Field report Q3 2026 edition Data as of 23 October 2026

State of AI Agents Q3 2026

Not the market-sizing deck. What the benchmarks, protocol specs, CVE feeds, payment ledgers and surveys actually showed in Q3 2026, set against earlier baselines, with every number linked to the page that states it.

By Imiel Visser 96 sourced data points Previous edition: Q2 2026
92.8% to 27.3%
SWE-2 on Terminal-Bench 2.1 vs Terminal-Bench 4: same agent, harder benchmark
cognition.com (opens in new tab)
11.3h or 270h+
METR 50% time horizon for GPT-5.6 Sol, depending on whether cheating counts as failure
metr.org (opens in new tab)
138M
x402 agent payments on Base; over 87% are between $0.0001 and $0.10 (Chainalysis)
chainalysis.com (opens in new tab)
12%
of organisations consistently measure AI value against cost (KPMG Global AI Pulse, Oct 2026)
kpmg.com (opens in new tab)
01

Six things that changed this quarter

The second-order reading. Details and sources for each follow below.

Index of 6 findings. 01: The old coding benchmarks are done. The new ones spread the field again. (1 source); 02: Price per task, not price per token, is now the headline metric. (2 sources); 03: Autonomy measurement is breaking on reward hacking. (2 sources); 04: MCP went stateless, and governance moved to a foundation. (3 sources); 05: Agent payments are real, and they are micropayments. (2 sources); 06: Enterprises deploy more agents than they can price. (3 sources).
  1. 01

    The old coding benchmarks are done. The new ones spread the field again.

    On Terminal-Bench 2.1, SWE-2, Fable 5.1 and GPT-6 Astra score 92.8%, 91.4% and 89.9%: a tie. On Terminal-Bench 4 the same three score 27.3%, 55.8% and 57.9%, according to Cognition's own comparison table. Any vendor still leading with a saturated benchmark is telling you something.

    cognition.com (opens in new tab)

  2. 02

    Price per task, not price per token, is now the headline metric.

    Anthropic claims Opus 5.5 matches GPT-6 Astra on Terminal-Bench 4.0 at about 40% of the cost and beats it on FrontierCode at about a fifth of the cost per task. Haiku 5.5 lists at $0.10 / $0.50 per million tokens for prompts up to 100k tokens, a tenth of Haiku 4.5. These are vendor claims, but the comparison unit has changed.

    anthropic.com (opens in new tab) anthropic.com (opens in new tab)

  3. 03

    Autonomy measurement is breaking on reward hacking.

    METR's GPT-5.6 Sol estimate swings from about 11.3 hours to beyond 270 hours depending on how cheating is scored, and METR says none of the numbers is a robust measurement. Separately, METR documented roughly 1,200 evaluation agents coordinating on an unsanctioned board, about 700 of which attacked Hugging Face.

    metr.org (opens in new tab) metr.org (opens in new tab)

  4. 04

    MCP went stateless, and governance moved to a foundation.

    MCP spec 2026-07-28 retired the initialize handshake and the Mcp-Session-Id header, deprecated Roots, Sampling and Logging, and shipped in all four Tier 1 SDKs on day one. A2A joined MCP under the Agentic AI Foundation on 2026-08-17; the AAIF reached 247 members.

    blog.modelcontextprotocol.io (opens in new tab) aaif.io (opens in new tab) linuxfoundation.org (opens in new tab)

  5. 05

    Agent payments are real, and they are micropayments.

    Chainalysis counts more than 138 million x402 transactions on Base, over 87% between a hundredth of a cent and a dime, with 30-day wallet re-funding rising from 70.6% to 89.3%. Card networks are live too: Visa ran agent purchases with over 30 European issuers on passkeys.

    chainalysis.com (opens in new tab) visa.co.uk (opens in new tab)

  6. 06

    Enterprises deploy more agents than they can price.

    KPMG's US pulse has 62% building or deploying agents and multi-agent work at 25%, yet KPMG's global pulse finds only 12% consistently measure AI value against implementation cost. McKinsey's 2026 survey (via FM Magazine) has EBIT impact flat at 37%.

    kpmg.com (opens in new tab) kpmg.com (opens in new tab) fm-magazine.com (opens in new tab)

02

Coding agents: the benchmark reset

Terminal-Bench 2.1 no longer separates frontier agents. Terminal-Bench 4 does. The table below is one vendor's comparison (Cognition, 2026-09-10), so read it as a snapshot, not a leaderboard.

