LMArena Leaderboard 2026: Top 10 AI Models by Elo and Cost

TL;DR (verified 2026-08-31): Claude Opus 5 tops both task benchmarks — 97.00% on vals.ai SWE-Bench Verified and 63.1 on the Artificial Analysis Intelligence Index — at $5/$25 per million tokens. DeepSeek V4 Pro 0813 is second on SWE-Bench at 96.40% for $1.32/$3.96, which makes it the best score-per-dollar at the frontier. On LMArena, where humans vote blind, the order is different: Claude Fable 5 leads at 1507 Elo and Opus 5 High is seventh. Seven of the 86 models vals.ai evaluates now clear 95% on SWE-Bench, so the interesting question stopped being “which model is best” and became “which model is best per dollar.”

Top 10 At a Glance

Ranked by vals.ai SWE-Bench Verified, the strongest independent coding measure. Prices are ofox.ai catalog rates.

SWE-Bench column: vals.ai SWE-bench Verified, leaderboard updated 2026-08-26, 86 models evaluated. AA column: Artificial Analysis Intelligence Index, pulled 2026-08-31, effort tier noted per row because the same model scores differently at different reasoning settings. Prices: ofox.ai catalog, 2026-08-31. Em-dashes mean the model is not currently carried on that leaderboard, not that it scored zero.

Opus 4.8 is included as row 11 for one reason: it led this benchmark at 88.60% when the previous version of this page was written, and it is now eleventh. That is how fast the top of this list moves.

How These Rankings Work

Three independent leaderboards measure different things, and they disagree — which is the point.

LMArena (formerly LMSYS Chatbot Arena) uses blind human preference votes. Two models answer the same prompt; users pick the better response without knowing which model is which. Scores are Elo ratings, the same system used in chess. The public text leaderboard currently carries 395 model entries.

SWE-Bench Verified measures whether a model can resolve real GitHub issues. An agent gets a repo, a bug report, and a test suite. It passes if the tests go green. No partial credit. vals.ai runs it with a deliberately minimal harness — the model gets one tool, bash, and has to navigate, search, and edit with standard command-line utilities. That choice matters: SWE-Bench scores published by model vendors usually come from richer scaffolding, so vendor numbers and vals.ai numbers are not comparable even when they name the same benchmark.

Artificial Analysis Intelligence Index aggregates multiple evaluations into a single composite score. Its rows are per-effort-tier, not per-model: Claude Opus 5 appears separately at max, xhigh, high, and medium effort, spanning 63.1 down to 58.6. Quoting “the Opus 5 score” without naming an effort tier is quoting one of four different numbers.

One structural caveat before the tables. SWE-Bench Verified is now saturating — the leader sits 3.00 points from a perfect score and seven models are inside the top five points. At that compression, a one-point gap is inside the noise of harness luck, and the cost column separates models better than the accuracy column does.

LM Arena — Human Preference Leaderboard

LMArena text leaderboard, pulled 2026-08-31 (source):

Three things worth knowing before you use this table.

Anthropic holds six of the top ten, and four of those six are Opus 4.6 and 4.7 variants — models that sit well down the coding leaderboard. Arena rewards conversational polish. SWE-Bench rewards finished code. A model can be excellent at one and unremarkable at the other, and these two lists are the cleanest demonstration of that available.

Opus 5 ranks seventh here and first on both task benchmarks. If you pick a model off LMArena for a coding agent, you are optimising for the wrong thing.

Some entries are codenames. andwise-evfd is a Google entry under a test alias, and the leaderboard carries several others like it. Codenamed rows are real evaluations, but you cannot buy the model behind them, and the name will change when it ships. Vote counts vary by two orders of magnitude across the top ten — muse-spark-1.2-xhigh-text at 3,247 votes carries a much wider confidence interval than claude-opus-4-6 at 76,079.

For teams evaluating xAI’s API — Grok 4.6 pricing, model IDs, and a working setup — see the Grok 4.6 API pricing guide for the 200K context cliff that doubles your bill mid-request.

Best for Coding — SWE-Bench Verified

Claude Opus 5 leads at 97.00% (vals.ai, leaderboard updated 2026-08-26). The column that actually decides procurement is the one next to it.

Cost-per-test and latency are vals.ai’s own measurements from the same runs that produced the accuracy column, so the three numbers in each row are internally consistent. They are not derived from list prices.

