Agent × LLM: The Leading 2026
"Framework + Model" Combos — Evaluation & Selection Guide

Three findings that matter

• ① The combo — not the model — is the unit of selection. Swap the scaffolding around one model and its Terminal-Bench score can move 5–30 points, more than a full model generation delivers. All 12 leaders on the official Terminal-Bench 2.0 board are submitted as "harness × model," and the top entry is a dedicated harness (NexAU-AHE × GPT-5.5, 84.7%), not any bare model.
• ② The Anthropic protocol has become the de-facto standard, and "Claude Code with a Chinese Mainland brain" is this cycle's biggest structural shift. GLM, Kimi and Qwen all expose Anthropic-compatible endpoints; two environment variables let Claude Code talk directly to GLM-5.3. Under our weighting, Claude Code + GLM-5.3 takes #2 at 84.3 at roughly a quarter of the cost of the official Opus 5 combo.
• ③ Open weights have reached the frontier band, but the "peak" is still held by closed models. Kimi K3 became the first open model to top an LMArena coding board (Frontend Code, 1679 Elo), and its AA Intelligence Index of 59.70 sits within 4 points of the closed first tier; yet on hard benchmarks such as SWE-bench Pro and GAIA, Claude / GPT-5.6 still hold a 5–15 point lead.

The one-line selection takeaway

For the quality ceiling choose Claude Code + Opus 5 / Fable 5; for value choose Claude Code + GLM-5.3 or Codex + DeepSeek V4 Pro; for high throughput on a tight budget choose Codex + GPT-5.6 Luna or Gemini 3.7 Flash; for self-hosting and data sovereignty choose OpenHands + GLM-5.2 (MIT); in regulated Chinese Mainland industries use the official compliant channels of Qwen / DeepSeek / GLM / Kimi. There is no "best" combo, only the combo that matches your budget, compliance constraints and task shape — Section 08 holds the full decision matrix.

From "strongest model" to "strongest combo"

Model evaluation in 2024–2025 rested on an implicit assumption: capability lives in the model, so buying the best model is enough. The 2026 data overturns it. On Princeton's HAL board, the real reason Anthropic swept the top six is the HAL Generalist Agent scaffolding — on the same GAIA problem a bare model sits around 44.8%, the HAL harness reaches 74.6%, and a freely engineered system tops 92%, a spread of more than 30 points across the three protocols. On the official Terminal-Bench 2.0 board, the neutral reference harness Terminus 2 × Claude Opus 4.7 manages only 69.4%, while the same-generation models on dedicated harnesses generally clear 80%. The LangChain team once lifted a Top-30 submission into the Top 5 simply by adjusting prompts and tool configuration.

At the same time, the model layer is rapidly becoming interchangeable. Once the Anthropic API protocol became the de-facto standard, GLM, Kimi and Qwen all shipped compatible endpoints, and officially bound CLIs such as Claude Code effectively became "universal framework shells." Evaluation sites completed the pivot ahead of procurement: the entries on Terminal-Bench, SWE-bench, SWE-Marathon and OSWorld are all "agent harness × model" combos.

The 2026.09 landscape at a glance

• Closed frontier: Anthropic holds the AA Intelligence Index top 7 with Fable 5 / Opus 5; the OpenAI GPT-5.6 family (Sol / Terra / Luna, three price tiers) went fully GA in July; Google's Gemini 3.1 Pro defends long context and scientific reasoning, while 3.7 Flash sets the value bar at $0.75/$3.75; xAI's Grok 4.6 enters the first tier at a low $2/$6 but lacks a mature official agent framework.
• Open weights: Kimi K3 (topped an LMArena coding board), GLM-5.3 (AA 59.51, weights not yet released), DeepSeek V4 (96.4% on the independent SWE-bench Verified test, best of the open set) and Qwen 3.8 (86.1% on OSWorld-Verified, #1 for computer use) form the four leading Chinese Mainland labs.
• Traffic structure: the OpenRouter August weekly token-volume top 10 is dominated by low-cost MoE models (DeepSeek V4 Flash 12.3T, GLM 5.3 Flash 6.2T) — agent traffic "runs the cheap ones and praises the expensive ones," while the reputation board (Stack Overflow 2026 most-loved tool, 46%) remains a Claude Code monopoly.

