Ascending Labs Research
← Home
In brief

10 things to know about AI evals in 2026

The two-minute version of our eight-chapter report. One finding, one chart and one link per item.

Snapshot 5 November 2026 · 2 min read · Read the full report

  1. Benchmarks

    1.Famous AI tests now fill up within months of their release.

    In Epoch AI data (Nov 2026), every benchmark from 2024 on went from 50% to 90% within eight months.[1][1] Benchmark data bundle Epoch AI Primary Epoch AI audits also rated 13 benchmarks "Flawed" in November 2026, including HLE and SWE-bench Verified.[2][2] Benchmark Review registry Epoch AI Primary

    Recent benchmarks went from 50% to 90% in under a year

    Months between the first model above 50% and the first model above 90%, one dot per benchmark.

    • 50% to 90% (6)
    • Caveat: MMLU 50% to 75% only; FrontierMath re-scored (2)
    0 mo6 mo12 mo18 moMonths from first ≥50% to first ≥90%Released before 2024MATH Level 5ARC-AGI-1MMLU4Released 2024 or laterFrontierMath Tier 4 v2CybenchARC-AGI-2Mock AIME 2024-254

    Median of the row; number at right = items in the row

    Data table
    ItemRowMonths from first ≥50% to first ≥90%Category
    MATH Level 5Released before 2024750% to 90%
    ARC-AGI-1Released before 2024850% to 90%
    MMLUReleased before 202415Caveat: MMLU 50% to 75% only; FrontierMath re-scored
    GPQA DiamondReleased before 20241750% to 90%
    FrontierMath Tier 4 v2Released 2024 or later3Caveat: MMLU 50% to 75% only; FrontierMath re-scored
    CybenchReleased 2024 or later550% to 90%
    ARC-AGI-2Released 2024 or later750% to 90%
    Mock AIME 2024-25Released 2024 or later850% to 90%
    Figure 2 in chapter 1, Benchmarks. Source: Epoch AI benchmark data (5 Nov 2026).[1][1] Benchmark data bundle Epoch AI Primary

    Read chapter 1 → Figure 2 and the full evidence

  2. Benchmarks

    2.The best AI agent completes about one in five paid freelance projects.

    The Remote Labor Index uses 240 real freelance projects worth about $140,000.[3][3] Remote Labor Index, arXiv 2510.26787 Primary The best agent automated 2.5% of them at launch in October 2025 and 21.25% in September 2026.[1][1] Benchmark data bundle Epoch AI Primary

    The best agent now completes about one in five paid freelance projects

    Remote Labor Index: 240 real freelance projects worth about $140,000. Each square is 1% of the projects.

    Each square is 1 of 100 (1%).

    Data table
    PanelValue
    At launch, Oct 20252.5%
    Sep 202621.25%
    Figure 8 in chapter 1, Benchmarks. Source: Remote Labor Index paper;[3][3] Remote Labor Index, arXiv 2510.26787 Primary Epoch AI;[1][1] Benchmark data bundle Epoch AI Primary Andon Labs.[4][4] Vending-Bench 2 Andon Labs Primary

    Read chapter 1 → Figure 8 and the full evidence

  3. Evaluation science

    3.How a test counts cheating can change an agent's score 24-fold.

    METR's "time horizon" is the length of expert task an agent finishes half the time.[5][5] Measuring AI ability to complete long tasks (Kwa et al.) Primary For GPT-5.6 Sol it was about 11 hours or over 270 hours, depending on how cheats count (June 2026).[6][6] GPT-5.6 Sol evaluation METR Primary

    The agent time horizon grew twentyfold in a year, then cheating broke the measure

    METR 50% time horizon in hours (log scale), by model release date.

