AI Shqipe
valon-hajredini.aishqipe.com/posts/ai-model-rankings-august-2026-text-coding-vision-image-video
Valon Hajredini
Subscribe

Who Leads AI Right Now? Model Rankings Across Text, Coding, Vision, Image and Video (August 2026)

One lab tops the chat and coding boards, another owns image generation, and text-to-video now belongs to China. A tour of the August 2026 leaderboards, with the numbers, the caveats and the charts.

Ask "which AI model is best?" in August 2026 and the honest answer is: best at what? The chat arena, the coding benchmarks, the vision board, the image arena and the video arena are each led by a different model, sometimes by a different company, and the gaps between them are not the same size. This post walks through those five angles, plus speed and price, using numbers read from Artificial Analysis, LMArena, Scale's SEAL SWE-bench Pro leaderboard and llm-stats on 21 August 2026. Every chart below is drawn from those numbers, so you can check them.

First, how to read a leaderboard

Three things decide whether a ranking means what it appears to mean. The first is what is being measured: an arena Elo is human preference in blind A/B votes, while a benchmark score is pass rate on a fixed task set, and the two regularly disagree. The second is who ran the test: a vendor-reported SWE-bench number comes from the lab's own tuned agent scaffold, and is routinely 10–30 points higher than the same model under a neutral harness. The third is the confidence interval. Arena scores carry ±6 to ±18 points of uncertainty, so a model that is "#3" with a rank spread of 1–12 is statistically tied with #1.

Everything below names its source and its date, and where two sources disagree I show both rather than pick one. Leaderboards move weekly; treat this as a dated snapshot, not a verdict.

Text and reasoning: Anthropic on top, a crowded second row

The Artificial Analysis Intelligence Index is a composite of reasoning, knowledge, maths and science evaluations (GPQA Diamond, Humanity's Last Exam, and others), run by Artificial Analysis themselves rather than self-reported. On 21 August 2026 the top of the table is Claude Opus 5 at 63 (both max and xhigh effort), Claude Fable 5 at 62, then a four-way tie at 61 between Claude Opus 5 (high), GPT-5.6 Sol (max) and Grok 4.6 (high). Kimi K3 and GLM-5.3 at 60 are the open-weight models closest to the frontier, and Qwen3.8 Max at 58 is only five points off the lead.

Claude Opus 5 (max)
63Anthropic. $2.34 per Intelligence Index task, 60 output tokens/s, 1M context. Ties its own xhigh setting for the top score.
Claude Fable 5
62Anthropic. $3.14 per task, 73 tokens/s — the most expensive model on the board per task.
Claude Opus 5 (high)
61Anthropic. $1.23 per task, 59 tokens/s. Two points below max effort at roughly half the cost.
GPT-5.6 Sol (max)
61OpenAI. $1.23 per task, 71 tokens/s. The default model in Codex.
Grok 4.6 (high)
61SpaceXAI. $0.84 per task, 68 tokens/s, 500k context. Cheapest way to buy a 61.
Kimi K3 (max)
60Moonshot, open weights. $0.84 per task, 40 tokens/s. Also leads the front-end code arena.
GLM-5.3 (max)
60Z AI, open weights. $0.68 per task, 93 tokens/s — the cheapest 60 on the board.
Qwen3.8 Max
58Alibaba. $1.13 per task, 45 tokens/s. Also ranks #2 on the LMArena Vision Arena.
GPT-5.6 Terra (max)
57OpenAI. $0.51 per task, 115 tokens/s. OpenAI's everyday production tier.
Gemini 3.7 Flash (high)
56Google. $0.40 per task and about 390 output tokens/s — six times the speed of the models above it.
Claude Sonnet 5 (max)
55Anthropic. $1.72 per task, 87 tokens/s.
DeepSeek V4 Pro 0813 (max)
53DeepSeek, open weights. $0.25 per task, 77 tokens/s, 1M context.
Artificial Analysis Intelligence Index, top models by score, read 21 August 2026. Composite of reasoning, knowledge, maths and science evals run by Artificial Analysis. Hover or tap a bar for cost and speed. Source: artificialanalysis.ai/leaderboards/models.

