Skip to content

Benchmarks · updated daily

AI benchmarks dashboard

A daily read of the benchmarks that still separate frontier AI models: who leads where, what changed, and links to every source.

Benchmarks tracked
65
Active · near-sat. · saturated
46 · 7 · 6
Data sources healthy
4/4
Leaderboards
19
Last updated

What changed

  • Page updatedArtificial Analysis Intelligence Index v4.3.2: the public page changed since the last check.
  • Page updatedGDPval / GDPval-AA v2.1: the public page changed since the last check.
  • Page updatedIFBench: the public page changed since the last check.
  • Page updatedImage arenas (Arena T2I, AA Image Arena): the public page changed since the last check.
  • Page updatedEpoch Capabilities Index (ECI) + Benchmarking Hub: the public page changed since the last check.
  • Page updatedEpoch Capabilities Index (ECI) + Benchmarking Hub: the public page changed since the last check.
  • Page updatedFrontierMath Tiers 1–4 (v2): the public page changed since the last check.
  • Page updatedSWE-bench Pro (Scale): the public page changed since the last check.
  • + 22 more in data/benchmarks.json

Source health

Daily link check: 76 of 82 reference pages reachable, 5 block automated checks, 34 changed since the last check.

Leaderboards

Top 5 per board, taken only from the named source and date. Compare models within one board, never across boards: harness and split change scores.

Aggregators

Epoch Capabilities Index (ECI)

  1. 1Claude Fable 5.2Anthropic173.4
  2. 2Gemini 4 CarbonGoogle DeepMind169.8
  3. 3GPT-6 AstraOpenAI169.0
  4. 4Claude Opus 5.5Anthropic168.5
  5. 5GPT-6.1 SolOpenAI168.5

ECI · higher is better · as of 2026-11-17
Source: Epoch AI, Benchmarking Hub (CC BY 4.0)

Aggregators

Arena Text (overall)

  1. 1Gemini 4 Carbon (High)google1538
  2. 2Gemini 4 Argon (High)google1530
  3. 3Claude Opus 5.5 (High)anthropic1514
  4. 4Claude Fable 5.1 (Max)anthropic1512
  5. 5Claude Opus 5 (Max)anthropic1506

Arena score · higher is better · as of 2026-11-11
Source: Arena / LMArena leaderboard-dataset (CC BY 4.0)

General intelligence & reasoning

Humanity's Last Exam

  1. 1GPT-6 AstraOpenAI54.8%
  2. 2Claude Fable 5.1 (xhigh)Anthropic46.5%
  3. 3Gemini 3.1 Pro PreviewGoogle DeepMind46.4%
  4. 4Claude Sonnet 5.5 (xhigh)Anthropic46.1%
  5. 5Gemini 3.8 FlashGoogle DeepMind44.5%

% · higher is better · as of 2026-11-17 · newest entry 2026-10-07
Source: Epoch AI, Benchmarking Hub (CC BY 4.0)

General intelligence & reasoning

ARC-AGI-2

  1. 1GPT-6 Astra (max)OpenAI95.0%
  2. 2GPT-6.1 Sol (max)OpenAI94.2%
  3. 3Gemini 4 Carbon (medium)Google DeepMind94.2%
  4. 4Claude Fable 5.2 (xhigh)Anthropic93.8%
  5. 5Claude Opus 5.5 (high)Anthropic93.3%

% · higher is better · as of 2026-11-17 · newest entry 2026-11-10
Source: Epoch AI, Benchmarking Hub (CC BY 4.0)

Math

FrontierMath Tier 4 (v2)

  1. 1Claude Fable 5.2 (max)Anthropic100.0%
  2. 2GPT-6.1 Sol (max)OpenAI100.0%
  3. 3GPT-6 Astra (medium)OpenAI97.6%
  4. 4Gemini 4 Carbon (high)Google DeepMind95.1%
  5. 5Claude Opus 5.5 (max)Anthropic95.0%

% · higher is better · as of 2026-11-17 · newest entry 2026-11-06
Source: Epoch AI, Benchmarking Hub (CC BY 4.0)

Coding / software engineering

Terminal-Bench 2.0

  1. 1GPT-5.5NexAU-AHE · OpenAI84.7%
  2. 2GPT-5.4ForgeCode · OpenAI81.8%
  3. 3Gemini 3.1 Pro PreviewTongAgents · Google DeepMind80.2%
  4. 4Claude Opus 4.7WOZCODE · Anthropic80.2%
  5. 5Claude Opus 4.6ForgeCode · Anthropic79.8%

% · higher is better · as of 2026-11-17 · newest entry 2026-05-14
Source: Epoch AI, Benchmarking Hub (CC BY 4.0)

Agentic, computer use & long-horizon

OSWorld 2.0

  1. 1Claude Opus 5 (max)Anthropic31.4%
  2. 2GPT-5.6 Sol (max)OpenAI27.3%
  3. 3Claude Opus 4.8Anthropic20.6%
  4. 4Claude Opus 4.7 (max)Anthropic18.2%
  5. 5GPT-5.5 (xhigh)OpenAI13.0%