Questions this report answers

• How do the world's mainstream "Agent framework × model" combos rank on overall capability? (Section 06)
• How does the ranking shift under different weight preferences (quality-first vs cost-first)? (Sensitivity analysis, Section 06)
• What should different audiences, company sizes, industries and professional backgrounds choose? (Section 08)
• Which data-protocol risks should be hedged in advance? (Section 09)

Source list and tiers

Normalization and weighting

• Normalization: raw benchmark scores are linearly mapped to 0–100 with "current best within the same protocol = 100." Where a benchmark has both a vendor self-report and an independent retest, the independent protocol is preferred (Vals AI, AA, the official uniform-harness board); the self-report is used only as a labeled range reference.
• Two-level weights: at the dimension layer — general intelligence 25% / agentic capability 25% / coding 20% / cost-value 15% / ecosystem & engineering 15% (a balanced scheme aligned to the report's agent focus but letting no single dimension dominate). At the source layer — independent third party 70%, official verified board 20%, vendor self-report 10%.
• Combo protocol: the model-side dimensions of a combo (intelligence, coding, cost) use model data; agentic and ecosystem dimensions are scored on the "model × framework" interaction (the same model scores differently on Claude Code versus its native CLI).
• Cost protocol: assessed on a "per-task" basis — 40K input + 8K output tokens (a typical agent task: repo context + plan + diff), accounting for cache discounts and subscription plans rather than list price alone.

Six families, 15 combos

Five-dimension framework

5.1 General intelligence: Anthropic holds the top seven; the four Chinese Mainland labs enter the frontier band

• The AA Index top 7 are all Anthropic models, and five models are packed within 4.7 points at the top — general intelligence has entered a neck-and-neck "no significant difference" zone, so choices within the closed first tier should be driven by coding/agentic sub-scores and price.
• Same protocol on the LMArena text board: claude-opus-5-max leads at 1569 Elo, while qwen3.8-max (1535), kimi-k3-max (1529) and glm-5.3-max (1528) have entered the 1520+ frontier band.
• A notable divergence: DeepSeek V4 Pro's composite of 53 trails Kimi K3 by ~7 points, yet it posts 96.4% on the independent SWE-V test — the gap between specialist and composite is the norm for Chinese Mainland models, which makes choosing by task more important than choosing by total score.

5.2 Coding: the harness protocol is worth 15–20 points, and independent retesting is the only hard currency

• SWE-bench Pro is the real watershed: Verified is near saturation (the top 5 sit within 4 points), while on Pro, Claude Mythos 5 / Fable 5 / Opus 5 lead at 79–80%, GPT-5.6 Sol is at 64.6% and Qwen 3.8 Max at 67.7%; under Scale's standardized protocol the top score drops sharply to 59.1% (GPT-5.4). Topping the public set ≠ performing on a private codebase.
• OpenAI stopped submitting Verified in 2026.02 citing test-set contamination; a Scale AI audit in 2026.07 found ~30% of Pro's public tasks broken. This report down-weights any vendor self-report without an independent protocol.
• On the LMArena WebDev board (human blind test): Opus 5 Max 1688 > Kimi K3 Max 1674 > Qwen 3.8 Max 1669 — Chinese Mainland models hold two of the coding blind-test top three; Kimi K3 became the first open model to top an LMArena coding board in July at 1679.
• The Aider Polyglot official board stopped updating after mid-2025 with all frontier models absent, so it no longer discriminates and is not used here.

5.3 Agentic capability: the combo protocol's home turf — and the dimension with the widest disagreement

This is where the "framework × model" interaction is strongest, and it shares the top weight (25%) with general intelligence. Three sets of key evidence:

• The official combo board proves the harness is the first variable: the Terminal-Bench 2.0 top four are NexAU-AHE × GPT-5.5 (84.7%), LemonHarness × multi-model (84.5%), Capy × GPT-5.5 (83.1%), Codex CLI × GPT-5.5 (82.2%); Claude Opus 4.8 has three public numbers — Anthropic's chart 74.6%, OpenAI's chart 78.9%, Vals independent 71.9% — one model, three stories.
• Long horizon is the real bottleneck: SWE-Marathon (avg 27.2M tokens/task) has a ceiling of just 42.0% (Kimi K3, verified), and the combo entry Claude Code + Opus 4.8 reaches only 26%; on τ³-Bench's new banking_knowledge domain the best model manages 26.7%; on BFCL V4 multi-turn function calling the best is 68.4% — multi-turn, long-horizon, knowledge-intensive tasks are far from solved.
• Computer use has been claimed by Chinese Mainland labs: on OSWorld-Verified, Qwen 3.8 Max overtakes Claude Mythos 5 (85.0%) at 86.1% to take #1 (2026.08.15), already past the human baseline of 72.36%. Scenarios needing desktop automation should re-evaluate Chinese Mainland options.
• Tool-call discipline: BFCL V4 shows every frontier model degrades markedly on multi-turn scenarios that require "deciding not to call a tool" (o3 only 14.8% multi-turn); the τ²-Bench telecom domain is near-saturated at 99%, but under the pass^8 reliability protocol the leaders fall below 25% — ask vendors for pass^k data in procurement.

5.4 Cost & value: a 45× spread in per-task cost; subscription anchors at $20/$100/$200

• A 45× range in per-task cost (Fable 5 $0.80 vs Luna $0.018), while the independent-protocol capability gap (SWE-V 95 vs 93) is just 2 points — the value story of cheap models is inherently favored in weighted scoring, which is exactly why the Section 06 sensitivity analysis exists.
• Cache economics sets the real agent bill: multi-turn agents replay the same system prompt repeatedly; cache-hit pricing is ¥0.025/M for DeepSeek, $0.30/M for Kimi and $0.26/M for GLM. The nominal 2.5× gap between K3 and GLM-5.3 shrinks to ~1.15× under cache-dominated traffic. Evaluate on cost per completed task, not unit price.
• The subscription market has converged: Claude / Codex / Devin / Factory / Cursor Max all anchor to three tiers at $20/$100/$200; Chinese Mainland vendors attack as compatible alternatives at a 5–10× price gap — Kimi starts at ¥39/month domestically, the GLM Coding Plan is ¥118/month on Tmall (launched 2026.09, adapting to 20+ agents incl. Claude Code / Codex / ZCode).
• Hedge promo-price risk in advance: GPT-5.6 Sol's promo runs to 2026.11.21 and Gemini 3.7/3.6 Flash intro pricing to 2026.12.31; budgets jump 50–100% afterward. Stress-test long contracts at list price.

5.5 Ecosystem & engineering: protocol decoupling and "choose the model by cloud"

• The Anthropic protocol is the de-facto standard: GLM, Kimi and Qwen all expose Anthropic-compatible endpoints (via Z.ai /api/anthropic, two environment variables connect Claude Code straight to GLM-5.3); framework-model decoupling is this cycle's biggest structural change.
• Claude Code is the double champion of reputation and protocol: 46% in Stack Overflow 2026 "most-loved tool" (Cursor 19%, Copilot 9%); tied for first with Cursor (18%) in the JetBrains workplace-adoption survey; #4 on OpenRouter at 289B tokens/day; ~75% penetration in startups (<50 people).
• Codex CLI is the most open first-line framework: Apache-2.0 (Rust, 62K+ stars), with community consensus that it is strong on DevOps/CI and 2–4× more token-efficient than Claude Code.
• Gemini CLI went through turbulence: 106K stars (Apache-2.0) but on 2026.06.18 it closed personal free/Pro/Ultra hosted channels, steering individuals to the closed successor Antigravity CLI while enterprise Code Assist licensing remains — its ecosystem score is marked down substantially.
• OpenHands is the default neutral multi-model answer: MIT-licensed, 84.5K+ stars, 490+ contributors, connects directly to OpenRouter/Ollama, and now brings Claude Code / Codex / ZCode under a unified governance console — the pragmatic choice for framework governance in large teams.
• Enterprises choose the model by cloud: Claude on Bedrock serves 100k+ enterprise customers (AWS added a ten-year, hundred-billion-dollar-class Anthropic commitment), Vertex pulls new users with Gemini long context + BigQuery, Azure hosts GPT — 87% of large enterprises use multiple clouds, and compliance and existing cloud footprint weigh on selection no less than the model itself.
• Traffic and reputation are inverted: the OpenRouter August weekly token top 10 is led by cheap MoE (DeepSeek V4 Flash 12.3T, GLM 5.3 Flash 6.2T; vendor share DeepSeek 27.1%, Google 25.3%, OpenAI 16.9%) — token volume reflects "how many tasks are running," not "quality preference."