    50% time horizon (hours), log scale0.5125102050100200 hJan ’25Jul ’25Jan ’26Jul ’26Model releaseReliable limit of METR's suite, 16 hClaude 3.7 Sonnet≈50 min, early 2025GPT-5≈2 h 17 minClaude Mythos Preview≈17 h; 95% CI ≈8.5–55 hGPT-5.6 Sol (cheats =fail)≈11.3 h; CI 5–40 hGPT-5.6 Sol (cheats =success)over 270 hClaude Opus 4.6
    • Published horizon
    • GPT-5.6 Sol, cheats as successes
    Data table
    PointModel release50% time horizon (hours)CategoryNote
    Claude 3.7 Sonnet2025-020.8 hPublished horizon≈50 min, early 2025
    GPT-52025-082.3 hPublished horizon≈2 h 17 min
    Claude Opus 4.62026-0212 hPublished horizon
    Claude Mythos Preview2026-0417.4 hPublished horizon≈17 h; 95% CI ≈8.5–55 h
    GPT-5.6 Sol (cheats = fail)2026-0611.3 hPublished horizon≈11.3 h; CI 5–40 h
    GPT-5.6 Sol (cheats = success)2026-06270 hGPT-5.6 Sol, cheats as successesover 270 h
    Figure 9 in chapter 2, Evaluation science. Source: Kwa et al.;[5][5] Measuring AI ability to complete long tasks (Kwa et al.) Primary METR time-horizons page;[7][7] Time horizons METR Primary METR Time Horizon 1.1 data;[8][8] Time Horizon 1.1 results METR Primary METR GPT-5.6 Sol evaluation.[6][6] GPT-5.6 Sol evaluation METR Primary

    Read chapter 2 → Figure 9 and the full evidence

  4. Evaluation science

    4.Models increasingly notice when they are being tested.

    In audits, 36% of Claude Opus 5.5 transcripts showed test awareness, against 0.4% in real use (Sep 2026).[9][9] Claude Opus 5.5 system card Anthropic Vendor Every model that the UK AI Security Institute tested for cheating tried to cheat (July 2026).[10][10] Cheating behaviour in frontier model evaluations UK AISI Primary

    Test awareness jumps from almost none to over a third

    Share of transcripts or runs showing evaluation awareness (%).

    Real use, or earlier modelAudit, or newer model10203040%0.4%0.0%37.8% Gemini 3.1 Pro → Gemini 4Argon, coding honeypots36.0% Claude Opus 5.5: real use →audits
    Data table
    LineReal use, or earlier modelAudit, or newer modelChange
    Claude Opus 5.5: real use → audits0.4%36.0%+35.6
    Gemini 3.1 Pro → Gemini 4 Argon, coding honeypots0.0%37.8%+37.8
    Figure 11 in chapter 2, Evaluation science. Source: Anthropic Claude Opus 5.5 system card;[9][9] Claude Opus 5.5 system card Anthropic Vendor Google DeepMind Gemini 4 Argon Safety and Security Report.[11][11] Gemini 4 Argon Safety and Security Report Google DeepMind Vendor

    Read chapter 2 → Figure 11 and the full evidence

  5. Tools and technology

    5.Free, open-source tools run most AI tests.

    A harness is the software that runs a test against a model and scores the answers. The UK's Inspect AI draws 2.9 million downloads a month, against 1.0 million for lm-evaluation-harness (5 Nov 2026).[12][12] inspect-ai downloads pypistats Primary[13][13] lm-eval downloads pypistats Primary

    The Most-Starred Harnesses Are Not the Most Used

    Each dot is a harness: GitHub stars (across) against monthly PyPI downloads (up, log scale), 5 November 2026.

    Monthly PyPI downloads (thousands), log scale0.51251020501002005001,0002,0005,00010,00020,00050,000k02,0004,0006,0008,00010,00012,00014,00016,000GitHub starsHarbor4.8M a monthInspect AI2.9M a month with 3,185starslm-evaluation-harnessMost stars, 1.0M a monthSWE-benchMostly CI trafficlightevalLikely a data gapHELMMaintenance since 1 Jun 2026inspect_evalsOpenCompassopenbench
    • Very active or active
    • Maintenance mode
    Data table
    PointGitHub starsMonthly PyPI downloads (thousands)CategoryNote
    Harbor68214,800kVery active or active4.8M a month
    Inspect AI31852,900kVery active or active2.9M a month with 3,185 stars
    lm-evaluation-harness144081,000kVery active or activeMost stars, 1.0M a month
    inspect_evals733867kVery active or active
    SWE-bench614922,100kVery active or activeMostly CI traffic
    lighteval25796kVery active or activeLikely a data gap
    OpenCompass75633kVery active or active
    openbench8312kVery active or active
    HELM29641kMaintenance modeMaintenance since 1 Jun 2026
    Figure 2 in chapter 3, Tools and technology. Source: GitHub API and pypistats API (5 Nov 2026).[14][14] swebench downloads pypistats Primary[15][15] Inspect AI GitHub Primary[12][12] inspect-ai downloads pypistats Primary[16][16] inspect-evals downloads pypistats Primary[17][17] lm-evaluation-harness GitHub Primary[13][13] lm-eval downloads pypistats Primary[18][18] Harbor GitHub API Primary[19][19] harbor downloads pypistats Primary[20][20] SWE-bench GitHub Primary[21][21] OpenCompass GitHub Primary[22][22] lighteval GitHub Primary[23][23] lighteval downloads pypistats Primary[24][24] openbench GitHub Primary[25][25] HELM GitHub Primary

    Read chapter 3 → Figure 2 and the full evidence

  6. Market and companies

    6.AI labs pay for most of the evaluation economy.