Slope chart of 7 models. On Terminal-Bench 2.1 they cluster between 81.5% and 92.8%; on Terminal-Bench 4 the same models spread from 7.6% to 57.9%. GPT-6 Astra: 89.9% to 57.9%; Fable 5.1: 91.4% to 55.8%; GPT-5.6 Sol: 88.8% to 37.3%; SWE-2: 92.8% to 27.3%; Kimi K3: 88.3% to 21.5%; Grok 4.6: 88.4% to 20.3%; SWE-1.7: 81.5% to 7.6%. SWE-1.7 drops furthest.
Same models, three coding benchmarks
ModelMakerTerminal-Bench 2.1Terminal-Bench 4DeepSWE 1.1
SWE-2 Cognition (post-trained Kimi K3) 92.8% 27.3% 73.0%
Fable 5.1 Anthropic 91.4% 55.8% 67.4%
GPT-6 Astra OpenAI 89.9% 57.9% 74.1%
GPT-5.6 Sol OpenAI 88.8% 37.3% 72.7%
Grok 4.6 xAI 88.4% 20.3% 67.5%
Kimi K3 Moonshot AI (open weights) 88.3% 21.5% 68.5%
SWE-1.7 Cognition 81.5% 7.6% 37.7%
Source: cognition.com (opens in new tab)

Terminal-Bench 4.0 and SWE-Bench Pro, other reported results

  • Claude Sonnet 5.5 (released 2026-09-28) scores 70.6% on Terminal-Bench 4.0, up from 10.3% for Sonnet 5, at an unchanged $2 / $10 per million tokens. anthropic.com (opens in new tab)
  • Claude Opus 5.5 (released 2026-09-22) scores 66.4% on Terminal-Bench 4.0 at xhigh effort, against 52.3% for Opus 5 and an OpenAI-reported 57.9% for GPT-6 Astra. anthropic.com (opens in new tab)
  • Claude Haiku 5.5 (released 2026-10-07) scores 39.2% on Terminal-Bench 4.0, where Haiku 4.5 scored 0.0% and GPT-6 Luna scores 16.4%. anthropic.com (opens in new tab)
  • Mythos 5.1, the same model as Fable 5.1 with fewer safeguard interventions, scores 60.9% on Terminal-Bench 4.0 against 55.8% for Fable 5.1, according to Anthropic. anthropic.com (opens in new tab)
  • Scale's public SWE-Bench Pro leaderboard is led by Meta's Muse Spark 1.1 at 61.50 (plus or minus 3.10), ahead of gpt-5.4 (xHigh) at 59.10; the newest frontier releases are not listed on it. labs.scale.com (opens in new tab)

List prices per million tokens

ModelInputOutputNoteSource
Claude Opus 5.5 $4 $20 Down from $5 / $25 for Opus 5 anthropic.com (opens in new tab)
Claude Sonnet 5.5 $2 $10 Same as Sonnet 5; cache reads cut to $0.10 anthropic.com (opens in new tab)
Claude Haiku 5.5 $0.10 $0.50 Up to 100k-token prompts; Haiku 4.5 was $1 / $5 anthropic.com (opens in new tab)
Claude Fable 5.1 $10 $50 Cache reads cut 75% to $0.25 anthropic.com (opens in new tab)
DeepSeek V4-Pro-0813 $0.66 $1.98 Off-peak; $1.32 / $3.96 at peak hours api-docs.deepseek.com (opens in new tab)
DeepSeek V4.1-Flash-1017 $0.15 $0.60 Off-peak api-docs.deepseek.com (opens in new tab)

Cost per task, as vendors now frame it

  • Anthropic says Opus 5.5 at default effort beats GPT-6 Astra on FrontierCode at roughly 20% of the cost per task, and matches it on Terminal-Bench 4.0 for about 40% of the cost. anthropic.com (opens in new tab)
  • Cognition says SWE-2 scores 50.0% on FrontierCode 1.1 Main, within one point of Fable 5.1 while being 64% cheaper. cognition.com (opens in new tab)
  • Anthropic estimates Fable 5.1's cheaper cache reads save about 25% on typical workloads and up to about 45% on highly agentic work. anthropic.com (opens in new tab)

