SShogo Bench
Agent Index v1.0 · updated Nov 20, 2026

Which model should run your agents?

The Shogo Agent Index ranks models only on agentic work: fixing real repos, staying on task for hours, producing professional deliverables and driving a computer. 17 benchmarks, weighted, every score sourced, with the price an agent actually pays.

Top 25 models for agents

All 92 ranked models →
ModelAgent IndexCodingLong-hznKnow. workComp. useSec/ctxMeasuredAgent $/MContext
Claude Fable 5.2
AnthropicSupported
76.4
71.7–81.2
867867776532%$3.211M
Gemini 4 Carbon
GoogleSupported
73.6
67.2–80
828063686327%$0.841M
Claude Opus 5.5
AnthropicSupported
70.9
64.8–76.9
817658666232%$1.691M
GPT-6 Astra
OpenAISupported
70.9
66.5–75.4
808056676338%$4.621.1M
GPT-6.1 Sol
OpenAIEstimated
70.0
62.2–77.7
807854666218%$0.841.1M
Gemini 4 Argon
GoogleSupported
69.5
62.9–76.2
777859616132%$0.841M
Gemini 4 Hydrogen
GoogleEstimated
69.4
61.8–77.1
787858616123%$0.631M
GPT-6.1 Terra
OpenAIEstimated
69.0
60.5–77.5
797456636214%$0.421.1M
Claude Sonnet 5.5
AnthropicEstimated
69.0
61.3–76.7
797557616123%$0.841M
Claude Fable 5.1
AnthropicSupported
68.2
62.7–73.7
807155626136%$4.011M
Claude Opus 5
AnthropicSupported
65.4
59.8–71
747452565939%$2.311M
Claude Fable 5
AnthropicSupported
64.1
59–69.3
746952555938%$4.621M
GPT-6 Sol
OpenAIEstimated
63.6
54.2–73
727648545914%$0.921.1M
GPT-5.6 Sol
OpenAISupported
61.8
57.1–66.5
717246535843%$1.851.1M
Muse Wonder
MetaEstimated
60.9
51.9–69.9
687050505722%$1.691M
GPT-5.6 Terra
OpenAIEstimated
58.8
50.7–66.9
666947485624%$1.021.1M
DeepSeek V4.1 Flash 1017
DeepSeek· openEstimated
57.6
44.4–70.8
636649455412%$0.0501M
Claude Haiku 5.5
AnthropicEstimated
57.5
48.4–66.5
676246475523%$0.0501M
GPT-5.5
OpenAISupported
56.7
52.5–61
636746435560%$2.551.1M
Claude Opus 4.8
AnthropicSupported
56.3
50.9–61.8
636643475544%$2.311M
Gemini 3.8 Flash
GoogleSupported
56.2
48.7–63.7
636447445334%$0.351M
Grok 4.6
xAISupported
56.2
48.3–64
616847425229%$0.98500K
Gemini 3.7 Flash
GoogleEstimated
55.7
44.2–67.1
616448435324%$0.351M
Inkling 2 Small
thinkingmachines· openEstimated
55.2
42.3–68.2
626642445413%$0.141M
GLM-5.3
Z.ai (Zhipu)· openSupported
55.2
47.7–62.7
626843405133%$0.611M

Tick models to compare them side by side. Pillar scores in white include measured results; grey ones are fully estimated. Shaded band = 95% range.

Agent performance per dollar

Agents resend a long, mostly cached context every turn. We price each model at 20 input tokens per output token with 85% cache hits, which is closer to an agent’s bill than the usual 3:1 blend.

