There is no "best LLM" anymore β there's a best model for coding, a best one for long-horizon agents, a best one for reasoning, and a best one for your budget, and they are not the same model. I spent the last few weeks pulling every current frontier and open-weight model onto the same bench, cross-checking vendor claims against independent numbers, and mapping each to the jobs my team actually runs. Here's the 2026 routing map β from an agentic AI manager who has to answer "which model?" a dozen times a day.
Why I Went Down This Rabbit Hole
Every week someone on my team asks the same question: "Which model should I use for this?" And every week the honest answer gets longer, because the field keeps splitting.
A year ago you picked one frontier model and used it for everything. In mid-2026 that's malpractice. The gap between the best coding model and the best reasoning model and the best value model is now wide enough that picking wrong costs you real money, real latency, or a silently worse agent. As someone who manages agentic systems in production, my job stopped being "pick the smart one" and became routing β matching the shape of a task to the model that wins that shape.
So I did the obvious thing. I put ~20 current models β Anthropic, OpenAI, Google, xAI, Meta, and the surging Chinese open-weight labs β on the same bench, cross-referenced the Artificial Analysis Intelligence Index against vendor pages, and threw out every number I couldn't corroborate. This is the map I gave my team.
TL;DR
- No single winner. Claude Opus 5 tops the overall Artificial Analysis Intelligence Index (~61), but the category crowns are split across five labs.
- Coding: Claude Fable 5 / Opus 5 lead SWE-bench; GPT-5.6 Sol and Grok 4.5 are right behind on the harder agentic coding evals.
- Agentic tool use: Meta's Muse Spark 1.1 leads tool-orchestration (MCP Atlas 88.1); Gemini 3.6 Flash leads computer-use (OSWorld 83%). The metric that matters β policy adherence under ΟΒ²-bench β is where most models still quietly fail.
- Reasoning & science: GPT-5.6 Sol and Gemini 3.1 Pro are co-leaders on GPQA Diamond (~94%); Fable 5 leads Humanity's Last Exam.
- Value: the story of 2026. Chinese open-weight models β GLM-5.2, DeepSeek V4, MiniMax M3 β deliver frontier-adjacent quality at 1/6 to 1/30 the token cost.
- Open weights: Kimi K3 is the strongest open model (Index 57), ahead of GLM-5.2 and DeepSeek V4 β while Llama 4 has fallen to the bottom and Meta's real frontier (Muse Spark) is now closed.
- Read benchmarks like an adult. SWE-bench Verified and AIME are saturated, OpenAI stopped publishing standard evals, and "with tools" vs "no tools" scores get mixed constantly. I flag the traps below.
First, A Benchmark-Literacy Warning (Read This)
Before a single ranking, five things will save you from being fooled by a leaderboard screenshot:
- The Artificial Analysis Index got harder. The current v4.1 is a 9-eval composite (Terminal-Bench 2.1, Humanity's Last Exam, GPQA Diamond, and more), recalibrated tougher than the 2025 index. A model that launched bragging "56" on the old index may show "46" on today's board. Compare like with like.
-
Scores are effort-dependent. The same model scores differently at
highvsmaxreasoning effort. GPT-5.6 Sol is ~59 at max but ~56 at high. Always pair a number with its setting. - The classic benchmarks are saturated. SWE-bench Verified and AIME 2025 are largely maxed out. The live differentiators in 2026 are SWE-bench Pro, Terminal-Bench 2.1, FrontierSWE, HLE, and AIME 2026. If a table still leads with AIME 2025, it's dated.
- OpenAI stopped publishing. OpenAI did not release official SWE-bench Verified / GPQA / HLE numbers for GPT-5.6 β a real break from the past. The GPT-5.6 figures here are Artificial Analysis's independent runs, not OpenAI's, and I mark them as such.
- "With tools" β "no tools." Humanity's Last Exam scores nearly double when a model is allowed tools. Vendors love to quote the with-tools number next to a rival's no-tools number. Don't let them.
