Recommendation for Agentic coding
Agentic Coding
The best LLM for agentic coding is Claude Fable 5, which dominates LMArena's agentic arena with a score of 13.9, nearly 50% higher than its nearest competitor. Claude Opus 4.8 takes second place with the strongest benchmark suite across SWE-bench Pro (69.2%), Terminal-Bench (85%), and NL2Repo (69.7%), making it the most well-documented workhorse for production agent work. GPT-5.5 claims the third spot with a 70% score on DeepSWE and solid Terminal-Bench performance (84%), though it trails Opus 4.8 on most other agentic benchmarks. Claude Opus 4.7 and Sonnet 5 round out the top five, offering strong agentic capabilities at different price points. The gap between Fable and everything else is substantial, but real-world reports suggest Opus 4.8 is the model people actually ship with. Sonnet 5 appears to be the cost-efficient workhorse, outperforming its predecessor Sonnet 4.6 at lower inference settings. Users report that agentic models can spiral without oversight, so reliability matters as much as raw benchmark numbers.
About this recommendation
- Updated
- Jul 17, 2026
- Evidence through
- Jul 17, 2026
- Sources
- 19
- Revision
- v5
Fable 5 is the benchmark king for agentic coding, sitting at #1 on LMArena's agentic arena with a score of 13.9, significantly ahead of every other model. Evidence suggests it drafts architecture diagrams better than junior developers and understands plain English context well. However, there's limited real-world production evidence compared to Opus 4.8, and one switcher moved from Fable to Grok 4.5. If you want the best raw tool-use and multi-step task performance, this is the pick.
Best when: You need maximum agentic performance for complex, multi-step coding tasks and want the top-ranked model on human preference benchmarks.
Tips
- Ranks #1 of 25 on LMArena's agentic arena with a score of 13.9, the highest by a wide margin.
- Excels at system architecture tasks, drafting design diagrams better than junior developers.
- Understands plain English context without requiring precise prompting.
Opus 4.8 is the most thoroughly documented agentic coding model, with benchmark wins across SWE-bench Pro (69.2%), Terminal-Bench (85%), and NL2Repo (69.7%). Real-world usage shows it can navigate complex development workflows independently, finding its way around Xcode tasks that previously required manual intervention. Some users don't find it significantly better than Sonnet 4.5, and it can slip as project context grows. Still, the breadth of evidence makes it the default choice for production coding agents.
Best when: You want the most battle-tested model with strong benchmarks across multiple agentic coding domains and proven real-world performance.
Tips
- Ranks #2 on LMArena's agentic arena with a score of 9.3.
- Leads on SWE-bench Pro (69.2%), Terminal-Bench (85%), and NL2Repo (69.7%), beating GLM-5.2 and GPT-5.5 on most measures.
- Can autonomously navigate Xcode workflows that previously required manual intervention, improving significantly from version 4.7.
- Offers 1M context window with only 3% typical utilization, leaving room for large codebases.
Watch out for
- One user reports it's not significantly better than Sonnet 4.5 for actual agentic coding work.
- Scores only 58% on DeepSWE, trailing GPT-5.5's 70%.
GPT-5.5 is a strong agentic coder with a 70% DeepSWE score, beating Opus 4.8 on that benchmark. It handles Terminal-Bench well (84%, just behind Opus's 85%) and gets out of the way during coding sessions. However, it trails Opus 4.8 significantly on SWE-bench Pro (58.6% vs 69.2%) and NL2Repo (50.7% vs 69.7%). The Codex integration has regressed with baffling approval prompts for basic actions. Good as a worker model or specialist, but not the top coordinator.
Best when: You need a model that performs well specifically on DeepSWE benchmarks or want to pair it with Claude for code review.
Tips
- Scores 70% on DeepSWE, beating Opus 4.8's 58% and substantially ahead of other competitors.
- Scores 84% on Terminal-Bench, competitive with Opus 4.8's 85%.
