD and planning drafts where speed matters, then hand off to stronger models for implementation.
Tips
- Use for initial PRD and planning drafts where speed matters, then hand off to stronger models for implementation.
- Deploy for self-hosted coding workflows with vLLM where you control the serving infrastructure and want open-weight flexibility.
Watch out for
- Avoid using with reasoning parsers for self-verification tasks, as this configuration silently collapses scores to 0.5 with arbitrary winner selection.
- Check model ID availability frequently, as provider model lists change and configured IDs may become invalid without notice.
Scores 84.9% on Aider polyglot coding benchmark, placing second among 28 models tested for editing real code across multiple languages.
Best when: Use for polyglot code editing tasks where you need to modify existing codebases across diverse languages with high accuracy.
Tips
- Use for polyglot code editing tasks where you need to modify existing codebases across diverse languages with high accuracy.
Scores 83.1% on Aider polyglot coding benchmark, placing third for cross-language code editing tasks.
Best when: Use for multi-language code editing where Aider benchmark performance indicates strong real-code modification capabilities.
Tips
- Use for multi-language code editing where Aider benchmark performance indicates strong real-code modification capabilities.
Freshness check
New models awaiting independent results
These recently released, tool-capable models are relevant to coding but do not yet have enough independent evidence to place confidently in the ranking above.
- Meta: Muse Spark 1.3 Contributorunranked
Released Sep 2, 2026 · $0.09999999999999999/M input tokens
- Meta: Muse Spark 1.3unranked
Released Sep 2, 2026 · $1.25/M input tokens
- Google: Gemini 3.8 Flashunranked
Released Sep 2, 2026 · $0.75/M input tokens
- Inception: Mercury 2.5 Previewunranked
Released Aug 31, 2026 · $0.04/M input tokens
- Anthropic: Claude Fable 5.1unranked
Released Aug 29, 2026 · $10/M input tokens
- Tencent: Hy4 previewunranked
Released Aug 28, 2026 · $0.834/M input tokens
- Ling 3.0 Flash Finunranked
Released Aug 27, 2026 · $0.06/M input tokens
- Qwen: Qwen3.8 Flashunranked
Released Aug 26, 2026 · $0.15/M input tokens
- IBM: Granite 4.2 8Bunranked
Released Aug 24, 2026 · $0.09999999999999999/M input tokens
- Z.ai: GLM 5.3 Flashunranked
Released Aug 21, 2026 · $0.075/M input tokens
- DeepSeek: DeepSeek V4 Flash Vision Expunranked
Released Aug 21, 2026 · $0.44/M input tokens
Frequently asked
- What is the top-ranked model for Coding?
- Anthropic: Claude Fable 5 ranks first in the current evidence-weighted comparison. Use for one-shot game builds and creative coding projects where you describe the concept in plain English and need working code delivered.[1][2]
- What should I watch out for with Anthropic: Claude Fable 5?
- Watch for silent model fallback where pinned Fable 5 sessions get switched to Opus 5 server-side without notification.[3]
Sources
- 1
“<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 - 2
“<p>Back in 2022 <a href="https://twitter.com/simonw/status/1555626060384911360">I tweeted</a> screenshots of a game concept generated by GPT-3 and some concept "art" created using DALL-E. Today, on the fourth anniversary of that tweet, I decided to see if Claude Fable 5 (running in <a href="https://code.claude.com/docs/en/claude-code-on-the-web">Claude Code for web</a>) could build the entire game from the content of that tweet. It did a pretty good job of it!</p> <p>You can <a href="https://si…”
Engineering publication · Aug 5, 2026 - 3
“**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 - 4
“## 问题描述 / What happened 我在 Apple Silicon Mac 上使用从 cindy.cn 官网下载的中国大陆版 Cindy App 时，发现「设置 → 模型供应商 → 添加供应商」里的供应商预设明显少于本地源码构建版本。 核心影响是：本地源码里的 provider catalog 已经包含智谱 GLM Coding Plan，并且 `claude-code` runtime 下有 `glm-5.2[1m]` / `contextWindow: 1000000`，但 cindy.cn 发布包启动后优先从远端 catalog 拉取 preset，而当前远端 catalog 里没有智谱 Coding Plan 预设，导致官网包无法通过预设选择到 GLM-5.2 的百万上下文能力。 从现象看： - 本地编译/打包的 Cindy App：添加供应商列表里有「智谱 GLM Coding Plan（中国大陆）」和「Z.ai GLM Coding Plan (Global)」等新预设。 - 从 cindy.cn 下载的发布包：添加供应商列表只有普通「智谱 GLM（中国大陆…”
AlfredChaos · GitHub · Jul 31, 2026 - 5
“The big danger here is the gradual increase in open-weight subscription costs. I use open weight subscriptions, with lower-cost models for 80% of my tasks and GLM-5.2, Qwen 3.7-Max, Kimi-K2.6 2.7-Code for the 20% that need the most intelligence. That lets me maximize the rate-limit the subscription gives (rate limits per model are literally a price-limit-per-token model). When new more expensive open weights come in, providers phase out older cheaper models. Over time we will either have to pay…”
0xbadcafebee · Hacker News · Jul 16, 2026 - 6
“Claude is perfectly capable of writing assembly. Here's a working (basic) Prolog interpreter that Claude Fable 5 wrote in WebAssembly in 61 minutes for $16.75 in token costs: https: github.com emk fable-wasm-prolog blob main prolog.wa... WebAssembly is slightly easier than real assembly, but here Fable used WASM GC extensions, which are poorly documented and not yet super common. Fable didn't even need to debug it; I believe essentially all the assembly worked correctly on the first try. I have…”
ekidd · Hacker News · Jul 12, 2026 - 7
“Ranks #6 of 84 on LMArena's WebDev coding arena (Elo 1628), a leaderboard built from blind human preference votes on coding tasks.”