% · higher is better · as of 2026-11-17 · newest entry 2026-07-24
Source: Epoch AI, Benchmarking Hub (CC BY 4.0)

Agentic, computer use & long-horizon

METR time horizon (50%)

  1. 1Claude Mythos Preview (Early)Anthropic1045 min
  2. 2Claude Opus 4.6Anthropic719 min
  3. 3Gemini 3.1 Pro PreviewGoogle DeepMind384 min
  4. 4GPT-5.2 (high)OpenAI352 min
  5. 5GPT-5.3 CodexOpenAI350 min

min · higher is better · as of 2026-11-17 · newest entry 2026-04-07
Source: Epoch AI, Benchmarking Hub (CC BY 4.0)

Multilingual (Swedish/Nordic)

EuroEval Swedish (generative)

  1. 1gpt-6.1-sol (zero-shot, val)1.22
  2. 2gpt-6-astra (zero-shot, val)1.24
  3. 3gemini/gemini-3.7-flash (zero-shot, val)1.28
  4. 4zai-org/GLM-5.3-Flash (val)1.30
  5. 5gemini/gemini-3.6-flash (zero-shot, val)1.33

rank score · lower is better · as of 2026-11-15
Source: EuroEval (euroeval.com)

More leaderboards (10)

Coding / software engineering

Arena WebDev

  1. 1Claude Fable 5.2 (xHigh)anthropic1826
  2. 2Claude Opus 5.5 (Max)anthropic1816
  3. 3Claude Sonnet 5.5 (xHigh)anthropic1772
  4. 4Gemini 4 Carbon (High)google1768
  5. 5GPT-6 Astra (Max)openai1763

Arena score · higher is better · as of 2026-11-15
Source: Arena / LMArena leaderboard-dataset (CC BY 4.0)

Coding / software engineering

SWE-bench Multilingual

  1. 1Gemini 3 FlashSWE-agent72.7%
  2. 2Claude 4.6 OpusSWE-agent72.0%
  3. 3Claude 4.5 OpusSWE-agent70.7%
  4. 4GLM 5SWE-agent69.7%
  5. 5Gemini 3 ProSWE-agent68.7%

% resolved · higher is better · as of 2026-02-20
Source: SWE-bench team (swebench.com)

Coding / software engineering

SWE-bench Verified (Epoch runs)

  1. 1Claude Opus 4.7 (max)Anthropic83.5%
  2. 2GPT-5.5 (xhigh)OpenAI80.6%
  3. 3Gemini 3.5 FlashGoogle DeepMind79.3%
  4. 4GLM-5.2Z.ai (Zhipu AI)78.7%
  5. 5Claude Opus 4.6 (no thinking)Anthropic78.7%

% · higher is better · as of 2026-11-17 · newest entry 2026-06-25
Source: Epoch AI, Benchmarking Hub (CC BY 4.0)

Coding / software engineering

SWE-bench Verified (official board)

  1. 1Sonar Foundation Agent + Claude 4.5 OpusSonar79.2%
  2. 2live-SWE-agent + Claude 4.5 Opus medium (20251101)UIUC79.2%
  3. 3TRAE + Doubao-Seed-CodeByteDance78.8%
  4. 4live-SWE-agent + Gemini 3 Pro Preview (2025-11-18)UIUC77.4%
  5. 5EPAM AI/Run Developer Agent v20250719 + Claude 4 SonnetEPAM Systems, Inc.76.8%

% resolved · higher is better · as of 2026-02-26
Source: SWE-bench team (swebench.com)

Cybersecurity

Cybench (unguided)

  1. 1Claude Opus 4.6Anthropic93.0%
  2. 2Claude Opus 4.5Anthropic82.0%
  3. 3Claude Sonnet 4.5Anthropic60.0%
  4. 4Grok 4xAI43.0%
  5. 5Claude Opus 4.1Anthropic42.0%

% · higher is better · as of 2026-11-17 · newest entry 2026-02-05
Source: Epoch AI, Benchmarking Hub (CC BY 4.0)

Design, UI & multimodal

Arena Vision

  1. 1Claude Fable 5 (High)anthropic1326
  2. 2Claude Opus 5.5 (High)anthropic1324
  3. 3Claude Fable 5.1 (Max)anthropic1322
  4. 4Claude Opus 4.7anthropic1319
  5. 5Claude Opus 5 (High)anthropic1318

Arena score · higher is better · as of 2026-11-09
Source: Arena / LMArena leaderboard-dataset (CC BY 4.0)

Knowledge, factuality & hallucination

SimpleQA Verified

  1. 1Claude Fable 5.2 (max)Anthropic94.0%
  2. 2Gemini 4 Carbon (high)Google DeepMind77.7%
  3. 3Gemini 4 Argon (high)Google DeepMind76.0%
  4. 4GPT-6 Astra (max)OpenAI75.6%
  5. 5GPT-6.1 Sol (max)OpenAI73.9%