Human preference tells the same story with a different texture. On LMArena's text leaderboard Claude Mythos 5 (1531), Claude Fable 5 (1525) and Claude Opus 5 (1522) hold the first three slots, with GPT-5.6 (1514), Claude Opus 4.8 (1512), GPT-5.5 (1506), Gemini 3.1 Pro Preview (1505) and Kimi K3 (1500) behind them. The whole top ten sits within about 40 Elo points, which at typical arena sample sizes means the ordering below #3 is soft. The chart below shows each model's distance from the leader rather than raw Elo, because a 1491-vs-1531 bar drawn from zero would look like a dead heat.

Claude Mythos 5
1531 EloAnthropic. Ceiling-capability model with limited access — approved partners only, so it tops the board without being generally usable.
Claude Fable 5
1525 EloAnthropic. 6 behind the leader — inside the confidence interval, so this is a tie. Also leads the Vision Arena and the vendor-reported SWE-bench Pro table.
Claude Opus 5
1522 EloAnthropic, released 24 July 2026. 9 behind, and top of the Artificial Analysis Intelligence Index at 63.
GPT-5.6
1514 EloOpenAI, July 2026. The highest-placed non-Anthropic model on the text board.
Claude Opus 4.8
1512 EloAnthropic, May 2026.
GPT-5.5
1506 EloOpenAI, April 2026.
Gemini 3.1 Pro Preview
1505 EloGoogle. Still the flagship Pro model, but scores 48 on the Intelligence Index — below Google's own Gemini 3.7 Flash at 56.
Kimi K3
1500 EloMoonshot AI, open weights, July 2026. 31 behind the leader — the closest open model on this board.
Grok 4.5
1499 EloSpaceXAI, July 2026. $2 in / $6 out per million tokens.
Qwen3.8 Max
1491 EloAlibaba, August 2026. 40 behind the leader closes out a top ten that spans only about 40 Elo points in total.
LMArena text leaderboard, Elo points behind the leader (Claude Mythos 5, 1531). Shorter is better. Hover or tap a bar for the absolute Elo. Scores as aggregated by Swfte's tracker on 17 August 2026; 95% intervals are roughly ±7 to ±11 points.

Two observations. The gap between closed and open weights at the top of the text board is now single digits of Elo and two or three index points: Kimi K3, GLM-5.3 and Qwen3.8 Max are genuinely frontier-adjacent. And the famous names are not where they were a year ago: Gemini 3.1 Pro Preview, still Google's flagship Pro model, scores 48 on the Intelligence Index, below Google's own Gemini 3.7 Flash at 56.

Coding: three numbers that all claim to be the best

SWE-bench Verified is effectively retired as a frontier benchmark: OpenAI stopped reporting it in early 2026 and the top models clear 80–94%. The live coding board is SWE-bench Pro, Scale AI's 1,865-task set across 41 repositories, and here you have to be careful, because three different numbers are all quoted as "the SWE-bench Pro score". Vendor-reported, on each lab's own agent scaffold, Claude Fable 5 leads at 80.0%, followed by Claude Opus 4.8 at 69.2%, Qwen3.8 Max at 67.7%, Grok 4.5 at 64.7% and GPT-5.6 Sol at 64.6%. Claude Opus 5, released 24 July, has no Pro entry yet.

