The question has a short answer — updated July 30, 2026, after a month that reshuffled half the board — as long as you say what you're optimizing for:
| If you optimize for | Pick | Headline benchmark | Price (USD per 1M in/out) |
|---|---|---|---|
| Best, full stop | Claude Fable 5 | 80.3% SWE-bench Pro | $10 / $50 |
| Frontier capability, sane bill | Claude Opus 5 | #1 AA Index (61) | $5 / $25 |
| Agentic & front-end ceiling | Kimi K3 | #1 Arena full-stack (1664) | $3 / $15 |
| Best per dollar (most teams) | GLM-5.2 | 62.1 SWE-bench Pro | $1.40 / $4.40 |
| Hard algorithmic problems | DeepSeek V4-Pro | 93.5 LiveCodeBench | $0.435 / $0.87 |
| Bulk & boilerplate | DeepSeek V4-Flash | good enough | $0.14 / $0.28 |
If you stopped reading here you'd make a fine choice. The rest explains when each pick wins, because "best for coding" hides at least three different jobs.
Job 1: resolving real issues in a real codebase
The benchmark that maps to this is SWE-bench Pro — multi-file GitHub issues, not toy functions. The standings: Claude Fable 5 at 80.3%, then a long gap, then GPT-5.6 Sol at 64.6, GLM-5.2 at 62.1, GLM-5.1 at 58.4, Kimi K2.6 around 58.6, DeepSeek V4-Pro at 55.4. July's two arrivals reshaped the top without touching the value verdict: Claude Opus 5 now tops the overall Artificial Analysis index at half of Fable's price — the new default frontier seat — and Kimi K3 took #1 on Arena's human-preference full-stack board (1664), making it the agentic and front-end ceiling among open models.
Two readings of that list. First, Fable 5 is genuinely a tier of its own — for a multi-day agent run or a risky migration across dozens of files, the extra capability pays for itself, and we've written about when the ceiling is worth it. Second, GLM-5.2 is the best non-frontier option by a clear margin, at about a tenth of Fable 5's output price. Its million-token context also means the whole service fits in one prompt — see the GLM-5.2 deep dive for how to use that well.
Job 2: hard, self-contained algorithmic problems
Different benchmark, different winner. On LiveCodeBench — competitive-programming-style problems — DeepSeek V4-Pro scores 93.5, the top score of any model, closed or open, with a 3206 Codeforces rating to match. If your workload looks like "implement this tricky algorithm correctly," V4-Pro at $0.435 / $0.87 per million tokens is the pick, and it isn't close once price enters the math.
Job 3: volume — completions, boilerplate, refactor grunt work
Most code tokens in a real org are not hard. Test scaffolding, CRUD endpoints, format conversions, comment generation: DeepSeek V4-Flash handles these at $0.14 / $0.28, roughly 30x cheaper than GLM-5.2 on output. The pattern that works is a two-line router: Flash by default, escalate to GLM-5.2 or V4-Pro when the prompt is genuinely hard. We show the code in the smart routing walkthrough.
The honest caveats
- Tool-heavy agent stacks: Kimi K2.6 posts the strongest tool-use and MCP scores of the Chinese models. If your bottleneck is orchestration rather than raw code quality, test it against GLM-5.2 on your own traces.
- Benchmarks saturate. A 62 vs 55 gap on SWE-bench Pro is meaningful; a 1-point gap is not. Run your own evals on your own repo before committing spend.
- This changes fast. GLM-5.2 went from unreleased to #1 open-weight in one week of June. Build so that the model is a config string, not an architecture decision.
How to actually run this from outside China
Fable 5 you can buy anywhere. The Chinese models are harder to reach directly — mainland sign-up, local payment rails. Turiloop is an OpenAI-compatible API gateway that puts GLM-5.2, DeepSeek V4-Pro/Flash, Kimi K2.6 and MiniMax behind one key, billed pay-as-you-go with an international card, no Chinese phone number. Point your OpenAI SDK at https://api.turiloop.com/v1 and the router above is all the integration you need.
FAQ
Is GLM-5.2 better than GPT-5.6 for coding? GPT-5.6 Sol edges it on SWE-bench Pro (64.6 vs 62.1) — a 2.5-point gap at 3.6× the input price and 6.8× the output price. Per dollar, GLM-5.2 wins clearly; for the hardest agentic runs, Sol or Opus 5 lead. Full tier-by-tier comparison.
What's the best free or cheapest option for coding? Self-hosting GLM-5.2's MIT-licensed weights is free if you own the GPUs. Via API, DeepSeek V4-Flash at $0.14/$0.28 per million tokens is the cheapest option that still writes usable code.
Should I just use Claude Fable 5 for everything? Only if the token bill is irrelevant next to engineer time. It's the strongest coding model, but at $50 per million output tokens, routing everyday work to GLM-5.2 or V4-Flash typically cuts the bill by 10-50x with little quality loss on routine tasks.
Which LLM is best for a coding agent? For long multi-file runs on a budget: GLM-5.2 (1M context, 62.1 SWE-bench Pro). For maximum reliability regardless of cost: Claude Fable 5. For tool-call-heavy orchestration: benchmark Kimi K2.6 against GLM-5.2 on your stack.
*Benchmarks and prices updated July 30, 2026 — sources: SWE-bench Pro, LiveCodeBench, Artificial Analysis, Arena, official provider price pages.*