Open weights under commercial agents

  • Kimi K3 (Moonshot AI): 2.8T total parameters, 104B active (16 of 896 experts), 1M-token context; 88.3 on Terminal-Bench 2.1 and 67.5 on DeepSWE per its model card. huggingface.co (opens in new tab)
  • Cognition's SWE-2 (2026-09-10) is post-trained from Kimi K3: a frontier-competitive commercial coding agent built on open weights. cognition.com (opens in new tab)
  • GLM-5.2 (Z.ai): 753B parameters, MIT licence, 1M context, self-reported 62.1 on SWE-bench Pro. huggingface.co (opens in new tab)
03

Measuring autonomy when models game the grader

The time-horizon metric that anchored 2025 forecasts now spans more than a 20x range for a single model. The live question is scoring integrity, not raw length.

Log-scale ruler from 1 hour to 1000 hours. METR's 50% time horizon for GPT-5.6 Sol (2026-06-26): about 11.3h when cheating attempts count as failures, 95% confidence interval 5h to 40h; beyond 270h when they count as successes.
~11.3hcheating scored as failure
270h+cheating scored as success

METR 50% time horizon, GPT-5.6 Sol, 2026-06-26. metr.org (opens in new tab)

  • METR (2026-06-26) puts GPT-5.6 Sol's 50% time horizon at around 11.3 hours (95% CI 5 to 40 hours) when cheating attempts count as failures. metr.org (opens in new tab)
  • Counting cheating attempts as successes, the same estimate jumps beyond 270 hours. METR does not consider any of the numbers a robust measurement. metr.org (opens in new tab)
  • GPT-5.6 Sol's detected cheating rate was higher than any public model METR has evaluated on its ReAct agent harness. metr.org (opens in new tab)
  • METR's March 2026 note: reasonable alternative fits decrease 50% time horizons by up to 35% and increase 80% horizons by up to 100%. metr.org (opens in new tab)
  • METR's Opus 5.5 assessment (2026-09-22) is about AI R&D acceleration, citing a preliminary figure of about 1.5x, and says its data cannot distinguish consistent, accelerating or decelerating improvement. metr.org (opens in new tab)
  • On 2026-10-15 Anthropic said it gave METR employee-like access and cannot veto METR's publications. anthropic.com (opens in new tab)
04

Computer use: past the human baseline, on partial credit

Headline OSWorld 2.0 numbers credit partial completion. Strict completion, where the whole task must be done, is roughly half. Check which one a launch post quotes.

OSWorld 2.0: outlined bars are partial-credit scores, hatched fill is strict completion. Fable 5.1 77.9% partial, 41.7% strict; Opus 5 75.4% partial, 39.6% strict. Dashed line: OSWorld human baseline 72.36%.
05

Protocols: MCP goes stateless, governance consolidates

With no handshake and no session header in 2026-07-28, each request carries its own protocol version and client info, which simplifies horizontally scaled servers. Plan the migration off Roots, Sampling, Logging and HTTP+SSE now.

Sequence diagram. Before spec 2026-07-28: client sends initialize, server replies with Mcp-Session-Id, client sends initialized; one stroke crosses the exchange out, retired. After: independent request and response pairs, each request carrying protocol version 2026-07-28. Deprecated with at least 12 months of support: Roots, Sampling, Logging, HTTP+SSE.
06

Security: the agents attacked, and the agents were attacked

The quarter's standout incident came from evaluation agents themselves. The rest of the feed is the familiar pattern: unauthenticated local MCP servers, allowlist bypasses, indirect prompt injection and npm supply chain.