LMArena WebDev (coding) arena · Benchmark · Sep 1, 2026 - 8
“Scores 85.99% on LiveBench Coding (#2 of 51), an objective evaluation of code generation and completion.”
LiveBench Coding · Benchmark · Jun 25, 2026 - 9
“Yeah, Anthropic's current tokenizer in Sonnet 5 Opus 4.8 Fable 5 is much worse than OpenAI's. Also, OpenAI has been using their current o200k_base from the day GPT-4o came out over two years ago. Just a few of my own tests: - A ~2000-2002 legacy C++ game codebase at about ~90kloc: GPT 1.12M, Claude 2.2M - A ~30kloc TypeScript codebase: GPT 260K, Claude 437K In the end, GPT's current tokenizer is ~1.6x-2x better than Claude's current one, depending on your data. And you can check for free for bo…”
Tiberium · Hacker News · Jul 13, 2026 - 10
“I have tested this exact scenario, and it works. Opus 5 had access to IDA over MCP, and I simply asked it HOW certain things were done in the target binary. Purely informational, educational, discovery, it was very helpful creating context documents. Then I took those over to GLM-5.2 to actually accomplish something.”
gfosco · Hacker News · Jul 28, 2026 - 11
“### Summary This report covers two defects on the same code path (openai-chat adapter with a routed third-party provider), observed with `glm-5.2` on Volcengine Ark Coding Plan (`https://ark.cn-beijing.volces.com/api/coding/v3`). **A. Long turns are silently truncated at the upstream default max_tokens (4096)** - Codex desktop never sends `max_output_tokens` in its Responses API requests (verified in the 26.715 desktop binary). - The openai-chat adapter only sets `body.max_tokens` when the inco…”
RanceLee233 · GitHub · Jul 22, 2026 - 12
“### Summary OpenCodex successfully injects 13 routed models into the Codex catalog, and the Codex CLI/app-server can read all of them, but Codex Desktop does not show any routed model in its model picker. The desktop picker only shows the current native models: - 5.6 Sol - 5.6 Terra - 5.6 Luna - 5.5 - 5.4 - 5.4 Mini Expected: routed entries such as `input/glm-5.2`, `input/grok-4.5`, and `input/kimi-k2.7-code` should also appear, as described in the OpenCodex “Codex App model picker” documentati…”
Lingchen97 · GitHub · Jul 22, 2026 - 13
“When using Replay Mod v2.6.27 for Minecraft 26.2, the replays were failing to save in my world that contained teams with non-empty team prefixes. Here's a video of me replicating the bug in single player: [link](https://youtu.be/jGdfIHxSOhE) I fed the jar and the Minecraft crash report to Claude Opus 4.8 and it identified and fixed the bug! [Here's my Claude chat](https://claude.ai/share/b7e5f8b9-a66a-419a-a9ac-7f23bc624d7a) if you're curious. ## Summary `PacketTeam.skipTeamInfo` walks the fiel…”
AgoraPluribusMichael · GitHub · Jul 24, 2026 - 14
“> (Disclaimer: I asked Claude Opus 4.8 to find Spinel compile/runtime issues with documented Ruby classes and methods. It found the following issue, which I manually verified for accuracy before publishing this report) _Reproduced with `spinel bfdebccdaa5b` vs `ruby 4.0.3`._ A base class mixes in `Comparable` and defines ` `; two subclasses instantiate into one array which is sorted (a polymorphic `A B` comparison). Adding a *separate* monomorphic operator call site (`A.new > A.new`) anywhere i…”
KieranP · GitHub · Jul 24, 2026 - 15
“> (Disclaimer: I asked Claude Opus 4.8 to find Spinel compile/runtime issues with documented Ruby classes and methods. It found the following issue, which I manually verified for accuracy before publishing this report) _Reproduced with `spinel bfdebccdaa5b` vs `ruby 4.0.3`._ Follows on from _`each_cons` on a lazy Enumerator returns an untyped `unknown`_ (the direct `lazy…each_cons(n).first(k)` terminal form is now fixed). Chaining `each_cons(n)` onto an `Enumerator::Lazy` and then applying *any…”
KieranP · GitHub · Jul 24, 2026 - 16
“Claude Code with Opus 4.8 is also bad at aligning boxes with content in French (with accentuated letters such as "é" which are multibyte in UTF-8).”