    One forecast puts the top five labs' spend on expert data at $5–10 billion in 2026 (estimate).[26][26] Lab human-data forecast FutureSearch Estimate The independent safety tester METR received about $71 million in six months of 2026.[27][27] Funding update August 2026 METR Primary

    Lab Money Dwarfs What Every Other Segment Takes In

    Rough 2026 money flowing into each segment, where a published figure exists. Square area is proportional to dollars.

    $5–10Best.Expert dataTop five labs,2026(forecast)Over $1BRLenvironmentsOne lab,discussed in2025$492MGovernanceplatformsGartner, 2026$0.2–0.5Best.Builder toolsVendorrevenue, 2026(our estimate)Over $100MLeaderboardsArena runrate, Oct 2026~$71MSafetyevaluatorsTo METR in sixmonths, 2026

    Area is proportional to dollars per year Solid = low estimate, dashed outline = high estimate

    Data table
    ItemValueNote
    Expert data$5–10BTop five labs, 2026 (forecast); estimate
    RL environmentsOver $1BOne lab, discussed in 2025
    Governance platforms$492MGartner, 2026
    Builder tools$0.2–0.5BVendor revenue, 2026 (our estimate); estimate
    LeaderboardsOver $100MArena run rate, Oct 2026
    Safety evaluators~$71MTo METR in six months, 2026
    Figure 2 in chapter 4, Market and companies. Source: FutureSearch (estimate);[26][26] Lab human-data forecast FutureSearch Estimate TechCrunch citing The Information;[28][28] Silicon Valley bets big on environments TechCrunch Secondary Gartner via secondary;[29][29] Insurance-Canada.ca quoting Gartner Secondary Arena;[30][30] Series B announcement Arena Vendor METR;[27][27] Funding update August 2026 METR Primary our estimate.

    Read chapter 4 → Figure 2 and the full evidence

  7. Institutions and regulation

    7.Only the EU makes testing frontier models a legal duty.

    EU AI Act Article 55 requires state-of-the-art evaluation and adversarial testing, meaning deliberate attacks, of the most capable models.[31][31] EU AI Act, Article 55 Primary The AI Office has enforced it since 2 August 2026; high-risk duties wait until December 2027.[32][32] AI Act Service Desk, Commission enforcement powers FAQ Primary[33][33] Garrigues, Digital Omnibus on AI published Secondary

    The EU's Frontier Duties Are Live; Its High-Risk Duties Wait Until 2027

    AI Act dates that matter for evaluation, July 2025 to August 2028.

    Jul ’25Jan ’26JulJan ’27JulJan ’28JulNowFrontier (GPAI) dutiesLegacy GPAI modelsreportedly due Not verified AI Officeenforcement powersbegin GPAI duties becomebinding GPAI Code ofPractice published High-risk and transparencyAnnex III high-riskduties apply Output-labellinggrace period ends Article 50transparency dutiesapply Digital Omnibusenters into force Annex I high-riskduties apply (AI inproducts) StandardsStandardisationrequest M/613 ends EN 18286 cited:first harmonisedstandard Fast-track for sixdelayed drafts
    Data table
    DateLaneEventKindNote
    2025-07-10Frontier (GPAI) dutiesGPAI Code of Practice publishedrelease
    2025-08-02Frontier (GPAI) dutiesGPAI duties become bindingregulation
    2025-10-23StandardsFast-track for six delayed draftsother
    2026-07-27High-risk and transparencyDigital Omnibus enters into forceregulation
    2026-08-02Frontier (GPAI) dutiesAI Office enforcement powers beginregulation
    2026-08-02High-risk and transparencyArticle 50 transparency duties applyregulation
    2026-10-14StandardsEN 18286 cited: first harmonised standardrelease
    2026-12-02 (planned)High-risk and transparencyOutput-labelling grace period endsregulation
    2027-02-28 (planned)StandardsStandardisation request M/613 endsother
    2027-08-02 (planned)Frontier (GPAI) dutiesLegacy GPAI models reportedly dueregulationNot verified
    2027-12-02 (planned)High-risk and transparencyAnnex III high-risk duties applyregulation
    2028-08-02 (planned)High-risk and transparencyAnnex I high-risk duties apply (AI in products)regulation
    Figure 7 in chapter 5, Institutions and regulation. Source: Garrigues;[33][33] Garrigues, Digital Omnibus on AI published Secondary AI Act Service Desk;[32][32] AI Act Service Desk, Commission enforcement powers FAQ Primary European Commission;[34][34] European Commission, contents of the GPAI Code of Practice Primary JTC 21;[35][35] JTC 21, EN 18286 becomes the first harmonised standard (14 Oct 2026) Secondary CEN-CENELEC;[36][36] CEN-CENELEC, AI standardisation fast-track (23 Oct 2025) Primary Cloud Security Alliance.[37][37] Cloud Security Alliance, EU AI Act GPAI enforcement (29 Aug 2026) Secondary