Claude Fable 5: 80%Claude Opus 4.8: 69.2%Qwen3.8 Max: 67.7%Grok 4.5: 64.7%GPT-5.6 Sol: 64.6%Claude Opus 4.7: 64.3%GPT-5.6 Terra: 63.4%Claude Sonnet 5: 63.2%GPT-5.6 Luna: 62.7%GLM-5.2 (open): 62.1%Muse Spark 1.1: 61.5%Qwen3.7 Max (open): 60.6%MiniMax M3 (open): 59%Gemini 3.6 Flash: 58.7%0%20%40%60%80%Claude Fable 5Grok 4.5GPT-5.6 TerraGLM-5.2 (open)MiniMax M3 (open)Pass@1Claude Fable 5: 80%Claude Opus 4.8: 69.2%Qwen3.8 Max: 67.7%Grok 4.5: 64.7%GPT-5.6 Sol: 64.6%Claude Opus 4.7: 64.3%GPT-5.6 Terra: 63.4%Claude Sonnet 5: 63.2%GPT-5.6 Luna: 62.7%GLM-5.2 (open): 62.1%Muse Spark 1.1: 61.5%Qwen3.7 Max (open): 60.6%MiniMax M3 (open): 59%Gemini 3.6 Flash: 58.7%Claude Fable 5Claude Opus 4.8Qwen3.8 MaxGrok 4.5GPT-5.6 SolClaude Opus 4.7GPT-5.6 TerraClaude Sonnet 5GPT-5.6 LunaGLM-5.2 (open)Muse Spark 1.1Qwen3.7 Max (open)MiniMax M3 (open)Gemini 3.6 Flash0%50%100%Pass@1
SWE-bench Pro, vendor-reported scores on each lab's own scaffold, as aggregated by llm-stats and read 10 August 2026. Comparable to each other, not to the standardized Scale numbers below. Claude Opus 5 has no entry yet.
Show the 14 numbers behind this chart
SWE-bench Pro, vendor-reported scores on each lab's own scaffold, as aggregated by llm-stats and read 10 August 2026. Comparable to each other, not to the standardized Scale numbers below. Claude Opus 5 has no entry yet. — the numbers this chart is drawn from
CategoryPass@1
Claude Fable 580%
Claude Opus 4.869.2%
Qwen3.8 Max67.7%
Grok 4.564.7%
GPT-5.6 Sol64.6%
Claude Opus 4.764.3%
GPT-5.6 Terra63.4%
Claude Sonnet 563.2%
GPT-5.6 Luna62.7%
GLM-5.2 (open)62.1%
Muse Spark 1.161.5%
Qwen3.7 Max (open)60.6%
MiniMax M3 (open)59%
Gemini 3.6 Flash58.7%

Scale's own standardized run, which puts every model through the same scaffolding, tells a humbler story: the public-set leader is GPT-5.4 (xHigh) at 59.1%, then Meta's Muse Spark at 55.0% and Claude Opus 4.6 (thinking) at 51.9%. Scale has not yet published standardized entries for Fable 5, Opus 4.8, GPT-5.5 or GLM-5.2, which is why they are absent. The most useful number is the private one: 276 tasks from proprietary codebases no model has seen. Scores drop hard and the ranking reshuffles, with Opus 4.6 holding best (47.1%, down only 4.8 points) while GPT-5.4 falls to third and Gemini 3.1 Pro loses 13.9 points.

GPT-5.4 (xHigh)
59.1% → 43.4%OpenAI. Leads the public set by 4.1 points but drops 15.7 points on unseen commercial code, falling to third. 95% CI on the public set is ±3.56.
Muse Spark
55.0% → 44.7%Meta. Second on both sets, losing 10.3 points between them.
Claude Opus 4.6 (thinking)
51.9% → 47.1%Anthropic. Third on the public set but first on private code, losing only 4.8 points — the smallest drop of any model here.
Gemini 3.1 Pro (thinking)
46.1% → 32.2%Google. Loses 13.9 points on unseen repositories, the second-largest fall on the board.
GPT-5.2 Codex
41.0% → 27.7%OpenAI. Its successor GPT-5.3 Codex reports 56.8% on OpenAI's own scaffold — a good illustration of the vendor-versus-standardized gap.
Gemini 3 Pro (preview)
43.3% → 18.0%Google. Drops 25.3 points on private code, the largest fall measured — a strong hint of public-set contamination.
SWE-bench Pro under Scale AI's standardized scaffolding: public set (731 tasks) against the private commercial set (276 unseen repositories), June 2026. The private bar is the better predictor of how a model will do on your codebase. Source: Scale SEAL leaderboards via morphllm.com.

Artificial Analysis also publishes a separate Coding Index. As surfaced on modelgrep's tracker, Claude Opus 5 leads it at 78.0, with GPT-5.6 Sol (77.4), Grok 4.6 (76.8), GPT-5.6 Terra (76.7), Claude Fable 5 (76.5), Kimi K3 (76.2) and Gemini 3.7 Flash (76.1) packed within two points. On the human-voted side, Claude Opus 4.7 held #1 on LMArena's Hard Prompts and Coding categories at the start of August, and Kimi K3 leads the front-end code arena. The practical reading: at the top, coding ability is close enough that scaffold quality, price and speed decide the outcome more than the model's rank.

