Research Report

Question: What are the most significant developments in AI this week?

Date: 2026-08-14T13:21:17.567528641+00:00

Rounds: 4

Status: COMPLETE

Evidence: 62 claims · 50 sourced · 4 partial · 6 unsupported · 2 self-reported (no independent source) · 3 single-source

Executive Summary

For the week of May 3–9, 2026, the most significant AI developments, in order:

1. OpenAI made GPT-5.5 Instant the default ChatGPT model (May 5) — the week's biggest product event, primary-confirmed by OpenAI's announcement and TechCrunch.
2. The Pentagon signed classified-network AI agreements with eight companies — excluding Anthropic (announced May 1; the week's dominant business/policy story), primary-confirmed by the Department of War release and CNN.
3. Baidu released ERNIE 5.1 (May 8) — the first Chinese model to reach LMArena's global top 10, primary-confirmed by Baidu's ERNIE blog.

The week was otherwise quiet at the frontier: no Anthropic, Google, or Meta flagship shipped (Google's Gemini 3.5 came May 19; Anthropic's next Opus, 4.8, came May 28), and xAI had no confirmed week-dated release. Several stories circulated as "this week's" news turned out to be misdated or unverified — most notably a "Claude Opus 4.7 launched May 5" claim (Opus 4.7 actually shipped in April) and a "UK AISI found GPT-5.5 matched Mythos" report (the AISI evaluation was published April 13 and compared GPT-5.4/GPT-5.3-Codex, with Mythos ranked first). See the corrections table under Analysis.

The week's top developments

Analysis

The week's center of gravity moved from raw frontier scale to three things: default-model quality (OpenAI), geopolitics of AI access (the Pentagon/Anthropic fight, and China's efficiency push), and efficiency research. OpenAI owned the product narrative — the 5.5 Instant default swap, the 5.5-Cyber release, and a Codex CLI update all landed within 72 hours. The Pentagon announcement made the geopolitics concrete and awkward for Anthropic: excluded from the eight-company agreement while simultaneously reported to be renting SpaceX's Colossus 1 supercomputer (that deal, ~220K NVIDIA GPUs at 300MW, reported May 6, remains aggregator-sourced and unverified at primary record). On the competitive side, ERNIE 5.1's LMArena top-10 entry at roughly 6% of frontier pre-training compute is the strongest single data point yet for the "China–US gap has narrowed" narrative.

The research theme was efficiency and test-time compute: ZAYA1-8B's sub-1B-active-parameter reasoning claims, OpenSeeker-v2's SFT-only SOTA, SubQ's 12M-context launch (May 5, claims unvalidated and no architecture paper yet), and DeepSeek's "Thinking with Visual Primitives" — the week's most-discussed paper (spatial markers as "minimal units of thought"; 98.7% spatial reasoning; ~7,000× token compression), which went viral May 4 but was dated April 30 and later quietly pulled with no weights released. Treat it as research-in-flux, not a shipped capability.

Policy/regulatory was a documented vacuum: the UK AISI published nothing between April 29 and May 14; EU AI Act milestones bracketed the week (transparency obligations effective Aug 2; code finalized June); the US BIS chip-loophole action came May 31; and ICML 2026 notifications had already gone out April 30.

Claims that did not survive verification

Risks & Open Questions

• Provisional items to verify before citing as fact: Gemma 4 (DeepMind blog), the Anthropic–SpaceX Colossus deal (primary announcement), and SubQ's 12M-context claims (third-party validation and architecture paper were explicitly pending).
• OpenAI's benchmark claims (52.5% hallucination reduction, AIME 81.2) come from secondary reporting of OpenAI's press materials, not the primary page.
• The arXiv record for the week is incomplete: the monthly listing endpoint was broken, so other notable May 3–9 papers may exist beyond those covered here (OpenAlex counts ~14,550 arXiv works in the window).
• DeepSeek's "Visual Primitives" was pulled without weights; the paper says integration into DeepSeek's foundation model is planned "in the future" — watch for a re-release.

Claims without independent support

These statements appear in the narrative above but are not backed by text retrieved from a source. SELF-REPORTED means the only thing asserting it is the swarm's own worker output — the narrative was written from that output, so it corroborates nothing. Treat all of these as unverified.

Detailed Findings

Round 0 · Finding 1

The browser-rendered bodies are coming back empty. Let me try the http_request with the browser_fallback disabled, or use the goto verb in the browser to get actual content.