1020304050607080$0.01$0.1$1$10Agent-effective price, $ per 1M tokens (log scale) →Claude Fable 5.2: 76.4 at $3.21/MClaude Fable 5.2Gemini 4 Carbon: 73.6 at $0.84/MGemini 4 CarbonClaude Opus 5.5: 70.9 at $1.69/MGPT-6 Astra: 70.9 at $4.62/MGPT-6.1 Sol: 70 at $0.84/MGemini 4 Argon: 69.5 at $0.84/MGemini 4 Hydrogen: 69.4 at $0.63/MGemini 4 HydrogenGPT-6.1 Terra: 69 at $0.42/MGPT-6.1 TerraClaude Sonnet 5.5: 69 at $0.84/MClaude Fable 5.1: 68.2 at $4.01/MClaude Opus 5: 65.4 at $2.31/MClaude Fable 5: 64.1 at $4.62/MGPT-6 Sol: 63.6 at $0.92/MGPT-5.6 Sol: 61.8 at $1.85/MMuse Wonder: 60.9 at $1.69/MGPT-5.6 Terra: 58.8 at $1.02/MDeepSeek V4.1 Flash 1017: 57.6 at $0.05/MDeepSeek V4.1 Flash 1017Claude Haiku 5.5: 57.5 at $0.05/MGPT-5.5: 56.7 at $2.55/MClaude Opus 4.8: 56.3 at $2.31/MGemini 3.8 Flash: 56.2 at $0.35/MGrok 4.6: 56.2 at $0.98/MGemini 3.7 Flash: 55.7 at $0.35/MInkling 2 Small: 55.2 at $0.14/MGLM-5.3: 55.2 at $0.61/MKimi K3: 54.9 at $1.39/MGrok 4.7: 54.8 at $0.98/MMuse Spark 1.3: 54.7 at $0.03/MMuse Spark 1.3Claude Opus 4.7: 54.6 at $2.31/MClaude Opus 4.6: 53.2 at $2.31/MClaude Sonnet 5: 52.4 at $0.92/MGPT-5.4: 51.6 at $1.27/MGPT-5.6 Luna: 51.4 at $0.1/MDeepSeek V4.1 Flash: 51.2 at $0.12/MGrok 4.5: 49.4 at $0.81/MGPT-5.3 Codex: 48.4 at $1.06/MGLM-5.3-Flash: 48.3 at $0.16/MGemini 3.6 Flash: 48.3 at $0.35/MMuse Spark 1.1: 48.3 at $0.5/MMuse Spark 1.2: 48.2 at $0.5/MGLM-5.2: 47.3 at $0.61/MGemini 3.5 Flash: 47.3 at $0.76/MGPT-5.2: 46.4 at $0.96/MGemini 3.1 Pro Preview: 46 at $1.02/MClaude Opus 4.5 (latest): 46 at $2.31/MClaude Sonnet 4.6: 44.9 at $1.39/MGemini 3 Pro Preview: 44.7 at $2.48/MGrok 4.20 (Non-Reasoning): 43.9 at $0.46/MKimi K2.6: 43.7 at $0.46/MGemini 3 Flash Preview: 43.7 at $0.21/MGLM-5.1: 42.5 at $0.46/MKimi K2.7 Code: 42.1 at $0.96/MDeepSeek V4 Pro: 40.7 at $0.5/MQwen3.6 Max Preview: 40.5 at $1.08/MGPT-5.1 Codex Max: 39.6 at $0.67/MGPT-5.1: 38.8 at $0.69/MClaude Sonnet 4.5 (latest): 38.7 at $1.39/MGPT-5: 37.8 at $0.76/MMiniMax-M3: 36.7 at $0.15/MGLM-5: 36.5 at $0.33/MKimi K2.5: 36.5 at $0.31/MInkling: 35.8 at $0.79/MQwen3.5 Plus: 32.7 at $0.2/MClaude Opus 4.1 (latest): 32.1 at $5.96/Mo3: 31.4 at $0.97/MGPT-5 Mini: 30.4 at $0.15/MKimi K2 Thinking: 30.3 at $0.35/MGLM-4.7: 30 at $0.24/MDeepSeek V3.2: 30 at $0.16/Mo4-mini: 29.8 at $0.53/MClaude-Opus-4: 27.9 at $5.96/MClaude Haiku 4.5: 27.1 at $0.46/MQwen3 Coder 480B A35B: 26.4 at $0.17/MClaude Sonnet 4: 26.2 at $1.2/MQwen3 Max: 25.7 at $0.18/MKimi K2 0905: 25.3 at $1.16/MGemini 2.5 Pro: 24.5 at $0.76/MClaude 3.7 Sonnet: 24.2 at $1.2/MMiniMax-M2: 23.9 at $0.4/MGemini 2.5 Flash: 23.6 at $0.19/Mo3-mini: 23.3 at $1.13/MGPT OSS 120B: 22.1 at $0.04/MDeepSeek-R1: 21.1 at $0.71/Mo1: 16.7 at $15.9/MClaude-Sonnet-3.5: 14.8 at $1.2/MGPT-4o (2024-08-06): 12.2 at $1.68/M
Dashed line: the price–performance frontier (no cheaper model scores higher). Click a dot to open the model.

The five pillars

  1. Agentic coding35%

    Fixing real repos, working in a terminal and shipping multi-file changes without a human in the loop.

  2. Long-horizon autonomy20%

    Staying on task for hours: long software tasks, running a business, training a model end to end.

  3. Knowledge-work agents25%

    Professional deliverables graded by experts: banking, consulting, law, research reports, paid freelance jobs.

  4. Computer use10%

    Operating a real desktop and browser through screenshots, mouse and keyboard.

  5. Security & context learning10%

    Capture-the-flag security tasks and learning new rules from long context on the fly.

Put the winner to work

Shogo runs your agents on whichever model fits the job, and routes each step to the cheapest model that can do it. Same models as this leaderboard, no API keys to juggle.

Build an agent in Shogo

Latest changes

  • Launched Shogo Agent Index v1.0: 92 ranked models, 564 models tracked, 46 benchmarks.
Changelog →