- Ranks #5 on LMArena's agentic arena with a 7.2 score.
- Works well in multi-agent setups where Fable controls GPT-5.5 agents for code generation.
Watch out for
- Trails Opus 4.8 significantly on SWE-bench Pro (58.6% vs 69.2%).
- Scores only 50.7% on NL2Repo compared to Opus 4.8's 69.7%.
- Codex integration has regressed, now rejecting basic actions that were previously allowed.
- One comparison shows Kimi K3 beating GPT-5.5 on 30 of 34 benchmarks.
Opus 4.7 sits at #4 on LMArena's agentic arena with a 7.7 score, ahead of GPT-5.5. It was a notable improvement over 4.6 for autonomous reasoning, though 4.8 has since surpassed it. The model can slip as project context grows, and users haven't found whipping or cursing helps. It's a solid predecessor to 4.8 but hard to recommend when 4.8 exists with better real-world navigation and stronger documentation.
Best when: You want a strong agentic model but 4.8 isn't available or you're optimizing cost within the Opus line.
Tips
- Ranks #4 on LMArena's agentic arena with a 7.7 score, ahead of GPT-5.5.
- Improved autonomous reasoning over Opus 4.6, handling Xcode workflows more smoothly.
Watch out for
- Slips and degrades as project context grows, requiring intervention.
Sonnet 5 is the cost-efficient workhorse of the Claude lineup. Running it at low outperforms Sonnet 4.6 at any level while costing less. It writes cleaner code than GLM 5.2 with about 20% less verbosity and fewer reasoning tokens. However, optimization for fully agentic development may have made it worse for agent-assisted work, and some users aren't positive about the direction. Best used as a worker under an Opus-planned task rather than as a top-level coordinator.
Best when: You need a fast, cost-efficient model for executing tasks planned by a stronger coordinator like Opus.
Tips
- Ranks #3 on LMArena's agentic arena with a score of 8.
- Running at low outperforms Sonnet 4.6 at any level while being cheaper.
- Writes cleaner code than GLM 5.2, using about 20% fewer tokens per submission.
- More capable than Sonnet 4.6 while being faster and slightly cheaper.
Watch out for
- Optimization for fully agentic development may make it worse for agent-assisted development.
- Still about 40% more expensive than expected in practice despite efficiency gains.
Grok 4.5 has a passionate endorsement as "the best coding AI agent" one user has tried, enough to switch away from Fable. It ranks #8 on LMArena's agentic arena with a 5.0 score, which puts it in the conversation but below the top tier. Earlier versions struggled with agentic tool calling, treating it as secondary to language generation. Worth testing if you're exploring options, but the benchmark gap behind Fable and Opus 4.8 is substantial.
Best when: You want to experiment with a non-Anthropic model or have had success with xAI's ecosystem.
Tips
- One user calls Grok Build with Grok 4.5 the best coding AI agent they've used, switching from Fable.
- Punches above its weight in language fluidity.
Watch out for
- Ranks #8 on LMArena's agentic arena with a 5.0 score, well behind Fable's 13.9.
- Earlier versions had trouble with agentic tool calling, treating it as a secondary modality.
Sonnet 4.5 remains a perfectly serviceable model for agentic coding, with one user unable to find significant differences between it and Opus 4.8. That said, it lacks the benchmark dominance of newer models and has been superseded by Sonnet 5 as the cost-efficient option. If you're already using it without issues, there's no urgent need to switch, but new projects should probably start with Sonnet 5 or Opus 4.8.
Best when: You're already using it in production and have it dialed in, or you don't need the absolute best performance.
Tips
- One user finds it not significantly different from Opus 4.8 for agentic coding in practice.
Frequently asked
- Which model has the highest agentic coding benchmark scores?
- Claude Fable 5 ranks #1 on LMArena's agentic arena with a score of 13.9. Claude Opus 4.8 leads on specific benchmarks like SWE-bench Pro (69.2%), Terminal-Bench (85%), and NL2Repo (69.7%).