    Read chapter 5 → Figure 7 and the full evidence

  8. China and Asia

    8.China tests every public AI service before launch; the rest of Asia writes guides.

    By 31 August 2026, 1,112 generative AI services had passed the filing with China's regulator.[38][38] Beijing News on CAC filings Secondary Each must pass a test bank of at least 2,000 questions with 90% acceptable answers.[39][39] CSET translation of GB/T 45654-2025 Primary

    One Country Tests Before Launch; the Rest of Asia Builds Tools

    Legal force of AI testing by country or region, November 2026.

    Hover over, tap or tab to a tile for details.

    Binding test duty?

    • Yes, de facto through filing
    • Self-assessment only
    • Voluntary; accredits private testers
    • Voluntary guides and benchmarks
    • Proposals only
    • Not researched
    Data table
    Country or regionBinding test duty?ValueNote
    ChinaYes, de facto through filing1,112 filedCAC filing; GB/T 45654-2025; mandatory agent standard in drafting.
    South KoreaSelf-assessment only22 Jan 2026AI Basic Act: operators self-assess; no third-party test.
    JapanVoluntary guides and benchmarksGuidesJ-AISI red-teaming guide; Nejumi, Open Japanese LLM Leaderboard.
    GulfNot researchedOpen Arabic LLM Leaderboard idle since Oct 2025.
    IndiaProposals only12.8%AISI proposed Jan 2026; IndicSafe, MILU, Inspect India Evals.
    Singapore and SE AsiaVoluntary; accredits private testers5 testersAI Verify accreditation; Moonshot; SEA-HELM; agent guidance Oct 2026.
    Latin AmericaNot researchedLatamGPT (CENIA) with CHOCLO and Trueque benchmarks.
    Figure 1 in chapter 6, China and Asia. Source: Beijing News;[38][38] Beijing News on CAC filings Secondary Cooley;[40][40] Cooley on Korea's AI Basic Act Secondary AI Verify Foundation;[41][41] AI Verify Foundation, AI Tester Accreditation Primary IndicSafe;[42][42] IndicSafe, arXiv 2603.17915 Primary Open Arabic LLM Leaderboard data;[43][43] Open Arabic LLM Leaderboard Space data Primary Buenos Aires Times.[44][44] Buenos Aires Times on LatamGPT Secondary

    Read chapter 6 → Figure 1 and the full evidence

  9. Multimodal, voice and embodied

    9.Voice agents lose most of the skill they show in text.

    In the τ-voice study (2026), GPT-5 with reasoning solved 85% of the tasks in text mode.[45][45] τ-voice, arXiv 2603.13686 Primary Voice agents solved only 26% to 51% of the same tasks, and noise and accents cost the most.[45][45] τ-voice, arXiv 2603.13686 Primary

    The same tasks drop from 85% in text to as low as 26% in voice

    τ-voice task success (pass@1): GPT-5 with reasoning in text mode, against the best and worst voice agents in two audio conditions.

    Text modeVoice mode020406080100%85%51% Clean audio, best voice agent38% Noise and accents, best voiceagent31% Clean audio, worst voice agent26% Noise and accents, worst voiceagentRealistic callers costvoice agents 47–59 pointsagainst text mode.
    Data table
    LineText modeVoice modeChange
    Clean audio, best voice agent85%51%−34
    Clean audio, worst voice agent85%31%−54
    Noise and accents, best voice agent85%38%−47
    Noise and accents, worst voice agent85%26%−59
    Figure 6 in chapter 7, Multimodal, voice and embodied. Source: τ-voice, arXiv 2603.13686;[45][45] τ-voice, arXiv 2603.13686 Primary Artificial Analysis speech-to-speech (Nov 2026).[46][46] Artificial Analysis, speech-to-speech Primary

    Read chapter 7 → Figure 6 and the full evidence

  10. Enterprise practice and sector rules

    10.The best teams build their tests from real traffic and real users.

