Coding Agent Candidate Scorecard: Grok 4.6 / Qwen3.8-Max / DeepSeek-V4 Pro / GLM-5.2

TL;DR: This is the targeted follow-up to “How to pick your daily Coding Agent model — stop worshipping SWE-bench #1” — we put four candidates (grok-4.6, qwen3.8-max, deepseek-v4-pro, glm-5.2) on official leaderboards and pinned down each score. None of the four beats the public frontier (claude-opus-5 74% / gpt-5.6-sol 73% / fable-5 70%); but on “cheap + good enough” there is a clear route: GLM-5.2 for 80% daily work, Grok 4.6 for hard tasks, DeepSeek-V4 Pro as the cheap sub-agent. The key red flag: GLM-5.2 ranks #4 at fixing fresh issues but sits near the bottom on original long-horizon engineering — and your daily work looks more like the latter, so don’t let the former fool you.

Why a separate post

The previous post answered “which public benchmarks should I track” (Terminal-Bench / SWE-rebench / DeepSWE + a cost dashboard). This one answers a more concrete question: of the four models I actually have on hand, who should be the primary? So the method is different — not skimming leaderboards in general, but going to each official page for these four specifically, using what’s there, and honestly writing “not on the board” when it isn’t. No fabrication.

Method: 4 parallel research agents via Grok web search → 4 verify agents re-checking each claim against the official source → one final manual pass against the official boards (swe-rebench.com / deepswe.datacurve.ai / tbench.ai). Snapshot: 2026-08-13.

The scorecard (after official-source verification)

First, an unexpected finding: DeepSWE is the only benchmark where all four candidates have official data today. Terminal-Bench 2.1 barely covers any of them (none on the official board), and SWE-rebench only covers two. So to compare these four horizontally, DeepSWE has to be the main axis, with rebench as a supplement.

Model	DeepSWE v1.1 (updated 8/13 · 113 tasks)	SWE-rebench (15/05–01/07 · 111 tasks)	Terminal-Bench 2.1
Grok 4.6	67%±2% [xhigh] · $5.50 · 71k tok · 87 steps ✅	4.6 not on board; 4.5=63.8%/$1.47	4.6 not on official board
DeepSeek-V4 Pro	63%±6% [max] · $0.06 · 106k · 155 steps ✅	40.2%/$0.15 (rank 14)	67.9% is vendor-reported, refuted ❌
Qwen3.8-Max	57%±3% [xhigh] · $3.73 · 95k · 111 steps ✅	no row	not on official board
GLM-5.2	44%±2% [max] · $3.92 · 78k · 129 steps ✅	62.9%/$1.40 (rank 4, Pass@5 81.1% — best on the board)	81%? third-party only, unconfirmed ⚠️

(✅ directly verified on the official board; ⚠️ third-party / unconfirmed; ❌ refuted by verification — vendor-reported number doesn’t match the official board.)

For context, the public frontier on the same DeepSWE board: claude-opus-5 [max] 74%/$11.84, gpt-5.6-sol [max] 73%/$8.39, claude-fable-5 [max] 70%/$21.63, kimi-k3 [max] 69%/$4.65. The best of the four, Grok 4.6 (67%), only ranks mid-upper — still a notch below the frontier.

The key red flag: GLM-5.2’s severe split

If you only look at SWE-rebench, GLM-5.2 is the prettiest of the four — rank 4 (62.9%), and its Pass@5 is the best on the entire board (81.1%), meaning given 5 retries it fixes more problems than anyone; plus $1.40/problem and already wired into the CPA gateway, it looks born to be the primary.

But switch to DeepSWE and it drops to 44%±2%, near the bottom of 23 models.

Why the gap? The two benchmarks measure different things:

SWE-rebench = you fix “fresh, real GitHub issues.” Models often have memory/familiarity with common issue patterns in public repos, so they fix them handily.
DeepSWE = “original long-horizon engineering tasks,” written from scratch, solutions need 5.5× the code and 2× the output tokens, and it’s contamination-free — tasks are constructed from scratch so the answer was never seen in pretraining. This measures real long-horizon engineering ability.