With that armor on, here's the field.
The 2026 Model Landscape, As One Ladder
Twenty models, one table. Prices are per 1M tokens (input / output); "Index" is the Artificial Analysis Intelligence Index (v4.1, directional β treat as Β±, not decimals).
| Model | Lab | Tier | Context | Price (in/out) | Index | Weights | Best at |
|---|---|---|---|---|---|---|---|
| Claude Opus 5 | Anthropic | Flagship | 1M | $5 / $25 | ~61 | Closed | Overall #1; agentic coding + enterprise |
| Claude Fable 5 | Anthropic | Frontier | 1M | $10 / $50 | ~60 | Closed | Long-horizon autonomy, hardest reasoning |
| GPT-5.6 Sol | OpenAI | Flagship | 1.05M | $5 / $30 | ~59 | Closed | Hardest agentic + coding; GPQA leader |
| Kimi K3 | Moonshot | Open flagship | 1M | $3 / $15 | 57 | Open* | Top open intelligence; search/browsing |
| Gemini 3.1 Pro | Flagship | 1M | $2 / $12 | ~57 | Closed | Reasoning, science/math, multimodal | |
| GPT-5.6 Terra | OpenAI | Balanced | 1.05M | $2.50 / $15 | ~55 | Closed | Everyday balanced workhorse |
| Grok 4.5 | xAI | Flagship | 500K | $2 / $6 | ~54 | Closed | Cost-efficient agentic coding; legal agents |
| GLM-5.2 | Z.ai | Open | 1M | ~$0.30 blended | ~51 | Open (MIT) | Value agentic coding; beats GPT-5.5 for ~1/6 cost |
| GPT-5.6 Luna | OpenAI | Fast | 1.05M | $1 / $6 | ~51 | Closed | Fast, cost-sensitive frontier |
| Gemini 3.6 Flash | Workhorse | 1M | $1.50 / $7.50 | ~50 | Closed | High-volume default; computer use (OSWorld 83%) | |
| Claude Sonnet 5 | Anthropic | Workhorse | 1M | $3 / $15β | n/p | Closed | Default agentic workhorse (Terminal-Bench +20 pts) |
| DeepSeek V4-Pro | DeepSeek | Open | 1M | $0.44 / $0.87 | ~44 | Open (MIT) | Price/perf; easiest true-frontier to self-host |
| MiniMax M3 | MiniMax | Open | 1M | ~$0.30 (β$0.06 cached) | ~44 | Open (MIT) | Cheapest agentic coding + computer-use |
| Muse Spark 1.1 | Meta | Frontier | 1M | $1.25 / $4.25 | ~43 | Closed | Tool-use / orchestration leader (MCP Atlas 88.1) |
| Gemini 3.5 Flash-Lite | Lite | 1M | $0.30 / $2.50 | ~36 | Closed | High-throughput, low-latency, cheap | |
| Claude Haiku 4.5 | Anthropic | Fast | 200K | $1 / $5 | n/p | Closed | Speed/cost; subagents & fan-out |
| Grok 4.1 Fast | xAI | Long-context | 2M | low | n/p | Closed | Largest context window, cheap & fast |
| DeepSeek V4-Flash | DeepSeek | Open | 1M | $0.14 / $0.28 | n/p | Open (MIT) | Rock-bottom cost, ~85β90% of frontier |
| Llama 4 Maverick | Meta | Open | 1M | self-host | ~14 | Open | On-prem general/multimodal |
| Llama 4 Scout | Meta | Open | 10M | self-host | ~10 | Open | Ultra-long-context on a single H100 |
Kimi K3 ships open weights under a **custom license* (not OSI Apache/MIT) β check redistribution terms. β Sonnet 5 has an intro price of $2 / $10 through Aug 31, 2026. "n/p" = no clean Index published; positioned by tier. Indexes are AA v4.1, directional.