Vision: reading images, charts and screenshots

LMArena's Vision Arena ranks multimodal models on tasks that start with an image: captioning, OCR, diagrams, homework, entity recognition. As of 6 August 2026 (1.19 million votes, 146 models) Claude Fable 5 leads at 1315 ±9, with Qwen3.8 Max (1301), Claude Opus 4.7 high (1301), Claude Opus 4.6 high (1300) and Claude Opus 5 high (1297) in a tight cluster behind it. Gemini 3.6 Flash high (1295), Meta's Muse Spark (1294), Gemini 3 Pro (1289), GPT-5.5 (1286) and Grok 4.5 (1285) complete the top ten. Four of the top six are Anthropic models, and the notable newcomer is Qwen3.8 Max at #2 with $2 per million input tokens.

Claude Fable 5
1315 Elo — the leaderAnthropic. This bar is empty because the chart measures distance behind the leader and Fable 5 is the leader: zero points behind itself. $10 in / $50 out per million tokens — the most expensive model
Qwen3.8 Max
1301 EloAlibaba. 14 behind, at about $2 per million input tokens — the standout price-to-quality point on this board.
Claude Opus 4.7 (high)
1301 EloAnthropic. Tied with Qwen3.8 Max on Elo; also held #1 on LMArena's Hard Prompts and Coding categories in early August.
Claude Opus 4.6 (high)
1300 EloAnthropic. 15 behind — inside the confidence interval, so positions 2 through 5 are a statistical tie.
Claude Opus 5 (high)
1297 EloAnthropic's newest model scores below two of its predecessors on vision, though the gap is within the interval.
Gemini 3.6 Flash (high)
1295 EloGoogle. 20 behind the leader at $0.38 in / $1.88 out per million tokens — roughly a twenty-fifth of Fable 5's price. Start here for OCR and document work at volume.
Muse Spark
1294 EloMeta. The best-placed model from a lab with no entry in the top five of the text board.
Gemini 3 Pro
1289 EloGoogle, November 2025. Still ahead of GPT-5.5 on images despite being the older model.
GPT-5.5
1286 EloOpenAI. 29 behind — OpenAI's highest-placed model on the vision board.
Grok 4.5
1285 EloSpaceXAI. Closes out a top ten spanning just 30 Elo points, on 1.19 million votes across 146 models.
LMArena Vision Arena, overall category, Elo points behind the leader (Claude Fable 5, 1315). Shorter is better, and the leader's bar is empty because it is zero points behind itself. Hover or tap for absolute Elo. Snapshot 6 August 2026, style control on; 95% intervals ±7 to ±18.

Vision is the board where the price spread is most dramatic. Fable 5 at the top costs $10 in and $50 out per million tokens; Gemini 3.6 Flash, 20 Elo behind, is $0.38 and $1.88. If your workload is OCR, document extraction or screenshot reading at volume, the cheap end of this list is where to look first, and you should run your own sample before paying for the top.

Image generation: OpenAI's biggest lead anywhere

On Artificial Analysis's text-to-image arena GPT Image 2 (high) leads at 1368 Elo, 47 points clear of Reve 2.1 (1321) and Google's Nano Banana 2 (1319), the largest first-to-second gap the board has recorded. Behind them sit GPT Image 1.5 (1310), Microsoft's MAI-Image-2.5 (1305), Nano Banana Pro (1298), MAI-Image-2.5-Pro (1292), Nano Banana 2 Lite (1289), Qwen-Image-3.0-Pro (1283) and ByteDance's Seedream 5.0 Pro (1279). llm-stats' independent arena agrees on the winner: GPT Image 2 at the top, then GPT Image 1.5 and MAI-Image-2.5. Hover or tap the points below to see each model's Elo and API price.