- Is GPT-5.5 good for agentic coding?
- GPT-5.5 scores 70% on DeepSWE, beating Opus 4.8's 58% on that specific benchmark. It ranks #5 on LMArena's agentic arena with a 7.2 score, making it competitive but not top-tier.
- What's the best budget-friendly model for coding agents?
- Claude Sonnet 5 at low inference settings outperforms Sonnet 4.6 at any level while being cheaper, making it a strong workhorse for Opus-planned tasks.
- Can coding agents work fully autonomously?
- Users report that models optimized for full autonomy can wreck codebases without oversight. Most successful setups use a frontier model as coordinator with human review, rather than fully hands-off operation.
Sources
- 1
“Ranks #1 of 28 on LMArena's agentic arena (score 13.2), measuring tool-use and multi-step task performance from human preference.”
LMArena agentic arena · Benchmark · Jul 20, 2026 - 2
“I think you are underestimating how fast this stuff is moving. Calling AI "just autocomplete" for system architecture is outdated. If you were to ask it (especially with Claude's newer model, Fable 5) to draft an architecture design diagram, it would do a better job than any junior dev ever could. Sure, giving context is still necessary, but you can just speak in plain, normal English and it will understand. AI is getting scary good at system design.”
maybiiLen · Hacker News · Jul 15, 2026 - 3
“Ranks #2 of 28 on LMArena's agentic arena (score 10), measuring tool-use and multi-step task performance from human preference.”
LMArena agentic arena · Benchmark · Jul 20, 2026 - 4
“Here are the numbers from their bar chart: 1. SWE-bench Pro Model Score (%) GLM-5.2 62.1 GLM-5.1 58.4 Claude Opus 4.8 69.2 GPT-5.5 58.6 Gemini 3.1 Pro 54.2 2. Terminal-Bench 2.1 Model Score (%) GLM-5.2 81.0 GLM-5.1 63.5 Claude Opus 4.8 85.0 GPT-5.5 84.0 Gemini 3.1 Pro 74.0 3. NL2Repo Model Score (%) GLM-5.2 48.9 GLM-5.1 42.7 Claude Opus 4.8 69.7 GPT-5.5 50.7 Gemini 3.1 Pro 33.4 4. DeepSWE Model Score (%) GLM-5.2 46.2 GLM-5.1 18.0 Claude Opus 4.8 58.0 GPT-5.5 70.0 Gemini 3.1 Pro 10.0 5. ProgramB…”
mlmonkey · Hacker News · Jun 25, 2026 - 5
“I've been using essentially this process (with Claude Code) for about six months. There are a couple of places where I've opened xcode; mostly to update the simulators for new betas of xcode 27, and once to add a target for Apple Watch (and I think something for HealthKit). Interestingly, since about Opus 4.6, Claude has been able to reason its way into this process on its own. It was clunky until 4.7, and in 4.8 it's managed to find its way around every reason I had to open xcode myself.”
Schiendelman · Hacker News · Jul 13, 2026 - 6
“This is interesting, because if I start a fresh session of Claude Code right now and run context, I see the following: Opus 4.8 (1M context) claude-opus-4-8[1m] 23k 1m tokens (2%) Estimated usage by category System prompt: 3.9k tokens (0.4%) System tools: 13.9k tokens (1.4%) Custom agents: 235 tokens (0.0%) Memory files: 28 tokens (0.0%) Skills: 4.9k tokens (0.5%) Messages: 8 tokens (0.0%) Compact buffer: 3k tokens (0.3%) Free space: 974k (97.4%) 4k tokens is 15-20kB. I'd ask you to paste that…”
mh- · Hacker News · Jul 12, 2026 - 7
“I've been running Opus 4.8 for agentic coding and I don't see it being significantly better than Sonnet 4.5 (not that I can tell). I find that pairing Google Gemini and Claude (having Gemini review Claude's code) seems to yield better results. Curious if this jump to 80.3% score in agentic coding will make me see a big difference in actual usage.”
yandie · Hacker News · Jun 9, 2026 - 8
“Ranks #6 of 28 on LMArena's agentic arena (score 7.7), measuring tool-use and multi-step task performance from human preference.”