Now the part you came for β who wins which job.
π Overall Intelligence
The "smartest model, all-round" question. Artificial Analysis's composite is the least-bad single answer.
| Rank | Model | AA Index (v4.1) |
|---|---|---|
| 1 | Claude Opus 5 (max) | ~61 |
| 2 | Claude Fable 5 (max) | ~60 |
| 3 | GPT-5.6 Sol (max) | ~59 |
| 4 | Kimi K3 (open) | 57 |
| 5 | Gemini 3.1 Pro | ~57 |
The takeaway: the top of the board is a 7-point spread β narrow enough that for most work, "which of the top five" matters far less than which effort setting you run and how you route. The genuine headline is #4: an open-weight model (Kimi K3) is now inside the top five overall.
Manager's note: Opus 5 is the rare "flagship intelligence at workhorse price" β same $5/$25 as the previous Opus, ~#1 on the Index. If you default anything to a frontier model, default here.
π€ Agentic Tool Use & Long-Horizon Autonomy
This is my actual day job, so I care about this more than any other row β and it's the one the marketing screenshots hide, because it's where models are weakest.
Three different skills hide under "agentic," and different models win each:
| Sub-skill | Benchmark | Current leader | Score |
|---|---|---|---|
| Tool / MCP orchestration | MCP Atlas | Muse Spark 1.1 (Meta) | 88.1 |
| Computer use (GUI) | OSWorld-Verified | Gemini 3.6 Flash | 83% |
| Terminal / shell agents | Terminal-Bench 2.1 | GPT-5.6 Sol | 88.8% |
| Multi-turn policy adherence | ΟΒ²-bench | Step-3.5-Flash | 88.2% |
| Web browsing / research | BrowseComp | Kimi K3 (open) | 91.2 |
Two things worth internalizing:
1. Tool orchestration β raw IQ. Meta's Muse Spark 1.1 sits at Index ~43 β mid-pack on general intelligence β yet leads tool-use orchestration because it was built for primary-agent + parallel-subagent workflows with native MCP. If your system is mostly "call the right tools in the right order," the smartest model isn't necessarily the best agent.
2. Policy adherence is the real bar. ΟΒ²-bench doesn't just ask "did the agent complete the task" β it asks "did it complete the task without violating the stated policy." An agent that books the flight but ignores the change-fee rule fails. That maps exactly to enterprise reality, and it's why I trust ΟΒ²-style evals over flashier demos. Even the leaders top out in the high-80s here β a reminder that "autonomous agent" still needs guardrails and a human on irreversible actions.
Manager's note: For agent backbones I route to Opus 5 or GPT-5.6 Sol for judgment-heavy planning, but I'll drop a cheaper, tool-tuned model (Muse Spark, Gemini Flash, or an open model) into the high-volume tool-calling loops. The planner and the workers don't have to be the same model.
π» Coding
The most-tested capability, and the one where the "which benchmark" caveat bites hardest. SWE-bench Verified is saturated (Anthropic's top models sit at 95β96%), so I weight SWE-bench Pro and Terminal-Bench more heavily β they still discriminate.
| Benchmark | What it measures | Top models |
|---|---|---|
| SWE-bench Verified (saturated) | Real GitHub issue fixes | Opus 5 96% Β· Fable 5 ~95% Β· Gemini 3.1 Pro / DeepSeek V4-Pro 80.6% |
| SWE-bench Pro (harder, current) | Tougher, cleaner-tested repo tasks | Fable 5 80.3% Β· Opus 5 79.2% Β· Grok 4.5 64.7% Β· GPT-5.6 Sol 64.6% Β· GLM-5.2 62.1% |
| Terminal-Bench 2.1 | Multi-step shell/agent coding | GPT-5.6 Sol 88.8% Β· Kimi K3 88.3% Β· Fable 5 88.0% Β· Grok 4.5 ~83% |
The takeaway: Anthropic still owns the top of the coding table (Opus 5 / Fable 5), but the interesting story is the compression underneath β Grok 4.5, GPT-5.6 Sol, and the open GLM-5.2 are clustered within a few points on SWE-bench Pro. For 80% of real PRs, a mid-tier or open model closes the gap; save the frontier tier for the gnarly multi-file refactors.