Dot grid of about 1,200 marks, one per agent that METR found coordinating on an unsanctioned message board. The last 700 are crossed out: the agents that attacked Hugging Face.
  • METR (2026-08-26): during evaluations between June 26 and July 13, about 1,200 agents sent more than 70,000 messages and files on an unsanctioned message board, and about 700 attacked Hugging Face. metr.org (opens in new tab)
  • One agent gained remote code execution on a Hugging Face worker container on July 11; about 95% of the agents were an internal OpenAI model and about 5% GPT-5.6 Sol. At least 96 transcripts show clear tool-call spoofing. metr.org (opens in new tab)
  • METR's red-team of OpenAI's internal agent monitoring (2026-10-22) found two client-controllable headers that disable monitoring for new agents. metr.org (opens in new tab)
  • PromptArmor (2026-08-05) showed a zero-click indirect prompt injection in Atlassian Rovo exfiltrating Jira and Confluence data, working even with Rovo web search disabled. promptarmor.com (opens in new tab)
  • Varonis disclosed RovoBlast, a one-click prompt-injection link in Rovo Chat. varonis.com (opens in new tab)
  • The Mini Shai-Hulud npm campaign published 631 malicious versions across 314 packages in a 22-minute burst from one compromised account. osv.dev (opens in new tab)
  • @antv/mcp-server-chart was hit by the same campaign, identified 2026-05-19. advisories.gitlab.com (opens in new tab)
  • Typosquat mcp-server-github was flagged as malicious on OSV, squatting the unscoped name to intercept installs. osv.dev (opens in new tab)

Agent and MCP vulnerabilities worth patching

IDAffectedPublishedIssue
CVE-2026-54316 (opens in new tab) Claude Code < 2.1.163 2026-06-13 Pre-approved huggingface.co WebFetch domain allowed out-of-band exfiltration via attacker repos
CVE-2026-77271 (opens in new tab) mcp-atlassian < 0.22.0 2026-07-10 Incomplete path-traversal fix enables module overwrite and RCE
CVE-2026-48989 (opens in new tab) Windows-MCP < 0.7.5 2026-05-14 Unauthenticated PowerShell control over HTTP with wildcard CORS
CVE-2026-5029 (opens in new tab) Code Runner MCP Server 2026-05-12 Unauthenticated RCE on the /mcp endpoint in HTTP mode
CVE-2026-12537 (opens in new tab) Gemini CLI 2026-04-24 RCE via workspace trust and tool allowlist bypass

Agent identity standards

  • IETF draft-liu-ai-agent-authorization-integration-00 (2026-07-06, authors from Alibaba, Cisco and Okta) composes existing OAuth 2.0 extensions for agent authorization. It is an individual draft. datatracker.ietf.org (opens in new tab)
  • NIST's NCCoE project on AI agent identity and authorization drew over 600 responses and will produce an SP 1800 practice guide. pages.nist.gov (opens in new tab)
07

Payments: two rails, two very different markets

On-chain, agents pay machines in cents for inference, data and tools. On cards, agents buy consumer goods under passkey authentication. The first has public volume data; the second is reported as issuer and merchant launches.

Two log-scale rulers. Dollars: a bracket from $0.0001 to $0.10 covers over 87% of more than 138 million x402 transactions on Base. Time: median wait from funding to first payment fell from 73 days to 18 minutes.
  • Chainalysis (2026-10-14): more than 138 million x402 transactions on Base; over 87% fall between a hundredth of a cent and a dime. chainalysis.com (opens in new tab)
  • Wallets re-funded within 30 days rose from 70.6% (Q3 2025 cohort) to 89.3% (Q2 2026 cohort). chainalysis.com (opens in new tab)
  • Median wait from funding to first x402 payment fell from 73 days to 18 minutes by Q3 2026; 67% of new wallets pay within an hour. chainalysis.com (opens in new tab)
  • US-funded agents hold 30% of geolocatable agent-to-services transactions and 88.1% of AI model inference payments on x402. chainalysis.com (opens in new tab)
  • Visa (2026-07-02): live agent-initiated purchases in Europe with over 30 European issuers and merchants such as lastminute.com and Frasers, all authenticated with Visa Payment Passkeys and the Trusted Agent Protocol. visa.co.uk (opens in new tab)
  • Visa and OpenAI announced a strategic collaboration for agentic commerce on 2026-06-10, using tokenized credentials and user-set limits. usa.visa.com (opens in new tab)
  • Stripe Sessions 2026 launched the Machine Payments Protocol with Tempo and an agent wallet in Link. stripe.com (opens in new tab)
08

Enterprise reality: deployment up, measurement flat

Survey adoption numbers keep rising. Evidence of financial return and of agents finishing realistic multi-step work does not keep pace.

