Same Repository, Same Ticket: Sol vs Grok vs GLM vs Claude🔗

GPT-5.6 Sol, Grok 4.5, GLM-5.2 and Claude Opus 4.8 are all sold as coding and agentic models for serious professional work. Our original plan was to plant defects in a small Next.js/Supabase repo and run all four through the same ticket three times each. That trial is still worth doing — but before spending the budget on it, it is worth collating what the published record already shows, with sources, and answering the question that actually motivated this: does self-hosting an open-weight model like GLM-5.2 beat paying for a hosted API on agentic coding work?

🖥️ Hint: Ask a naive question. Everyone quotes SWE-bench percentages. Almost nobody quotes the electricity bill.

The baseline question: is hosting GLM-5.2 yourself worth it?🔗

GLM-5.2 is Z.ai's MoE reasoning model — 753B total parameters, 40B active per token, a 1M-token context window, released 16 June 2026 and shipped MIT-licensed with public weights (Flowtivity, Emergent). Being open weight means the model itself is free. Running it is not.

Hosted, it is cheap. The direct Z.ai API prices GLM-5.2 at $1.40 per million input tokens and$4.40 per million output tokens ($0.26 for cached input), and OpenRouter resellers bring that to roughly$0.95 / $3.00 ([DeepInfra](https://deepinfra.com/blog/glm-5-2-pricing-benchmarks-cost-comparison), [OpenRouter](https://openrouter.ai/z-ai/glm-5.2)). The GLM Coding subscription tiers run from about$3–6/month (Lite) through $15–72/month (Pro) to$30–160/month (Max) (Lushbinary).

Self-hosted, it is a different order of spend. Running the full FP8 checkpoint (~750GB plus KV cache) in production realistically needs 8× NVIDIA H200 GPUs; a 4-bit GGUF cut can run on 4× H100; a 2-bit quantisation fits a single Mac Studio M3 Ultra (decodethefuture.org). What that costs, per month, depends entirely on how you acquire the hardware (ofox.ai):

The break-even point sits at roughly 3,000–10,000 prompts a day depending on the ownership model above, and that number only covers hardware — it excludes the team required to run it. Broader self-hosted-LLM cost research puts the practical crossover against premium hosted APIs at somewhere between 5–10 million tokens a month for a rented single GPU, and notes self-hosting typically runs 3–5× the raw GPU sticker price once you add a DevOps engineer (~$145K/year fully loaded in the US), model update cycles every 6–8 weeks, and the idle-GPU cost of a workload that is not flat 24/7 (braincuber.com, gigagpu.com).

The Mac Studio route is the interesting middle ground — near-zero monthly cost for a private, air-gapped setup, if a 2-bit quantisation of GLM-5.2 holds up for your use case. For everything else, the arithmetic is blunt: unless an engineering team is running enough concurrent coding agents to sustain tens of thousands of prompts a day, or has a compliance reason data cannot leave the building, the hosted API or coding-plan subscription is cheaper than owning GPUs, and dramatically cheaper than renting them.

The four contenders, as published🔗

Sources: Finout, pricepertoken.com, apidog, BenchLM, CodingFleet. Context-window figures vary between reseller pages and should be treated as approximate — check the provider's own docs before quoting a number in a proposal.

What the published benchmarks say about agentic coding🔗

Sources: Groundy, apidog, morphllm, dev.to. Two gaps worth naming honestly: no independently published SWE-bench Verified figure for GPT-5.6 Sol specifically was found at time of writing (the 88.7% belongs to GPT-5.5), and no Grok-4.5-specific coding score has been independently confirmed either — the 75% above is the earlier Grok 4 result. Benchmark harnesses and reasoning-effort settings differ between publishers, so treat this table as directional, not a leaderboard.

The number worth sitting with is MCP-Atlas, since it is closest to what an audit-log-and-RBAC ticket actually exercises — reading a repo, calling tools, orchestrating multiple steps. GLM-5.2 (77.0) sits within a point of Claude Opus 4.8 (77.8) there, at roughly a quarter to a sixth of the token cost.

Where each model actually fits🔗

• Best raw coding accuracy: Claude Opus 4.8 — leads SWE-bench Verified and Terminal-Bench 2.1 among the four.
• Best value on agentic/tool-use work: GLM-5.2 — near Opus 4.8 on MCP-Atlas at a fraction of the API price, hosted.
• Longest usable context: GPT-5.6 Sol (~1.05M) and GLM-5.2 (1M), both ahead of Opus 4.8's advertised window.
• Best open-weight / most deployable option: GLM-5.2, with the self-hosting caveat above.
• Cheapest entry tier: GLM-5.2 Coding Lite (~$3–6/month) or the GPT-5.6 Luna tier ($1/$6 per million tokens).

What I would deploy🔗

Route by task, not by favourite: Claude Opus 4.8 for architecture calls, security review, and the defect nobody else can close cleanly. GLM-5.2 — hosted via API or a coding-plan subscription, not self-hosted, until the volume math above actually clears — for the high-frequency, lower-stakes work: routine implementation, test updates, documentation. Reach for Sol or Grok specifically when their context window or tool-orchestration mode fits the job better than the other two. Only put GLM-5.2 on your own GPUs once monthly hosted spend is genuinely into five figures — below that, the DevOps headcount alone costs more than the tokens you'd save.

What this isn't yet🔗

This is a literature and pricing synthesis, not our own controlled trial. Every number above is a snapshot from mid-July 2026, sourced and linked, and this market moves in weeks not quarters — re-check before quoting any of it externally. The original plan still stands as the useful follow-up: plant the same defects in the same small repo, run all four models three times each under identical budgets, and report cost per accepted change first-hand instead of citing someone else's harness.

"Latency is a musical instrument." — Brian Eno

The same holds for a coding agent choosing between a cheap model and a fast one. Cost per accepted change only means something once you have actually shipped the change.

Sources🔗

• GLM 5.2 API & Pricing: GLM Coding Plan Guide — Lushbinary
• GLM-5.2 Pricing, Benchmarks, and Cost Comparison — DeepInfra
• GLM 5.2 — API Pricing & Benchmarks — OpenRouter
• GLM-5.2 GPU Requirements: Self-Host vs API Cost — decodethefuture.org
• Self-Host GLM 5.2: 8×H200 vLLM Cost vs $30/mo Cloud — ofox.ai
• GLM-5.2: The Open-Source AI Model Beating GPT-5.5 at 1/6th the Cost — Flowtivity
• GLM 5.2 Benchmark: Every Score Explained — Emergent
• GLM-5.2 Benchmarks and Specs: SWE-bench Pro — apidog
• GLM-5.2 Benchmarks: What 62.1% SWE-bench Pro and 99.2% AIME Mean — Groundy
• Claude Opus 4.8 Pricing 2026 — Finout
• Claude Opus 4.8 API Pricing — pricepertoken.com
• LLM API Pricing Comparison July 2026 — BenchLM.ai
• Grok 4.5 vs Claude Opus 4.8: Which One Is the Best — apidog
• GPT-5.6 Sol vs Claude Opus 4.8: Full Benchmark Comparison — CodingFleet
• Best LLM for Coding (2026): 12 Models Ranked by SWE-bench and Cost per Task — morphllm
• The Best LLMs for Agentic Coding in 2026 — dev.to
• Self-Hosted LLM vs API: The $4,200/mo Break-Even Point — braincuber.com
• Is Self-Hosting LLMs Cheaper Than APIs in 2026? — gigagpu.com