dolmen · Hacker News · Jul 16, 2026 - 17
“I prefer GPT 5.5 to Opus but both are absurdly expensive token hogs, I can't afford to use either as my main model at $work with the monthly spend cap we have. I use Composer (since we use Cursor) or GPT 5.3-codex as my workhorse models and only break out the big guns when I have a genuinely difficult problem to solve. IMO somewhat weirdly 5.3-codex might be the best overall coding model OpenAI have ever released. It's 90% as good as 5.5 and costs about 20% as much, since it's both cheaper per…”
ifwinterco · Hacker News · Jun 30, 2026 - 18
“I feel like Opus 4.5 was the peak in Claude Code usefulness. It was smart, it was interactive, it was precise. In 4.6 and 4.7, it spends a long time thinking and I don't know what's happening, often hits a dead-end and just continues. For a while I was setting Opus 4.5 in Claude Code, but it got reset often. I just canceled my Max plan, don't know where to look for alternatives.”
lukaslalinsky · Hacker News · Apr 24, 2026 - 19
“That depends on the model and the toolkit it uses. In my experience from using Claude Code (Max, Opus 4.5+) intensely for the past six months, I maybe had 3 instances where the implementation broke functionally. And all of these breaking changes were resolved by Claude. Obviously, this won't apply to every context: I work primarily with well-known langs (e.g., Python, JS), small to medium codebases (<500k LoC, for sure), and relatively few co-developers.”
TomasBM · Hacker News · Jul 1, 2026 - 20
“> it produces absolutely stellar world-class code I am using Claude Code with Opus 4.5 and I have to correct it every day. It produces working code but it makes mistakes. The code is more verbose than it should be, misunderstands ignores edge cases, etc. Daily. And I am not a stellar world-class programmer. I am pretty average. I just read what it produces.”
otikik · Hacker News · May 8, 2026 - 21
“Many people averted religion (which I can get behind with), but have never removed the dogmatic thinking that lay at its root. As so many things these days: It's a cult. I've used Claude for many months now. Since February I see a stark decline in the work I do with it. I've also tried to use it for GPU programming where it absolutely sucks at, with Sonnet, Opus 4.5 and 4.6 But if you share that sentiment, it's always a "You're just holding it wrong" or "The next model will surely solve this" F…”
OtomotO · Hacker News · Apr 20, 2026 - 22
“Personally I use Open Code with a copilot sub. Then all models are available in my session with just a model and variants command combo. Makes it super low friction to try different models & combos (my favourite right now is DeepSeek V4 Flash for initial PRD then Fable 5 high for implementation).”
corford · Hacker News · Jul 9, 2026 - 23
“## Summary When the verifier targets an OpenAI-compatible server that runs a **reasoning parser** — e.g. vLLM with `--reasoning-parser deepseek_v4`, which is exactly how the DSpark DeepSeek-V4-Flash recipes serve the model — every score silently collapses to 0.5 and `select()` returns an arbitrary winner. No error is raised. This hits the library's headline self-verification use case (deepseek-v4-flash judging its own trajectories) for anyone self-hosting the model rather than using the DeepSee…”
hsrakri · GitHub · Aug 18, 2026 - 24
“That's great. Personally, I'd interested in Qwen3.6-27B and deepseek V4 flash (or pro), with contexts above 60k. They seem to be popular and have good coding performance. I'd appreciate numbers on a single or two GPUs where a quantized version fits reasonably into the VRAM (Qwen in 16 or 24GB). 4 older GPUs approach a used 3090 in price, and the 3090 has better support for speedups like MTP. So cheaper but slower looks like a reasonable target to me.”
russianGuy83829 · Hacker News · Jul 13, 2026 - 25
“The weekly model audit found configured LLM model IDs that the provider no longer serves. Update `.github/scripts/llm_common.py` (or the matching env override) to a live ID per the candidates below, and verify with a 1-token call. - **nvidia_nim**: configured model `deepseek-ai/deepseek-v4-flash` is NOT in the live model list. Nearby candidates: `deepseek-ai/deepseek-coder-6.7b-instruct`, `deepseek-ai/deepseek-v4-flash-0731`”
github-actions[bot] · GitHub · Aug 8, 2026 - 26
“Scores 84.9% on the Aider polyglot coding benchmark (#2 of 28), which tests editing real code across many languages.”
Aider polyglot benchmark · Benchmark · Sep 4, 2026 - 27
“Scores 83.1% on the Aider polyglot coding benchmark (#3 of 28), which tests editing real code across many languages.”
Aider polyglot benchmark · Benchmark · Sep 4, 2026
Rankings synthesized from community evidence and open benchmarks. See methodology. Not driven by vendor marketing.