GPT Image 2 (high)
1368 EloOpenAI, April 2026. $211 per 1,000 images. Rank 1 with the largest first-to-second gap the board has recorded; best at text-in-image and complex multi-subject prompts.
Reve 2.1
1321 EloReve, July 2026. $200 per 1,000 images. Rank 2–3, statistically tied with Nano Banana 2.
Nano Banana 2 (Gemini 3.1 Flash Image)
1319 EloGoogle, February 2026. $67 per 1,000 images. Rank 2–4. The best quality-per-dollar point near the top of the board.
GPT Image 1.5 (high)
1310 EloOpenAI, December 2025. $133 per 1,000 images. Rank 3–5.
MAI-Image-2.5
1305 EloMicrosoft AI, June 2026. $48.10 per 1,000 images. Rank 4–6 at roughly a quarter of GPT Image 2's price.
Nano Banana Pro (Gemini 3 Pro Image)
1298 EloGoogle, November 2025. $134 per 1,000 images. Rank 5–8.
MAI-Image-2.5-Pro
1292 EloMicrosoft AI, July 2026. $108.50 per 1,000 images. Rank 6–8; scores below the cheaper non-Pro sibling.
Nano Banana 2 Lite (Gemini 3.1 Flash Lite Image)
1289 EloGoogle, June 2026. $33.60 per 1,000 images. Rank 6–9; the cheapest model in the top ten.
Qwen-Image-3.0-Pro
1283 EloAlibaba, July 2026. $43 per 1,000 images. Rank 7–11; added to the board in the last month.
Seedream 5.0 Pro
1279 EloByteDance, July 2026. $90 per 1,000 images. Rank 9–11.
Qwen-Image-3.0
1275 EloAlibaba, July 2026. $30 per 1,000 images. Rank 9–11.
FLUX.2 [max]
1226 EloBlack Forest Labs, December 2025. $70 per 1,000 images. Rank 13–21; the best FLUX variant on the board.
FLUX.2 [dev]
1200 EloBlack Forest Labs, November 2025, open weights. $12 per 1,000 images hosted, or free to self-host. Rank 30–41.
Text-to-image quality against price on the Artificial Analysis arena, 21 August 2026. Hover or tap a point for the model's Elo, price and rank range.

Two things stand out. The frontier of this chart is not one model but a curve: GPT Image 2 if you need the best, Nano Banana 2 or MAI-Image-2.5 if you want 95% of the quality at a third of the price, Nano Banana 2 Lite or Qwen-Image-3.0 for volume work at $30–34 per thousand. And open weights are further behind in images than in text: FLUX.2 [dev], the strongest open model most people can actually run, is about 170 Elo below GPT Image 2. Note that other arenas order the middle of the pack differently (CodeSOTA, for instance, puts Gemini 3.1 Flash Image first), but the identity of the leader is consistent across all of them.

Video: the board Western labs no longer own

Text-to-video has changed hands. On the Artificial Analysis text-to-video arena (with audio), Alibaba's Wan 3.0, released this month, leads at 1244 Elo, statistically tied with Google's Gemini Omni Flash at 1238. MiniMax H3 (1229), the best open-weight video model, and ByteDance's Seedance 2.0 (1221) complete a top four that is 70 points clear of everything else. The next tier is Alibaba again (Wan 2.7, HappyHorse 1.1 and 1.0), then Kling 3.0 at 1105, and only at #12 does Google's Veo 3.1 appear at 1090. OpenAI's Sora 2 was deprecated on 26 April 2026 and its API shuts down on 24 September; it no longer features.

Wan 3.0
1244 EloAlibaba, August 2026. API pricing not announced yet. Statistically tied with Gemini Omni Flash — the 95% interval is ±10.
Gemini Omni Flash
1238 EloGoogle, May 2026. $6.00 per minute of 1080p — the cheapest model in the top four and a quarter of Veo 3.1's price.
MiniMax H3
1229 EloMiniMax, July 2026, open weights. $7.80 per minute. The best open-weight video model by a wide margin.
Dreamina Seedance 2.0 720p
1221 EloByteDance Seed, March 2026. $9.07 per minute, 22,179 votes — the most-sampled model on the board.
Wan 2.7 (June build)
1154 EloAlibaba, June 2026. $9.00 per minute. The gap from here to the top four is about 70 Elo.
HappyHorse 1.1
1146 EloAlibaba ATH, June 2026. $9.90 per minute.
HappyHorse 1.0
1120 EloAlibaba ATH, April 2026. $13.20 per minute. llm-stats' smaller arena ranks this model second overall, so the ordering is sensitive to the vote pool.
Wan 2.7
1107 EloAlibaba, April 2026. $9.00 per minute.
Kling 3.0 1080p (Pro)
1105 EloKlingAI, February 2026. $20.16 per minute. llm-stats' arena puts Kling v3 first overall — the clearest disagreement between the two video boards.
Veo 3.1
1090 EloGoogle, January 2026. $24.00 per minute, the most expensive model on the board, for 12th place. Veo 3.1 Lite scores 1088 at $4.80.
grok-imagine-video
1062 EloSpaceXAI, January 2026. $4.20 per minute.
LTX-2.5 Fast
1061 EloLightricks, August 2026, open weights. $7.80 per minute. Second-best open-weight model after MiniMax H3.
Artificial Analysis text-to-video arena (with audio), Elo points behind the leader (Wan 3.0, 1244). Shorter is better. Hover or tap a bar for absolute Elo and price per minute. Read 21 August 2026; 95% intervals are ±5 to ±10 points, so the top two and positions 8–9 are ties.