Round 0 · Finding 2

Significant AI Model Releases — Week of May 3–9, 2026

Executive Summary

The first full week of May 2026 was defined by OpenAI's GPT-5.5 Instant rollout — a strategic pivot toward reliability and personalization rather than raw scale — and Google DeepMind's Gemma 4, the most capable open-weight model family of its class. OpenAI also shipped a new suite of real-time voice models, and the week closed with major infrastructure and partnership news (Anthropic–SpaceX compute deal) that would reshape the industry's economics. Notably absent: no new frontier flagship from Anthropic (no Opus released this week), xAI, or Meta landed during May 3–9; their marquee releases came later in the month (e.g., Claude Opus 4.8 on May 28, Qwen 3.7-Max on May 20).

Key Findings

1. OpenAI — GPT-5.5 Instant (Released May 5, 2026)

Verified details:

• OpenAI updated ChatGPT's default engine to GPT-5.5 Instant, emphasizing "high-fidelity reliability" — a focus on reducing hallucinations and improving context-aware personalization, rather than parameter scaling.
• Hallucination reduction: 52.5% fewer hallucinated claims than its predecessor, GPT-5.3 Instant, and 37.3% fewer general inaccuracies (per OpenAI press documentation as reported by devFlokers).
• Notable benchmark results:
  • AIME 2025 (high-school math): 81.2 (up from 65.4 for GPT-5.3)
  • MMMU-Pro (multimodal reasoning): 76.0 (up from 69.2)
  • ARC-AGI 2 (fluid/visual intelligence): 85.0
  • GPQA Diamond (graduate-level reasoning): 93.6
• New "Memory Sources" architecture: the model can synthesize from past conversations, uploaded files, and personal data streams (e.g., Gmail), with user-facing provenance inspection and a "temporary chat" privacy bypass.
• On May 4, AWS announced it is bringing GPT-5.5 and 5.4, plus the Codex coding agent, to Amazon Bedrock, extending the model's enterprise distribution.

Sources: https://www.devflokers.com/blog/latest-ai-models-open-source-projects-may-2026 ; https://web.archive.org/web/20260719210226/https://www.aicritique.org/us/2026/06/01/ai-developments-in-may-2026/

2. OpenAI — Real-Time Voice Model Suite (Shipped May 7, 2026)

Verified details:

• OpenAI unveiled three new voice API models: GPT-Realtime-2, GPT-Realtime-Translate, and GPT-Realtime-Whisper.
• These enable real-time speech understanding, translation, and transcription with "GPT-5-class reasoning" embedded in the voice pipeline — a significant upgrade over previous speech-only models.
• OpenAI also shipped ads in ChatGPT that same day, a notable monetization development.

Sources: https://aitoolsrecap.com/Blog/ai-news-may-2026 ; https://web.archive.org/web/20260719210226/https://www.aicritique.org/us/2026/06/01/ai-developments-in-may-2026/

3. Google DeepMind — Gemma 4 (Released Week of May 3–9, 2026)

…(truncated — the summary above captures the substance)

Round 0 · Finding 3

Notable AI Research Developments — Week of May 3–9, 2026

Executive Summary

Live research for the week of May 3–9, 2026 surfaced two fully verified research items on the primary record (arXiv): a reasoning-model efficiency breakthrough from Zyphra (ZAYA1-8B, posted May 6) and a 30-author ICML 2026 position paper on Bayesian orchestration for agentic AI (updated in-week, May 6), plus an ICML-accepted theory paper on Bayes-consistency (May 5). Notably, the week's most loudly "promoted" alleged breakthroughs in aggregator blogs — a DeepSeek "Thinking with Visual Primitives" multimodal reasoning paper and a sensational "Claude Mythos" cybersecurity model — could not be verified against any primary record; the arXiv API returned zero results for the former, and no primary source exists for the latter. Any week-summary built on those claims would be built on unverified material.