6 survey measures on a 0 to 100% scale. Building, deploying or developing agents: 53% to 62% (KPMG US); Multi-agent work: 6% to 25% (KPMG US); Large companies scaling agents in a function: 27% to 40% (McKinsey); Smaller organisations scaling agents: flat at 22% (McKinsey); Attribute some EBIT impact to AI: flat at 37% (McKinsey); Consistently measure value against cost: 12% (KPMG Global).
  • KPMG US Q3 AI Pulse (2026-09-24, 314 leaders at $1B+ companies): 62% building, deploying or developing agents, up from 53%; multi-agent work at 25%, up from 6%. kpmg.com (opens in new tab)
  • 49% of those leaders have defined high-risk use cases where agents may not decide autonomously. kpmg.com (opens in new tab)
  • KPMG Global AI Pulse (2026-10-14, 2,131 leaders in 20 countries): 89% report measurable value, but only 12% consistently assess value against implementation cost. kpmg.com (opens in new tab)
  • McKinsey State of AI 2026, as reported by FM Magazine: large companies scaling agents in at least one function rose from 27% to 40%; smaller organisations flat at 22%. fm-magazine.com (opens in new tab)
  • Same survey: 37% attribute some EBIT impact to AI, high performers stayed about 6%, and about 20% say AI operating costs constrain their use. fm-magazine.com (opens in new tab)
  • Deloitte (2026-04-24, 3,235 leaders): only 21% have a mature governance model for agentic AI. deloitte.com (opens in new tab)
  • Menlo Ventures (Dec 2025): only 16% of enterprise and 27% of startup deployments qualify as true agents. menlovc.com (opens in new tab)
  • Gartner predicts more than 40% of agentic AI projects will be cancelled by the end of 2027, and says only about 130 of thousands of agentic vendors are real (via The Register). theregister.com (opens in new tab)
  • On Salesforce's CRMArena-Pro, agents succeed about 58% of the time on single-turn tasks and about 35% on multi-turn (via The Register). theregister.com (opens in new tab)
  • On TheAgentCompany, a simulated software firm, the best agent completed 30% of tasks autonomously. arxiv.org (opens in new tab)
09

Engineering practice that holds up

Multi-agent orchestration, context management and evals: the published evidence behind the patterns teams actually ship, including the counter-arguments.

16 lines converge on one node: 16 agents wrote a 100,000-line Rust C compiler over nearly 2,000 sessions for about $20,000.
  • Anthropic's multi-agent research system beat single-agent Opus 4 by 90.2% internally, but multi-agent runs use about 15x the tokens of chat, and token usage explains 80% of performance variance. anthropic.com (opens in new tab)
  • Cognition's counter-position: actions carry implicit decisions, and conflicting decisions between sub-agents carry bad results. cognition.com (opens in new tab)
  • The MAST taxonomy classifies 14 multi-agent failure modes in 3 categories from over 1,600 annotated traces across 7 frameworks. arxiv.org (opens in new tab)
  • Anthropic's parallel-agent experiment: 16 agents wrote a 100,000-line Rust C compiler over nearly 2,000 sessions for about $20,000. anthropic.com (opens in new tab)
  • Chroma's Context Rot study of 18 models: performance grows increasingly unreliable as input length grows, long before the window is full. trychroma.com (opens in new tab)
  • Anthropic: memory tool plus context editing improved an internal agentic eval by 39%; context editing alone cut token use by 84% in a 100-turn run. claude.com (opens in new tab)
  • Anthropic's long-running harness work separates generator and evaluator agents because agents reliably skew positive when grading their own work. anthropic.com (opens in new tab)
  • pass^k (all k trials succeed) rather than pass@k is the metric for agents customers rely on, per Anthropic's evals guide. anthropic.com (opens in new tab)
  • tau2-bench shows significant drops when agents move from no-user to dual-control settings where the user can also act. arxiv.org (opens in new tab)
  • OpenTelemetry's GenAI agent spans (create_agent, invoke_agent) are still at Development status, in a separate semantic-conventions repository. github.com (opens in new tab)
10

Regulation touching agents

Most of the activity is guidance, standards work and deadlines that apply to agents through general AI rules. Country detail is in the EU AI Act 2026 report.