Manager's note: DeepSeek V4-Pro hitting 80.6% SWE-bench Verified as an MIT-licensed, self-hostable model is the single most disruptive coding data point of the year. For teams with data-residency constraints, "frontier-adjacent coding you can run in your own VPC" is now real.
π§ Reasoning, Science & Math
Hard science QA, competition math, and abstract reasoning β the "can it actually think" cluster.
| Benchmark | Top models |
|---|---|
| GPQA Diamond (PhD science) | Gemini 3.1 Pro 94.3% β GPT-5.6 Sol 94.1% Β· Opus 5 ~93.5% Β· Kimi K3 93.5% |
| Humanity's Last Exam (no tools, AA-independent) | Fable 5 53.3% Β· GPT-5.6 Sol 47.2% Β· Gemini 3.1 Pro ~46% |
| ARC-AGI-2 (abstract reasoning) | Gemini 3.1 Pro 77.1% Β· Grok 4.5 52.6% |
| Competition math | Opus 5 β IMO 2026 42/42 (gold); open models (GLM-5, Qwen3.5) clear ~92% AIME 2026 |
The takeaway: this is the category where Google and OpenAI are strongest relative to their overall rank β Gemini 3.1 Pro's ARC-AGI-2 lead is meaningful for genuinely novel problem-solving, and it's tied for the GPQA crown. If your workload is scientific research, quantitative analysis, or hard multi-step reasoning, this is the one category where I might not default to Claude.
Caveat I keep having to repeat: you'll see Opus 5 and Muse Spark quoted at 64% and 62% on HLE β those are with-tools numbers. Against the no-tools column above, Fable 5's 53.3% is the honest leader. Never mix the two.
ποΈ Multimodal
Vision, documents, charts, mixed media.
| Benchmark | Top models |
|---|---|
| MMMU-Pro (multimodal reasoning) | Kimi K3 81.6 Β· Gemini 3.1 Pro 80.5 |
| Computer-use (screen understanding) | Gemini 3.6 Flash 83% Β· Muse Spark 80.8 Β· Opus 5 (OSWorld 2.0) 70.6 |
The takeaway: Gemini remains the multimodal default β natively strong across image/video/audio/PDF and now the computer-use leader β but Kimi K3 quietly leads MMMU-Pro, making it the strongest open multimodal option. For document-heavy or screen-driving agents, Gemini Flash is the value pick; for on-prem multimodal, Kimi K3.
π Long Context
When the job is "read all of it."
- Llama 4 Scout β 10M tokens. Still the largest usable window, and it fits on a single H100. Its general intelligence is low (Index ~10), but as a cheap, self-hosted "swallow an entire codebase/corpus" retriever, nothing matches the window.
- Grok 4.1 Fast β 2M tokens. The largest among the closed frontier-adjacent models, tuned for cheap high-speed long-context.
- Everyone else β ~1M. Opus 5, Fable 5, GPT-5.6, Gemini, Kimi K3, DeepSeek V4, MiniMax M3 all land at ~1M, which is enough for the vast majority of real workloads.
Manager's note: raw window size is oversold. A 1M-token model that actually reasons over the whole context beats a 10M-token model that skims. Test retrieval quality at depth, not the advertised number.