Reading the ranking

• The champion does not win by a rout: Claude Code + Opus 5 (86.0) leads #2 by 1.7 points — under the "<3 points = no significant difference" protocol, the top four should be read as one tier, and the real difference is in dimension structure: the Opus 5 combo wins on quality and ecosystem, the GLM-5.3 combo on cost and protocol compatibility.
• Same model, different framework — the gap is clearly visible: Kimi K3 on Claude Code (81.4) is 2.9 points above Kimi K3 on Kimi CLI (78.5); the entire gap comes from harness maturity and ecosystem — quantitative proof that "the combo is the unit of selection."
• Gemini 3.1 Pro combo ranking last is not a model problem: the model itself is first-tier (AA 57.2 / SWE-V 80.6); the drag is the ecosystem turbulence after Gemini CLI closed personal channels (score 70) plus a relatively high per-task cost.

Sensitivity analysis: how the ranking moves when weights change

A single total is easily manipulated by weight design — the honest approach is to show the results under several schemes. Below, the top 8 combos are tested under three weighting schemes:

• Quality-first: the order becomes Opus 5 > Fable 5 > GPT-5.6 Sol > GLM-5.3 > Kimi K3 — closed flagships take the top two and "you get what you pay for" returns.
• Cost-first: DeepSeek V4 Pro (82.3), Luna (81.4) and GLM-5.3 (80.1) take the top three — exactly why cheap MoE dominates the OpenRouter traffic board.
• Robust conclusion: Claude Code + GLM-5.3 is the only combo that stays in the top four under all three schemes — if your demand profile is unclear, it is the lowest-risk default (while hedging the supply risks in Section 09).

8.1 By audience and company size

8.2 By industry (compliance as the first variable)

8.3 By application (decision matrix)

Three core calls

• Call one: the combo is the unit of selection, and protocol decoupling is a structural trend. Once the Anthropic protocol became the de-facto standard, the full Cartesian product of "framework × model" became composable for the first time — both freedom of choice and governance complexity. Every 2026 procurement decision should include a "brain-swap cost" assessment: keep interchangeability at the framework layer (OpenHands / compatible endpoints) and route models by task.
• Call two: closed models defend the quality peak, but open weights have reached the frontier band. The AA top 7 remain Anthropic and the SWE-Pro top three remain the Claude line; but Kimi K3 (topped a coding blind-test board), Qwen 3.8 (#1 on OSWorld) and DeepSeek V4 Pro (best open score of 96.4% on the independent SWE-V) mean the old assumption "open isn't production-ready" no longer holds. Open weights' real shortfall is in independent agentic protocols (DeepSeek 54.7) and ecosystem maturity, not coding.
• Call three: the cost band is being reshaped, with ~$0.02/task the new floor. Once Luna and DeepSeek V4 Flash push per-task cost to the $0.02 level, the economic boundary between "an expensive model gets it right once" and "a cheap model retries several times" must be recomputed per task — which is why this report sets the cost weight at 15% and no higher: cheapness alone is no advantage; cheap plus independent-protocol reliability is.

Signals to track next cycle (2026 Q4)

• Whether GLM-5.3 releases weights — deciding whether the "self-host route" gains a new flagship (currently only GLM-5.2 is downloadable).
• OpenAI pricing after the GPT-5.6 Sol promo expires (2026.11.21) — if $4/$20 becomes permanent, it rewrites the cost dimension.
• Progress on Terminal-Bench 3.0 and the τ³ banking_knowledge domain — where the ceiling on long-horizon, knowledge-intensive tasks sits.
• Convergence/divergence between Chinese Mainland vendors' AA Index share and their OpenRouter traffic share — whether "run the cheap, praise the expensive" gets broken.
• Whether Gemini 3.5 Pro ships and whether Antigravity CLI repairs the Gemini ecosystem score.