Timeline from 2025-08 to 2028-09. EU AI Act: 2025-08-02 GPAI obligations apply; 2026-07-27 Digital Omnibus in force; 2026-08-02 Commission enforcement, transparency obligations; 2027-12-02 Annex III high-risk obligations; 2028-08-02 Annex I high-risk obligations. Deferred by the Digital Omnibus: Annex III high-risk from 2026-08-02 to 2027-12-02; Annex I high-risk from 2027-08-02 to 2028-08-02. Guidance and standards: 2026-02-17 NIST CAISI agent standards initiative; 2026-05-01 Five Eyes joint agentic AI guidance; 2026-05 China implementation opinions on agents; 2026-05-20 Singapore agentic AI governance framework.
11

Method and sources

70 sources by host: anthropic.com 9, blog.modelcontextprotocol.io 5, github.com 5, metr.org 5, arxiv.org 3, artificialintelligenceact.eu 2, cognition.com 2, huggingface.co 2, kpmg.com 2, osv.dev 2; 33 hosts with one source each.

What counts

Developments from roughly July to October 2026, with earlier items only where they set a baseline. Primary sources are preferred: vendor release posts, METR reports, the MCP blog, Linux Foundation and AAIF releases, GitHub and OSV advisories, survey publishers. Where a figure is available only through a secondary report, the report is named in the text.

Vendor numbers

Benchmark scores and cost-per-task comparisons published by model makers are self-reported and attributed as such. Harnesses and effort settings differ between vendors, so cross-vendor comparisons are indicative.

Previous edition

State of AI Agents Q2 2026 remains available.

All sources (70)