Key Findings

1. Zyphra's ZAYA1-8B — small-MoE reasoning model that claims parity with much larger models (VERIFIED, high confidence)

• Posted May 6, 2026 on arXiv: https://arxiv.org/abs/2605.05365 (also dated May 6 in Sebastian Raschka's curated 2026 paper list: https://magazine.sebastianraschka.com/p/llm-research-papers-2026-part1)
• Claim: a reasoning-focused Mixture-of-Experts model with only 700M active / 8B total parameters (Zyphra's MoE++ architecture), pretrained end-to-end on a full-stack AMD compute platform, "matches or exceeds DeepSeek-R1-0528 on several challenging mathematics and coding benchmarks" and stays competitive with far larger open-weight reasoning models.
• Novel method: Markovian RSA, a test-time-compute technique that recursively aggregates parallel reasoning traces while carrying forward only bounded-length reasoning tails. Authors claim it lifts ZAYA1-8B to 91.9% on AIME'25 and 89.6% on HMMT'25 with only a 4K-token tail, "narrowing the gap" to Gemini-2.5 Pro, DeepSeek-V3.2, and GPT-5-High.
• Why it matters per authors: frontier-level reasoning from a sub-1B-active-parameter model trained on non-NVIDIA hardware — a direct challenge to the "compute is the only moat" assumption, and evidence that efficiency + test-time compute is where reasoning gains are being found.

…(truncated — the summary above captures the substance)

Round 0 · Finding 4

Most Significant AI Business Developments — Week of May 3–9, 2026

Based on live web research into daily AI news roundups, startup funding trackers, and venture analysis covering the target week.

Executive Summary

The week of May 3–9, 2026 was defined by three interlocking themes: (1) the largest compute-acquisition deal ever struck in AI, with Anthropic committing to rent SpaceX's entire Colossus 1 supercomputer; (2) a Pentagon AI contracting fight that pointedly excluded Anthropic while granting deals to SpaceX, OpenAI, Google, Microsoft, Nvidia, AWS, Oracle, and Reflection; and (3) an extraordinary burst of venture and revenue momentum across the sector — including ElevenLabs raising $500M, Anthropic topping OpenAI in quarterly revenue, OpenAI hitting $25B ARR, and Meta launching its Muse Spark product line.

Key Findings (with Confidence Levels)

1. Anthropic–SpaceX Colossus Compute Deal (HIGH confidence — reported May 6, corroborated by subsequent S-1 disclosures)

On May 6, 2026, Anthropic signed a deal to rent SpaceX's entire "Colossus 1" supercomputer — 220,000+ NVIDIA GPUs at 300MW capacity. Claude Code rate limits were doubled across all paid tiers the same day, signaling expanded compute capacity. Subsequent disclosures (SpaceX S-1, May 21) revealed Anthropic pays approximately $1.25 billion/month through May 2029 — a ~$45B total commitment, far above the $3–6B annual figure analysts had originally estimated (wikipedia-quality estimate ~$15B/year).

• Source: https://aitoolsrecap.com/Blog/ai-news-may-6-2026 and https://aitoolsrecap.com/Blog/ai-news-may-2026

2. Pentagon AI Contracts Exclude Anthropic (MEDIUM confidence — reported May 5)

The Pentagon signed AI agreements with SpaceX, OpenAI, Google, Microsoft, Nvidia, AWS, Oracle, and Reflection — pointedly excluding Anthropic amid ongoing litigation. This marks a notable business development given Anthropic's parallel $45B compute commitment to a SpaceX subsidiary.

• Source: https://aitoolsrecap.com/Blog/ai-news-may-5-2026 (referenced in May calendar)

3. Anthropic Topped OpenAI in Revenue (MEDIUM confidence — reported May 4)

Anthropic reportedly generated more revenue than OpenAI at this point, driven by its Claude Code product and enterprise deployments. This was later corroborated on May 11, when Anthropic disclosed Q1 revenue grew 80x year-over-year with ARR above $44B.

• Note: The 80x/ARR figures were disclosed May 11 (outside target week); the reversal of top-line standings was reported during the week.
• Source: https://aitoolsrecap.com/Blog/ai-news-may-4-2026 and https://aitoolsrecap.com/Blog/ai-news-may-11-2026 (calendar entry), https://aitoolsrecap.com/Blog/ai-news-may-2026

…(truncated — the summary above captures the substance)

Round 1 · Finding 1

The DeepMind blog returned empty content via browser rendering. Let me check the blog listing directly.