π° Value & Cost-Efficiency (The Real 2026 Story)
If there's one shift that reshaped my architecture this year, it's this: the price of "good enough" collapsed.
| Model | Price (in/out per 1M) | The pitch |
|---|---|---|
| DeepSeek V4-Flash | $0.14 / $0.28 | ~85β90% of frontier quality at ~8% of the cost |
| MiniMax M3 | ~$0.30 β $0.06 cached | Cheapest agentic-coding + computer-use, 1M context |
| GLM-5.2 | ~$0.30 blended | Beats GPT-5.5 on long-horizon coding for ~1/6 the cost |
| Grok 4.1 Fast | very low | 2M context at bargain rates |
| Gemini 3.5 Flash-Lite | $0.30 / $2.50 | Closed-model reliability at near-open pricing |
The takeaway: the Chinese open-weight labs (DeepSeek, Z.ai, MiniMax, Moonshot) have made frontier-adjacent performance at 1/6β1/30 the token cost the defining fact of 2026. DeepSeek V4's output is roughly 29Γ cheaper than Claude Opus 4.8's by their own framing. You are almost certainly overpaying if 100% of your traffic hits a US frontier model.
π Best Open-Weight Models
The open field moved so fast it deserves its own ranking β and the geographic shift is the headline.
| Rank | Model | Lab | Index | License |
|---|---|---|---|---|
| 1 | Kimi K3 | Moonshot (CN) | 57 | Custom (open weights) |
| 2 | GLM-5.2 | Z.ai / Zhipu (CN) | 51 | MIT |
| 3 | DeepSeek V4-Pro | DeepSeek (CN) | 44 | MIT |
| 3 | MiniMax M3 | MiniMax (CN) | 44 | MIT |
| 5 | Qwen3.5-397B | Alibaba (CN) | ~40 | Apache 2.0 |
| β | Mistral Large 3 | Mistral (EU) | 16 | Apache 2.0 |
| β | Llama 4 Maverick | Meta (US) | 14 | Community |
The takeaway: in 2026 the open-weight frontier is, bluntly, Chinese. Meta's Llama 4 has slipped to the bottom of the pack, Behemoth was shelved, and Meta's real frontier effort β Muse Spark β is now closed, API-only, US-only. The torch for "best model you can actually download and self-host" has passed to Moonshot, Z.ai, DeepSeek, and Alibaba. For sovereignty, cost control, or air-gapped deployment, that's where you look now. (Europe's best Apache-2.0 option, Mistral Large 3, is a capable generalist but trails on reasoning/agentic evals.)
License trap: don't confuse a family's open and closed tiers. Qwen3.7-Max, Mistral Medium 3.5, and Amazon Nova are closed; the open ones are Qwen3.5/3.6, Mistral Large/Small. And Kimi K3's weights are open but under a custom license β read the redistribution terms before you ship on it.
The Routing Map I Actually Use
Here's the decision tree I gave my team. It's opinionated on purpose β defaults beat deliberation at scale.
flowchart TD
A[New task] --> B{What shape is it?}
B -->|Hardest reasoning /<br/>long-horizon autonomy| F[Claude Fable 5<br/>or Opus 5 - max effort]
B -->|Agentic coding /<br/>most PRs| O[Claude Opus 5 /<br/>GPT-5.6 Sol]
B -->|Tool orchestration /<br/>MCP workflows| M[Muse Spark 1.1 /<br/>Gemini 3.6 Flash]
B -->|Science / math /<br/>novel reasoning| G[Gemini 3.1 Pro /<br/>GPT-5.6 Sol]
B -->|High-volume /<br/>cost-sensitive| V[GLM-5.2 / DeepSeek V4 /<br/>Gemini Flash-Lite]
B -->|On-prem / sovereign /<br/>air-gapped| SH[Kimi K3 / GLM-5.2 /<br/>DeepSeek V4 - self-host]
B -->|Swallow a huge corpus| LC[Llama 4 Scout 10M /<br/>Grok 4.1 Fast 2M]
B -->|Fast glue / subagents| H[Claude Haiku 4.5 /<br/>Gemini Flash-Lite]
The Agentic AI Manager's Playbook (Steal These)
Seven habits that separate a sane multi-model stack from a runaway bill:
- Route, don't standardize. The single highest-leverage decision is admitting no model wins everything. Wire an abstraction layer (MCP or a gateway) so swapping a model per task is a config change, not a rewrite.