    Cursor retrained its code-completion model on live user feedback: suggestions fell 21% and the accept rate rose 28% (Sep 2025).[47][47] Cursor, Online RL for Tab Primary Shopify builds its test sets from samples of real merchant conversations, not hand-written examples (Aug 2025).[48][48] Shopify Engineering, Building production-ready agentic systems Primary

    At Cursor's scale, users are the eval

    Change after Cursor's online-RL Tab model replaced the previous model, September 2025. Both arrows start at the previous model, indexed to 100.

    7085100115130Index, previous model = 100ChangeSuggestions shown100−21%−21Previous modelOnline-RL modelAccept rate100+28%+28
    Data table
    RowPrevious modelOnline-RL modelChange
    Suggestions shown100−21%−21
    Accept rate100+28%+28
    Figure 4 in chapter 8, Enterprise practice and sector rules. Source: Cursor, Online RL for Tab.[47][47] Cursor, Online RL for Tab Primary

    Read chapter 8 → Figure 4 and the full evidence

Go deeper: eight chapters, one directory.

Each chapter has key findings, a ten-minute report with figures, and "Go deeper" sections with the full tables and sources.

  1. Benchmarks Static exams have hit their ceilings, audits show many ceilings were fake, and only long, expert-built tasks still spread the frontier.
  2. Evaluation science LLM judges fall near chance on hard tasks, models notice tests and cheat, and the best teams check the instrument first.
  3. Tools and technology Inspect, lm-evaluation-harness and Harbor became the standard runners; larger firms bought the platforms.
  4. Market and companies Expert data and RL environments earn billions from labs; builder tools stay small and keep getting bought.
  5. Institutions and regulation Government institutes anchor credibility, the AI Act's Article 55 is enforced from August 2026, and testing standards are thin.
  6. China and Asia China sets numeric pass marks and files every service; Japan and Korea write guides; Singapore accredits private testers.
  7. Multimodal, voice and embodied Vision tests saturated, video arenas disagree, deepfake detectors fail in the wild, and voice agents lose most of their text skill.
  8. Enterprise practice and sector rules Software teams gate releases on regressions and calibrated judges; banks validate each use case; insurers face dense testing duties.

Also in this report

Sources for this page

The numbers below apply to this page only. Each chapter carries its own numbered sources.