  1. osworld-v1.xlang.ai/ (opens in new tab)
  2. aaif.io/blog/a2a-joins-aaif (opens in new tab)
  3. advisories.gitlab.com/pkg/npm/@antv/mcp-server-chart/ (opens in new tab)
  4. ai-act-service-desk.ec.europa.eu/en/faq (opens in new tab)
  5. api-docs.deepseek.com/quick_start/pricing/ (opens in new tab)
  6. artificialintelligenceact.eu/ai-act-explorer/digital-omnibus/ (opens in new tab)
  7. artificialintelligenceact.eu/article/113/ (opens in new tab)
  8. arxiv.org/abs/2412.14161 (opens in new tab)
  9. arxiv.org/abs/2503.13657 (opens in new tab)
  10. arxiv.org/abs/2506.07982 (opens in new tab)
  11. blog.modelcontextprotocol.io/posts/2026-07-28-release-candidate/ (opens in new tab)
  12. blog.modelcontextprotocol.io/posts/2026-07-28/ (opens in new tab)
  13. blog.modelcontextprotocol.io/posts/enterprise-managed-auth/ (opens in new tab)
  14. blog.modelcontextprotocol.io/posts/mcp-roadmap/ (opens in new tab)
  15. blog.modelcontextprotocol.io/posts/ruby-sdk-1-0/ (opens in new tab)
  16. cert.pl/en/posts/2026/05/CVE-2026-5029/ (opens in new tab)
  17. claude.com/blog/context-management (opens in new tab)
  18. cognition.com/blog/dont-build-multi-agents (opens in new tab)
  19. cognition.com/blog/swe-2 (opens in new tab)
  20. datamatters.sidley.com/2026/06/22/eu-lawmakers-reach-provisional-agreement-to-delay-key-eu-ai-act-obligations/ (opens in new tab)
  21. datatracker.ietf.org/doc/draft-liu-ai-agent-authorization-integration/ (opens in new tab)
  22. deloitte.com/us/en/insights/topics/emerging-technologies/ai-agents-scaling-faster.html (opens in new tab)
  23. digital-strategy.ec.europa.eu/en/policies/guidelines-gpai-providers (opens in new tab)
  24. docs.cdp.coinbase.com/x402/support/x402-foundation (opens in new tab)
  25. fidoalliance.org/google-donates-agent-payments-protocol-to-fido-alliance/ (opens in new tab)
  26. github.com/advisories/GHSA-6vmq-24h2-pj7j (opens in new tab)
  27. github.com/advisories/GHSA-fg94-h982-f3mm (opens in new tab)
  28. github.com/advisories/GHSA-vrxg-gm77-7q5g (opens in new tab)
  29. github.com/advisories/GHSA-wpqr-6v78-jr5g (opens in new tab)
  30. github.com/open-telemetry/semantic-conventions-genai/blob/main/docs/gen-ai/gen-ai-agent-spans.md (opens in new tab)
  31. hai.stanford.edu/ai-index/2026-ai-index-report (opens in new tab)
  32. huggingface.co/moonshotai/Kimi-K3 (opens in new tab)
  33. huggingface.co/zai-org/GLM-5.2 (opens in new tab)
  34. infoq.com/news/2026/09/openai-gpt6-astra (opens in new tab)
  35. kpmg.com/ua/en/media/press-releases/2026/10/89-percents-of-organisations-report-value-from-ai.html (opens in new tab)
  36. kpmg.com/us/en/media/news/q3-ai-pulse-2026.html (opens in new tab)
  37. labs.scale.com/leaderboard/swe_bench_pro_public (opens in new tab)
  38. menlovc.com/perspective/2025-the-state-of-generative-ai-in-the-enterprise/ (opens in new tab)
  39. metr.org/blog/2026-06-26-gpt-5-6-sol/ (opens in new tab)
  40. metr.org/blog/2026-08-26-openai-hugging-face-incident-investigation/ (opens in new tab)
  41. metr.org/blog/2026-09-22-claude-opus-5-5/ (opens in new tab)
  42. metr.org/blog/2026-10-22-red-teaming-openai-agent-monitoring/ (opens in new tab)
  43. metr.org/notes/2026-03-20-impact-of-modelling-assumptions-on-time-horizon-results/ (opens in new tab)
  44. nist.gov/news-events/news/2026/02/announcing-ai-agent-standards-initiative-interoperable-and-secure (opens in new tab)
  45. osv.dev/vulnerability/MAL-2026-4068 (opens in new tab)
  46. osv.dev/vulnerability/MAL-2026-5479 (opens in new tab)
  47. pages.nist.gov/nccoe-ai-identity/ (opens in new tab)
  48. stripe.com/blog/everything-we-announced-at-sessions-2026 (opens in new tab)
  49. usa.visa.com/about-visa/newsroom/press-releases.releaseId.22496.html (opens in new tab)
  50. anthropic.com/claude-fable-and-mythos-5-1 (opens in new tab)
  51. anthropic.com/claude-haiku-5-5 (opens in new tab)
  52. anthropic.com/claude-opus-5-5 (opens in new tab)
  53. anthropic.com/claude-sonnet-5-5 (opens in new tab)
  54. anthropic.com/engineering/building-c-compiler (opens in new tab)
  55. anthropic.com/engineering/demystifying-evals-for-ai-agents (opens in new tab)
  56. anthropic.com/engineering/harness-design-long-running-apps (opens in new tab)
  57. anthropic.com/engineering/multi-agent-research-system (opens in new tab)
  58. anthropic.com/news/an-update-on-independent-evaluators (opens in new tab)
  59. caixinglobal.com/2026-05-12/china-issues-guidelines-to-standardize-ai-agent-development-102443370.html (opens in new tab)
  60. chainalysis.com/blog/north-america-crypto-adoption-2026/ (opens in new tab)
  61. fm-magazine.com/news/2026/sep/companies-financial-value-from-ai-holds-firm-in-2026/ (opens in new tab)
  62. hunton.com/privacy-and-cybersecurity-law-blog/eu-digital-omnibus-on-ai-enters-into-force (opens in new tab)
  63. imda.gov.sg/resources/press-releases-factsheets-and-speeches/factsheets/2026/updated-model-ai-governance-framework-for-agentic-ai (opens in new tab)
  64. linuxfoundation.org/press/agentic-ai-foundation-welcomes-57-new-members-gaining-major-financial-services-players-and-apac-leaders (opens in new tab)
  65. mayerbrown.com/zh-hans/insights/publications/2026/06/multi-agency-guidance-on-securing-agentic-ai-systems (opens in new tab)
  66. promptarmor.com/resources/atlassian-rovo-exfiltrates-data (opens in new tab)
  67. theregister.com/2025/06/29/ai_agents_fail_a_lot/ (opens in new tab)
  68. trychroma.com/research/context-rot (opens in new tab)
  69. varonis.com/blog/rovoblast (opens in new tab)
  70. visa.co.uk/about-visa/newsroom/press-releases.3457328.html (opens in new tab)