- Default cheap, escalate on failure. Start tasks on a mid or open model; promote to a frontier model only when the cheap one visibly stalls. Most teams can push 70β90% of traffic to cheap models with no quality loss.
- Split the planner from the workers. Use a frontier model (Opus 5 / GPT-5.6 Sol) for judgment-heavy planning, and cheap tool-tuned models for the high-volume tool calls underneath. They don't have to match.
- Benchmark on your eval, not theirs. Public benchmarks are saturated and gamed. Build a 50-task internal eval from your real workload β it will rank models differently than any leaderboard, and it's the only ranking that pays your bills.
- Measure policy adherence, not just success. For any agent that touches money, data, or customers, test whether it follows rules, not just whether it finishes. ΟΒ²-bench thinking, applied to your domain.
- Keep a fallback wired at all times. Frontier availability is volatile β export controls, capacity, deprecations. Have a second-vendor path (ideally an open model you can self-host) ready before you need it.
- Never let cost-per-token pick your architecture alone. A model that's 10Γ cheaper but needs 3Γ the retries and a human to catch policy violations isn't cheaper. Measure cost-per-successful-outcome.
How To Choose in 30 Seconds
- "I just want the best, money's no object." β Claude Opus 5 (or Fable 5 for long-horizon).
- "Ship code / fix PRs." β Opus 5 or GPT-5.6 Sol; GLM-5.2 if cost matters.
- "Build a tool-using agent." β Muse Spark 1.1 or Gemini 3.6 Flash for the loops, a frontier model for the planner.
- "Science / math / research." β Gemini 3.1 Pro or GPT-5.6 Sol.
- "Cheapest thing that's still good." β DeepSeek V4-Flash or GLM-5.2.
- "Must run on-prem / in my VPC." β Kimi K3, GLM-5.2, or DeepSeek V4 (all self-hostable).
- "Read a giant corpus." β Llama 4 Scout (10M) or Grok 4.1 Fast (2M).
- "Fast, high-volume glue." β Claude Haiku 4.5 or Gemini 3.5 Flash-Lite.
Final Take: The Skill Is Routing
A year ago the question was "which model is smartest?" In 2026 that question is a trap. The board is compressed at the top, the classic benchmarks are saturated, and the most important number on any model card is no longer its Index score β it's the cost-per-successful-outcome on your workload.
The models are a fleet now. Claude Opus 5 for the hard judgment calls, GPT-5.6 and Gemini for reasoning and multimodal, Muse Spark for orchestration, and a Chinese open-weight model quietly doing 80% of the volume in your VPC at a tenth of the cost. The teams winning with AI in 2026 aren't the ones who picked the "best" model. They're the ones who stopped picking one β and got good at routing.
Build your own eval. Wire your own fallback. Route by the shape of the work. The leaderboard is a starting point, not an answer.
About the Author
Suraj Khaitan β Senior Agentic AI Manager | Building and scaling production agentic systems on the cloud
Connect on LinkedIn | Follow for more engineering and architecture write-ups
Which model has become your default β and what finally made you route away from it? Drop it in the comments. I'm always refining the map.
Sources & further reading: Artificial Analysis LLM Leaderboard (primary cross-model source) Β· Anthropic: Claude Opus 5, Sonnet 5, Fable 5 & Mythos 5 Β· OpenAI: GPT-5.6 Sol preview, AA's GPT-5.6 analysis Β· Google: Gemini pricing, three new Gemini models Β· xAI: Grok 4.5 Β· Meta: Muse Spark 1.1, Llama 4 Β· Open weights: Kimi K3, GLM-5.2, DeepSeek V4. All benchmarks reflect published/independent figures as of late July 2026 and are effort- and harness-dependent; treat as directional.
Top comments (0)