  1. [1]
    Benchmark data bundle, Epoch AI. Dated scores (5 Nov 2026) and ceilings for most benchmarks here. Primary epoch.ai
  2. [2]
    Benchmark Review registry, Epoch AI. Lists 13 Flawed, 5 Verified and 8 unrated benchmarks in November 2026. Primary epoch.ai
  3. [3]
    Remote Labor Index, arXiv 2510.26787. It describes the projects and the launch automation rate. Primary arxiv.org
  4. [4]
    Vending-Bench 2, Andon Labs. It gives final balances and the frontier trend. Primary andonlabs.com
  5. [5]
    Measuring AI ability to complete long tasks (Kwa et al.). It defines the 50% time horizon. Primary arxiv.org
  6. [6]
    GPT-5.6 Sol evaluation, METR. It shows how cheat scoring changes the time horizon. Primary metr.org
  7. [7]
    Time horizons, METR. It tracks the 50% time horizon of frontier models. Primary metr.org
  8. [8]
    Time Horizon 1.1 results, METR. It lists time horizons and intervals for each model. Primary metr.org
  9. [9]
    Claude Opus 5.5 system card, Anthropic. It reports capability, cyber, alignment and evaluation-awareness results. Vendor anthropic.com
  10. [10]
    Cheating behaviour in frontier model evaluations, UK AISI. It reports cheating by every tested model. Primary aisi.gov.uk
  11. [11]
    Gemini 4 Argon Safety and Security Report, Google DeepMind. It reports evaluation awareness and names the external testers. Vendor storage.googleapis.com
  12. [12]
    inspect-ai downloads, pypistats. It gives the monthly PyPI download count. Primary pypistats.org
  13. [13]
    lm-eval downloads, pypistats. It gives the monthly PyPI download count. Primary pypistats.org
  14. [14]
    swebench downloads, pypistats. It gives the monthly PyPI download count for SWE-bench. Primary pypistats.org
  15. [15]
    Inspect AI, GitHub. It gives stars, licence and release data. Primary github.com
  16. [16]
    inspect-evals downloads, pypistats. It gives the monthly PyPI download count. Primary pypistats.org
  17. [17]
    lm-evaluation-harness, GitHub. It gives stars, licence and release data. Primary github.com
  18. [18]
    Harbor, GitHub API. It gives stars, licence and the move from laude-institute. Primary api.github.com
  19. [19]
    harbor downloads, pypistats. It gives the monthly PyPI download count. Primary pypistats.org
  20. [20]
    SWE-bench, GitHub. It gives stars, licence and push data. Primary github.com
  21. [21]
    OpenCompass, GitHub. It gives stars, licence and release data. Primary github.com
  22. [22]
    lighteval, GitHub. It gives stars, licence and release data. Primary github.com
  23. [23]
    lighteval downloads, pypistats. It gives a monthly count that looks too low. Primary pypistats.org
  24. [24]
    openbench, GitHub. It gives stars, licence and release data. Primary github.com
  25. [25]
    HELM, GitHub. It gives stars, licence and last push. Primary github.com
  26. [26]
    Lab human-data forecast, FutureSearch. It estimates 2026 human-data spending by the top five labs. Estimate futuresearch.ai
  27. [27]
    Funding update August 2026, METR. It lists METR's new commitments, funders and funding policy. Primary metr.org
  28. [28]
    Silicon Valley bets big on environments, TechCrunch. It reports Anthropic's discussed RL-environment budget, citing The Information. Secondary techcrunch.com
  29. [29]
    Insurance-Canada.ca quoting Gartner. It quotes Gartner's forecast for AI governance platform spending. Secondary insurance-canada.ca
  30. [30]
    Series B announcement, Arena. It gives Arena's valuation, revenue and usage in October 2026. Vendor arena.ai
  31. [31]
    EU AI Act, Article 55. It gives the duties of providers of GPAI models with systemic risk. Primary artificialintelligenceact.eu
  32. [32]
    AI Act Service Desk, Commission enforcement powers FAQ. It explains the powers that apply from 2 August 2026. Primary ai-act-service-desk.ec.europa.eu
  33. [33]
    Garrigues, Digital Omnibus on AI published. It gives the new AI Act deadlines after the Omnibus. Secondary garrigues.com
  34. [34]
    European Commission, contents of the GPAI Code of Practice. It lists the chapters and the signatories. Primary digital-strategy.ec.europa.eu
  35. [35]
    JTC 21, EN 18286 becomes the first harmonised standard (14 Oct 2026). It reports the citation and the remaining standardisation work. Secondary jtc21.eu
  36. [36]
    CEN-CENELEC, AI standardisation fast-track (23 Oct 2025). It announces the exceptional measures for delayed drafts. Primary cencenelec.eu
  37. [37]
    Cloud Security Alliance, EU AI Act GPAI enforcement (29 Aug 2026). This secondary note covers fines, notification and non-signatories. Secondary labs.cloudsecurityalliance.org
  38. [38]
    Beijing News on CAC filings. It reports 1,112 filed services and 731 registrations by 31 August 2026. Secondary m.bjnews.com.cn
  39. [39]
    CSET translation of GB/T 45654-2025. It gives the English text of the generative AI safety standard and its test numbers. Primary cset.georgetown.edu
  40. [40]
    Cooley on Korea's AI Basic Act. It summarises the Act's duties, thresholds and fines. Secondary cooley.com
  41. [41]
    AI Verify Foundation, AI Tester Accreditation. It lists the accredited firms and the programme terms. Primary aiverifyfoundation.sg
  42. [42]
    IndicSafe, arXiv 2603.17915. It tests culturally grounded safety prompts in 12 Indic languages. Primary arxiv.org
  43. [43]
    Open Arabic LLM Leaderboard Space data. It shows the Space's last update. Primary huggingface.co
  44. [44]
    Buenos Aires Times on LatamGPT. It reports the LatamGPT launch in Santiago. Secondary batimes.com.ar
  45. [45]
    τ-voice, arXiv 2603.13686. Primary arxiv.org
  46. [46]
    Artificial Analysis, speech-to-speech. Primary artificialanalysis.ai
  47. [47]
    Cursor, Online RL for Tab. Company self-report. Primary cursor.com
  48. [48]
    Shopify Engineering, Building production-ready agentic systems. Primary shopify.engineering