% · higher is better · as of 2026-11-17 · newest entry 2026-11-06
Source: Epoch AI, Benchmarking Hub (CC BY 4.0)

Economic & real-world work value

APEX-Agents

  1. 1Gemini 4 CarbonGoogle DeepMind84.2%
  2. 2Gemini 4 ArgonGoogle DeepMind82.4%
  3. 3Claude Fable 5.2Anthropic80.1%
  4. 4Claude Sonnet 5.5 (max)Anthropic75.5%
  5. 5Claude Opus 5.5 (max)Anthropic73.5%

% · higher is better · as of 2026-11-17 · newest entry 2026-11-10
Source: Epoch AI, Benchmarking Hub (CC BY 4.0)

Economic & real-world work value

GDPval (win rate)

  1. 1GPT-5.2 (none)OpenAI49.7%
  2. 2Claude Opus 4.5 (no thinking)Anthropic45.5%
  3. 3Claude Opus 4.1Anthropic43.6%
  4. 4Claude Sonnet 4.5 (no thinking)Anthropic42.5%
  5. 5Gemini 3 Pro PreviewGoogle DeepMind40.3%

% · higher is better · as of 2026-11-17 · newest entry 2025-12-11
Source: Epoch AI, Benchmarking Hub (CC BY 4.0)

Economic & real-world work value

Remote Labor Index

  1. 1Claude Fable 5.2Anthropic41.3%
  2. 2Claude Opus 5.5Anthropic21.2%
  3. 3GPT-6 AstraOpenAI20.8%
  4. 4Claude Fable 5.1Anthropic17.9%
  5. 5Gemini 4 ArgonGoogle DeepMind17.1%

% · higher is better · as of 2026-11-17 · newest entry 2026-11-09
Source: Epoch AI, Benchmarking Hub (CC BY 4.0)

Status by category

  • Aggregators6
  • General intelligence & reasoning6
  • Math4
  • Coding / software engineering9
  • Agentic, computer use & long-horizon12
  • Cybersecurity9
  • Design, UI & multimodal7
  • Knowledge, factuality & hallucination2
  • Long context2
  • Instruction following2
  • Safety, alignment & robustness5
  • Economic & real-world work value5
  • Multilingual (Swedish/Nordic)2

Active / newNear-saturatedSaturatedFrozen / static / other

References

Every benchmark in the catalogue, grouped by category. Links open the official page; third-party marks pages not run by the benchmark authors.

Aggregators6
General intelligence & reasoning6
Math4
Coding / software engineering9
Agentic, computer use & long-horizon12
Cybersecurity9
Design, UI & multimodal7
Knowledge, factuality & hallucination2
Long context2
Instruction following2
Safety, alignment & robustness5
Economic & real-world work value5
Multilingual (Swedish/Nordic)2

Catalogue: Hieg research, ranked by relevance (9 Oct 2026). Numbers are reproduced only from sources that publish them openly, with attribution and date; nothing is estimated. Raw data (JSON).

Stockholm · Europe/Stockholm

Clarity for AI, security, and systems that stay yours.

Hieg is a consulting practice for organisations that need sharp judgment on artificial intelligence, cybersecurity, and local sovereign infrastructure — without the slide-deck theatre.

Focus

Where the work lands

Five overlapping domains. Engagements often cut across more than one.

AI

Model selection, evaluation, and integration into real workflows. Grounded advice on what is ready, what is noisy, and what will actually ship.

Cybersecurity

Threat-aware design, hardening, and review for systems that handle sensitive data — including AI pipelines and edge deployments.

Consulting

Short, high-signal engagements: architecture reviews, build-vs-buy decisions, and executive briefings that survive contact with engineering.

Local sovereign AI

On-prem and private-cloud models when data residency, latency, or policy demand control. From hardware sizing to operational runbooks.

Robotics

Practical bridging of perception, control, and AI — for teams shipping robots into the physical world, not demos that only work on stage.

Approach

Fewer slides. More durable decisions.

Hieg is run by Gustavo A from Stockholm. The practice exists for teams that already know the generic playbook and need someone who can argue trade-offs, challenge vendors, and stay accountable when the system goes live.

Work is scoped tightly. Deliverables are written for the people who build and operate — not for a marketing archive.

  1. Sovereignty by default

    Prefer architectures that keep critical inference, logs, and keys under your control when the risk profile demands it.

  2. Security as design, not garnish

    Threat models early. Least privilege, measurable surfaces, and no “we’ll harden it later” as a roadmap item.

  3. Evidence over fashion

    Benchmarks, failure modes, and operational cost beat trend reports. If it cannot be measured, it is not a recommendation.

Contact

Tell me what you are trying to decide.

Describe the problem, constraints, and timeline. Serious enquiries get a reply; vague mass outreach does not.

Prefer a short brief over a long pitch. If NDA is required before details, say so in the message.
No newsletter. No shared inbox spam.