Price per minute of 1080p video is as spread out as the quality. Gemini Omni Flash is $6.00 per minute, MiniMax H3 $7.80, Seedance 2.0 $9.07 and the Wan family around $9.00, while Kling 3.0 Pro is $20.16 and Veo 3.1 is $24.00 per minute for a model 150 Elo behind the leader. Veo 3.1 Lite at $4.80 scores almost the same as full Veo 3.1. One caveat on this board specifically: llm-stats' smaller video arena (1,392 votes) puts Kling v3 first with HappyHorse 1.0 second, so with fewer votes and a different prompt mix the order at the top can flip. What no arena disputes is that the Chinese labs now hold most of the top ten.

Speed and price: the other two axes

Intelligence is one axis; how fast the tokens arrive and what they cost are the other two, and at the frontier they trade off sharply. Among frontier-tier models Gemini 3.7 Flash is the outlier at roughly 390 output tokens per second, six times Claude Opus 5 (60 t/s) and five times GPT-5.6 Sol (71 t/s). Below the frontier, Inception's diffusion model Mercury 2 streams at about 1,044 t/s and Ling 3.0 Flash at 407 t/s; Artificial Analysis lists Celeris-1 faster still. Speed, remember, is measured per provider and changes week to week.

Mercury 2: 1044 t/sLing 3.0 Flash: 407 t/sGemini 3.7 Flash (high): 390 t/sGemini 3.6 Flash: 218 t/sGPT-5.6 Luna (max): 151 t/sDeepSeek V4 Flash (max): 133 t/sGemini 3.1 Pro Preview: 130 t/sGPT-5.6 Terra (max): 115 t/sGLM-5.3 (max): 93 t/sClaude Sonnet 5 (max): 87 t/sClaude Fable 5: 73 t/sGPT-5.6 Sol (max): 71 t/sGrok 4.6 (high): 68 t/sClaude Opus 5 (max): 60 t/sQwen3.8 Max: 45 t/sKimi K3 (max): 40 t/s0 t/s500 t/s1000 t/s1500 t/sMercury 2DeepSeek V4 Flash (max)Claude Fable 5Kimi K3…Output tokens per secondMercury 2: 1044 t/sLing 3.0 Flash: 407 t/sGemini 3.7 Flash (high): 390 t/sGemini 3.6 Flash: 218 t/sGPT-5.6 Luna (max): 151 t/sDeepSeek V4 Flash (max): 133 t/sGemini 3.1 Pro Preview: 130 t/sGPT-5.6 Terra (max): 115 t/sGLM-5.3 (max): 93 t/sClaude Sonnet 5 (max): 87 t/sClaude Fable 5: 73 t/sGPT-5.6 Sol (max): 71 t/sGrok 4.6 (high): 68 t/sClaude Opus 5 (max): 60 t/sQwen3.8 Max: 45 t/sKimi K3 (max): 40 t/sMercury 2Ling 3.0 FlashGemini 3.7 Flash (hig…Gemini 3.6 FlashGPT-5.6 Luna (max)DeepSeek V4 Flash (ma…Gemini 3.1 Pro PreviewGPT-5.6 Terra (max)GLM-5.3 (max)Claude Sonnet 5 (max)Claude Fable 5GPT-5.6 Sol (max)Grok 4.6 (high)Claude Opus 5 (max)Qwen3.8 MaxKimi K3 (max)0 t/s500 t/s1000 t…1500 t…Output tokens per second
Median output speed in tokens per second, Artificial Analysis, read 21 August 2026. Frontier-tier models plus the two fastest models on the board for scale. Measured on each model's first-party API.
Show the 16 numbers behind this chart
Median output speed in tokens per second, Artificial Analysis, read 21 August 2026. Frontier-tier models plus the two fastest models on the board for scale. Measured on each model's first-party API. — the numbers this chart is drawn from
CategoryOutput tokens per second
Mercury 21044 t/s
Ling 3.0 Flash407 t/s
Gemini 3.7 Flash (high)390 t/s
Gemini 3.6 Flash218 t/s
GPT-5.6 Luna (max)151 t/s
DeepSeek V4 Flash (max)133 t/s
Gemini 3.1 Pro Preview130 t/s
GPT-5.6 Terra (max)115 t/s
GLM-5.3 (max)93 t/s
Claude Sonnet 5 (max)87 t/s
Claude Fable 573 t/s
GPT-5.6 Sol (max)71 t/s
Grok 4.6 (high)68 t/s
Claude Opus 5 (max)60 t/s
Qwen3.8 Max45 t/s
Kimi K3 (max)40 t/s