LMArena agentic arena · Benchmark · Jul 20, 2026 - 9
“We did the same thing. It took me a few iterations to get the library right, but overall I used approx 10 billion tokens here across a codex and Claude max plan The first iteration was a rewrite using codex goal but it ended up producing junk The second iteration was fable controlling gpt-5.5 agents to write the code with really strong acceptance tests. This pushed the library in the right direction and is basically what you see here today”
suchintan · Hacker News · Jul 16, 2026 - 10
“Apart from the 'Approve for me' in Codex where it has massively regressed. With GPT 5.5 it never got in the way. Now it's infuriatingly deciding to reject the most basic actions used hundreds of times before. It just gave me this gem: > The push to GitLab was blocked because the repository's privacy status couldn't be confirmed. Since the code is private, do you explicitly authorize pushing it to the configured origin on gitlab.com, so the merge request can be opened? This is not a new project,…”
user43928 · Hacker News · Jul 11, 2026 - 11
“Ranks #5 of 28 on LMArena's agentic arena (score 8.3), measuring tool-use and multi-step task performance from human preference.”
LMArena agentic arena · Benchmark · Jul 20, 2026 - 12
“I use Claude Opus 4.7 on max thinking inside Claude Code and I gotta tell you, as context of the project grows, it starts slipping. No amount of whipping and cursing has helped. Currently looking to start making my own hooks setup so it can be safer but nothing concrete yet.”
pdimitar · Hacker News · May 27, 2026 - 13
“Ranks #4 of 28 on LMArena's agentic arena (score 9.1), measuring tool-use and multi-step task performance from human preference.”
LMArena agentic arena · Benchmark · Jul 20, 2026 - 14
“Looking at some of the agentic coding benchmarks on the system card[0], pages 117-118, it seems that running it at low outperforms Sonnet 4.6 at any level, and is a good deal cheaper as well. So on low it could be a good workhorse for an Opus-planned task. [0] https: www.anthropic.com claude-sonnet-5-system-card”
phainopepla2 · Hacker News · Jun 30, 2026 - 15
“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 - 16
“Claude Sonnet 5 is built to be the most agentic Sonnet model yet. It can make plans, use tools like browsers and terminals, and run autonomously at a level that, just a few months ago, required larger and more expensive models. I have been using Sonnet 4.6 more than Opus, because I'm mostly doing agent-assisted development and not fully agent-driven development. This announcement does not make me positive, I have found that the more models are optimized for fully agentic development, the worse…”
microtonal · Hacker News · Jun 30, 2026 - 17
“Grok Build with Grok 4.5 is the best coding AI agent I have ever had the pleasure of using. Stopped using Fable after it.”
hsn915 · Hacker News · Jul 16, 2026 - 18
“Interesting. I experimented with Grok 4 for openclaw when they made clear they wanted to bring claw users in the fold. It was (as expected) more verbally fluid than 5.5, but had real trouble with agentic tool calling - the model felt like it hadn't been trained to think of tool calling as one of its primary modalities. I'll give this a try, the speed and the benchmarks look good. In my experience, Grok slightly punches above its weight in language fluidity, and seems to not benchmaxx on coding,…”
vessenes · Hacker News · Jul 8, 2026 - 19
“Ranks #8 of 28 on LMArena's agentic arena (score 6.4), measuring tool-use and multi-step task performance from human preference.”
LMArena agentic arena · Benchmark · Jul 20, 2026
Rankings synthesized from community evidence and open benchmarks. See methodology. Not driven by vendor marketing.