The accuracy spread across the top ten is 8.2 points. The cost spread is 129x. DeepSeek V4 Flash 0731 resolves 88.80% of tasks at $0.010 per test; Claude Fable 5 resolves 95.00% at $2.047. You are paying roughly two hundred times more per attempt for 6.2 points of accuracy. Whether that trade is correct depends entirely on what a failed task costs you — for a supervised coding assistant where a human reviews every diff, it rarely is; for an unattended migration running overnight, it often is.

Open weights are at the top now, not near it. DeepSeek V4 Pro 0813 is second overall, 0.60 points behind the closed leader, and GLM-5.3 and Kimi K3 are both inside the top eight. vals.ai’s own summary notes that closed models still trend higher on this benchmark overall, but the gap at the very top has closed.

Latency does not track price or accuracy. GLM-5.3 is the sixth most accurate model and the slowest in the top ten at 841 seconds per task. GPT-5.6 Terra is a fifth of a point behind it and finishes in 180 seconds. If your agent runs interactively, that difference is the whole product experience.

For the deeper cost breakdown on the two Chinese-lab entries, see the DeepSeek V4 Pro 0813 price and benchmark guide and the GLM-5.3 API pricing and reasoning-effort guide, which covers why an unset effort parameter defaults to max and bills accordingly.

Best for Reasoning — Composite Intelligence Index

The Artificial Analysis Intelligence Index aggregates several reasoning evaluations into one composite. Pulled 2026-08-31, restricted to non-deprecated models with measured (not estimated) scores:

Blended price is Artificial Analysis’s own metric: a 3:1 input-to-output weighting applied to each vendor’s list price, so (3 x input + output) / 4. It does not match the gateway rates in the tables above, because those are what ofox charges and this is what the vendor lists. Use it to compare rows within this table, not against anything else.

Two observations. Grok 4.6 at high effort ties GPT-5.6 Sol at max effort (60.9 each) for 37% of the blended price. And GLM-5.3 at max reaches 59.5 for $2.15 blended, which is the cheapest entry anywhere near the top of this index — a fifth of Opus 5, within 3.6 points.

The effort-tier point deserves repeating because it is the most common way this leaderboard gets misquoted. Opus 5 spans 63.1 (max) to 58.6 (medium). Grok 4.6 spans 60.9 (high) to 59.0 (medium). Reading a single number off a screenshot without the tier label tells you almost nothing. Our own Opus 5 versus Grok 4.6 cost measurement covers what those tiers do to a real bill rather than a benchmark.

Best Value — Price-Performance

Prices and context lengths from the ofox.ai catalog on 2026-08-31. Context is what the gateway advertises; if you are building close to the limit, confirm against the model provider’s own card before you rely on it.

The two rows that matter. GPT-5.6 Luna resolves 93.00% of SWE-Bench tasks at $0.20/$1.20 — four points behind the leader at a twentieth of the input price. GLM-5.3 resolves 95.40% at $1.26/$3.96, 1.6 points behind Opus 5 at a quarter of the price. Neither is a fringe pick; both sit in the top ten of an 86-model independent leaderboard.

Where the money still buys something. Opus 5’s 97.00% is 1.6 points above GLM-5.3 and 0.6 above DeepSeek V4 Pro. On a hundred-task batch that is one or two extra solved tasks. If those tasks are load-bearing and a human has to clean up the misses, the premium pays for itself; if you are generating boilerplate at volume, it does not.

Context is no longer the differentiator it was. Twelve of the fifteen rows above carry a million tokens or more. Grok 4.6’s 500K is the outlier at the top of the list, and its pricing doubles above a 200K threshold, which is the real constraint rather than the window size.

Best Open-Weight Model

DeepSeek V4 Pro 0813 is the strongest open-weight model on any independent leaderboard right now: second overall on vals.ai SWE-Bench Verified at 96.40%, 0.60 points behind Claude Opus 5, at $1.32/$3.96 through the API. Weights are published as deepseek-ai/DeepSeek-V4-Pro-0813 on Hugging Face, dated 2026-08-13, with unsloth GGUF and NVIDIA NVFP4 conversions already up. That is a different situation from the one this page described three months ago, when the best open model sat well below the closed frontier and was “worth considering for cost reasons.” It is now competitive on the merits.

GLM-5.3 is the second name in this band — 95.40% SWE-Bench, 59.5 AA Index at max effort, $1.26/$3.96 — and the weights landed on 2026-08-25 as zai-org/GLM-5.3, with GLM-5.3-Flash alongside it. GGUF and FP8 conversions followed within days. Check the GLM-5.3 benchmarks and access guide for the API side.