Which one is your daily work? Opening Claude Code on a real repo, editing across files, running tests, reading errors, re-editing, touching Docker/Git/WSL/deploy — that’s long-session, original-edit work, closer in shape to DeepSWE than to fixing a ready-made issue on SWE-rebench.

So the high rebench score is misleading for your daily work. 44% is the warning line to watch: GLM-5.2 is good at “fixing fresh issues” but bad at “building long-horizon engineering from scratch.”

That doesn’t mean it can’t be the primary — it’s cheap, integrated, and Pass@5 #1 means it’s strong in retry-heavy scenarios (debugging with repeated retries). But don’t expect it to carry the hardest long-horizon original tasks — route those to Grok 4.6 (67%) or straight to the frontier.

Cost/success: the KPI you should actually optimize

The previous report’s core claim: don’t look at raw accuracy, look at “how much does one successfully completed real task cost.” A rough DeepSWE calc ($/task ÷ Pass@1):

DeepSeek-V4 Pro: ~$0.10/success — absurdly cheap (but low success rate; only pays off at scale)
Qwen3.8-Max: ~$6.5/success
Grok 4.6: ~$7.5/success
GLM-5.2: DeepSWE cost not extracted; rebench standard ~$2.2/success (note: issue-fixing standard, not long-horizon)

On “cost per successful long-horizon task,” DeepSeek-V4 Pro is anomalously cheap but with a low success rate; Grok 4.6 / Qwen3.8-Max have a higher success rate but cost ~60–75× more. That’s exactly why you should route rather than dogmaticly stick to one model.

Routing recommendation

1
2
3
4
Primary (80% daily / retry-heavy issue):   GLM-5.2        # cheap + Pass@5 best, but low long-horizon ceiling
Hard / long-horizon original engineering:  Grok 4.6 or claude-opus-5/gpt-5.6 frontier  # DeepSWE 67%+
Worker/Cheap (bulk / sub-agent):           DeepSeek-V4 Pro   # $0.06/task, absurdly cheap (verify tool format first)
Challenger (pending private audit):       Qwen3.8-Max       # mid-pack on public data, not enough to decide

A few things that must be said clearly:

Verify tool format before using DeepSeek-V4 Pro as a sub-agent. The flash tier was tested before: 4.8× slower, 5.8× more verbose, unfit for “fast and short” sub-agents; v4-pro is the reasoning tier, needs a separate test — don’t just reuse that conclusion.
Grok 4.6 has only one data point (DeepSWE), no rebench/TB cross-check. And xAI/grok has a history of quota and stability pain in a self-hosted proxy stack — go in with eyes open.
I could not confirm GLM-5.2’s “Terminal-Bench 81%” on the official board — it’s a third-party site (layer3labs) figure, so don’t treat it as a TB championship.
DeepSeek’s “TB 67.9%, rank #1” was a sub-agent hallucination. The research phase reported it; the verify phase refuted it: it’s not on the official 2.1 board, and 67.9% is vendor-reported. This is exactly why the verification phase can’t be skipped.

Honest limits

All of the above are public-harness scores. Princeton HAL proved long ago: the same model under a different scaffold (tool config, retry policy, context management, patch strategy) can swing by tens of percentage points. So —

Public boards are only a shortlist. The final verdict must come back to your own Personal Suite: a fixed Claude Code (or your main harness) + your own repo tasks, logging each model’s “successes / $ / tokens / steps / derailments / regressions.”

GLM-5.2’s rebench↔DeepSWE split is exactly the illustration of “you can’t pick a primary from a single benchmark” — especially when one benchmark happens to flatter you and another happens to expose the weakness.

Every 1–2 months, re-run GPT / Claude / Gemini / DeepSeek / Kimi / Qwen / GLM / MiniMax through a fixed suite and look at “how much money, time, and human hand-holding each successful real task costs.” That’s the real scale for picking a primary.