On cost, Artificial Analysis reports a cost-per-task figure: what it actually spent, input plus output, to run one Intelligence Index task on each model. That is more honest than list price because reasoning models burn very different numbers of output tokens. Claude Fable 5 costs $3.14 per task and Claude Opus 5 (max) $2.34; Opus 5 at high effort drops to $1.23 for 61 points; GPT-5.6 Sol (max) is also $1.23 for 61; Grok 4.6 and Kimi K3 are $0.84 for 61 and 60; GLM-5.3 is $0.68 for 60. At the far end, GPT-5.6 Luna scores 50–52 for two to five cents a task. The interactive chart below plots every model on both axes; hover a point for its numbers and switch the x-axis to speed.

Interactive · runs in a sandboxed frameisolated

The pattern the scatter makes visible is a Pareto frontier with a knee. From GPT-5.6 Luna at a few cents to GLM-5.3 and Grok 4.6 at under a dollar, each step up in cost buys several index points. From there to Claude Opus 5 and Claude Fable 5 the last two or three points cost two to four times as much. Whether those points are worth it depends entirely on whether your task is one where the top models succeed and the rest do not; for long agentic coding runs they often are, for summarising support tickets they almost never are.

So which model should you use?

For agentic coding Claude Fable 5 or Claude Opus 5 if the task is long and failure is expensive; Claude Sonnet 5, GPT-5.6 Terra or Grok 4.5 for everyday work; GLM-5.2 or Kimi K3 if you need open weights. Judge by the private-set SWE-bench Pro score and your own repo, not by vendor numbers.
For images and video GPT Image 2 when quality or text-in-image matters, Nano Banana 2 or MAI-Image-2.5 for near-frontier quality at a third of the price, FLUX.2 [dev] to self-host. For video, try Gemini Omni Flash, MiniMax H3 or Seedance 2.0 before Veo or Kling; they score higher and cost less per minute.
For vision, volume and speed Claude Fable 5 leads the Vision Arena, but Gemini 3.6 Flash and Qwen3.8 Max are within 20 Elo at a fraction of the price. For throughput, Gemini 3.7 Flash is the frontier-quality speed outlier; for cost, GPT-5.6 Luna and DeepSeek V4 Flash deliver 50+ index points for cents.

One closing caution. Every board above moved in the last four weeks: Wan 3.0, Claude Opus 5, DeepSeek V4 Pro 0813 and LTX-2.5 all arrived in July or August. If you are reading this more than a month after 21 August 2026, reopen the sources, because at least one of these leaders will have changed.

Sources and dates

Text and reasoning, speed and cost: Artificial Analysis LLM leaderboard, read 21 August 2026. Arena Elo: LMArena via Swfte's tracker, 17 August 2026. Coding: SWE-bench Pro scores and Scale SEAL standardized results, 10 August 2026, and the Artificial Analysis Coding Index via modelgrep. Vision: LMArena Vision Arena, 6 August 2026. Images: Artificial Analysis text-to-image leaderboard and llm-stats image arena, 21 August 2026. Video: Artificial Analysis text-to-video leaderboard and llm-stats video arena, 21 August 2026.

Share

Related

Claude Sonnet 5.5: Opus-Level Scores at Half the Price, and Where It Still Falls Short
Gemini 4 Argon: Google's New Frontier Model, the Benchmarks, the Price and What It Can Do
Claude Opus 5.5: Every Benchmark, From Anthropic, Cursor, Artificial Analysis, Sonar and ARC Prize