Kimi K3 weights are published — the moonshotai/Kimi-K3 repository on Hugging Face carries 118 files and has passed 2.7 million downloads, with community quantisations from unsloth, nvidia and others. At 93.40% SWE-Bench and 59.7 AA Index it is the strongest model you can actually download, though at its parameter count self-hosting means a cluster, not a workstation. The Kimi K3 access guide covers the routes in.

Which Model Should You Use?

Pick by task, not by headline ranking.

Coding agents, quality first. Claude Opus 5 — 97.00% SWE-Bench Verified, $5/$25, 1M context. Budget $1.29 per resolved task at vals.ai’s measured rate.

Coding agents, cost first. DeepSeek V4 Pro 0813 at 96.40% for $0.103 per test, or GLM-5.3 at 95.40% for $0.338. Both are within 1.6 points of the leader at a tenth of the cost per attempt. This is the single biggest change on this page since the last update.

Interactive tools where latency shows. GPT-5.6 Terra: 95.40% at 180 seconds per task, the fastest in the top ten by a wide margin. GLM-5.3 matches its accuracy and takes 841 seconds.

High-volume production. GPT-5.6 Luna ($0.20/$1.20, 93.00%) or DeepSeek V4 Flash 0731 ($0.44/$1.32, 88.80%). Escalate the failures to Opus 5 rather than paying frontier rates for every call.

Reasoning at frontier. Claude Opus 5 at max effort (63.1) or Grok 4.6 at high (60.9 for a third of the blended price). Name the effort tier in your config; the default is not always what you think.

Human-facing chat and writing. This is where LMArena is the right leaderboard and the coding tables are the wrong one — Claude Fable 5, Opus 4.6 and 4.7 variants dominate the top of the Elo list.

Multimodal and OCR. The picks shift on this axis. See the best AI model for OCR.

The full Claude vs. GPT vs. Gemini comparison guide goes deeper on use-case-specific recommendations across the three frontier families.

Where to Access These Models

Every model in the tables above is available through ofox.ai with a single API key and OpenAI-compatible endpoints, including Claude Fable 5 — a live call to anthropic/claude-fable-5 returned normally on 2026-08-31 at the standard catalog rate.

One key, one billing account, one SDK integration. Switch models by changing one string:

Python (OpenAI SDK):

from openai import OpenAI

client = OpenAI(
    base_url="https://api.ofox.ai/v1",
    api_key="your-ofox-key",
)

# Switch between any model on this leaderboard by changing `model`.
response = client.chat.completions.create(
    model="anthropic/claude-opus-5",  # or "openai/gpt-5.6-sol", "z-ai/glm-5.3"
    messages=[{"role": "user", "content": "Refactor this Django view to use async..."}],
)
print(response.choices[0].message.content)

Node.js (OpenAI SDK):

import OpenAI from "openai";

const client = new OpenAI({
  baseURL: "https://api.ofox.ai/v1",
  apiKey: process.env.OFOX_API_KEY,
});

const response = await client.chat.completions.create({
  model: "deepseek/deepseek-v4-pro-0813", // or "moonshotai/kimi-k3", "x-ai/grok-4.6"
  messages: [{ role: "user", content: "Refactor this Django view to use async..." }],
});
console.log(response.choices[0].message.content);

Model IDs in the tables above (provider/model-version) are what go in the model parameter — they are not OpenAI’s or Anthropic’s native naming. For a complete setup guide, including failover and per-team budget routing, see AI API aggregation: access every model through one endpoint.

References

• vals.ai SWE-bench Verified
• LMArena text leaderboard
• Artificial Analysis Intelligence Index
• ofox.ai model catalog
• moonshotai/Kimi-K3 on Hugging Face
• deepseek-ai/DeepSeek-V4-Pro-0813 on Hugging Face
• zai-org/GLM-5.3 on Hugging Face

Last updated: August 31, 2026. SWE-Bench figures from the vals.ai leaderboard as updated 2026-08-26 (86 models, bash-only harness); LMArena Elo and Artificial Analysis Index pulled 2026-08-31; prices from the ofox.ai catalog on 2026-08-31. For the June 2026 snapshot with Claude Opus 4.8 in the top slot, and the April ranking before it, see AI Model Rankings May 2026.