Round 1 · Finding 2

Findings: AI Policy, Regulatory, Safety, Export-Control, and Chip Infrastructure Developments — May 3–9, 2026

Executive Summary

I was tasked with filling the policy/regulatory/safety/export-control/hardware gap for the week of May 3–9, 2026. Across multiple search-engine queries (EU AI Act enforcement May 2026, NVIDIA export controls May 2026, US AI chip export controls China May 2026, AI governance policy White House May 2026, EU AI Act GPAI code of practice May 2026, plus site-specific queries), the available search tooling returned only evergreen, generic content — Wikipedia pages, vendor homepages, and dict/calendar entries — rather than dated news items from that specific week. A direct fetch of the NVIDIA Newsroom archive (nvidianews.nvidia.com/news) returned the live archive, whose visible top items were dated August 2026 (e.g., the Aug 10–11, 2026 NVIDIA/Apollo/BlackRock/KKR AI-financing announcements, Aug 11–12 Glasdoors/CEO and Nemotron 3.5 items); I could not page back to the May 3–9 window within my budget.

Net assessment: I could not identify, confirm, or corroborate any specific AI policy, regulatory, safety, export-control, or chip-infrastructure development dated to May 3–9, 2026. Every candidate I was asked to check (EU AI Act enforcement actions, US chip export-control changes, NVIDIA/TSMC hardware announcements, model safety incidents for the week) remains unverified. Per the research protocol, I will flag these explicitly rather than repeat aggregator claims as fact. The absence of hits is not proof that nothing happened — it reflects the limits of the search index available in this session — but I will not fabricate details to fill the gap.

Key Findings (with confidence levels)

1. No primary-record evidence found for week-dated policy/hardware events — LOW/MEDIUM confidence that significant events occurred but were missed; HIGH confidence that I could not verify any. All search-engine queries returned non-news, evergreen pages. No primary regulator (EU Commission/European Parliament, NIST, Commerce/BIS, White House) or chip company (NVIDIA/TSMC newsroom) week-dated item was surfaced in my fetch window. Confidence in the absence of found items: HIGH. Confidence that I captured the full picture: LOW.

…(truncated — the summary above captures the substance)

Round 1 · Finding 3

Resolution of the two contradictory claims

Claim 1 — "Anthropic released Claude Opus 4.7 on May 5, 2026 at a Wall Street finance event": NOT SUPPORTED — no such model exists in the record.

Claim 2 — "Claude Mythos, a restricted cybersecurity model, exists": CONFIRMED — it is real, but the cited AISI details (May 5 date, "GPT-5.5 matched Mythos") are wrong.

Findings

1. Claude Mythos is real and was publicly disclosed on April 7, 2026 — not a hoax, and not a May-week event.

• Anthropic maintains a product page for it: https://www.anthropic.com/claude/mythos (snippet dated Jul 1, 2026: "Claude Mythos 5 is the latest update for Mythos Preview, seeing gains in cybersecurity, biology, and healthcare…"). Anthropic also lists it in its system cards: https://www.anthropic.com/system-cards
• The existence of the model leaked March 26, 2026 (Fortune: https://fortune.com/2026/03/26/anthropic-leaked-unreleased-model-exclusive-event-security-issues-cybersecurity-unsecured-data-store/) and Anthropic disclosed it on April 7, 2026, launching "Project Glasswing" — a restricted-access consortium (40+ companies including Microsoft, Apple, Google, AWS, Nvidia, Broadcom, the Linux Foundation) that uses Mythos to find software vulnerabilities. Access was deliberately withheld from the public because of the model's offensive-cyber capabilities (NYT, Apr 7, 2026: https://www.nytimes.com/2026/04/07/technology/anthropic-claims-its-new-ai-model-mythos-is-a-cybersecurity-reckoning.html; WSJ: https://www.wsj.com/tech/ai/anthropic-set-to-preview-powerful-mythos-model-to-ward-off-ai-cyberthreats-75683cf5; Wired: https://www.wired.com/story/anthropic-mythos-preview-project-glasswing/).
• Reputable tech press coverage corroborates: BBC (Apr 17, 2026): https://www.bbc.com/news/articles/crk1py1jgzko; The Guardian (Apr 22, 2026): https://www.theguardian.com/technology/2026/apr/22/what-is-anthropic-mythos-ai-threat-global-cybersecurity; Forbes (Apr 8, 2026): https://www.forbes.com/sites/jonmarkman/2026/04/08/what-is-claude-mythos-and-why-anthropic-wont-let-anyone-use-it/; Scientific American (Apr 17, 2026): https://www.scientificamerican.com/article/what-is-mythos-and-why-are-experts-worried-about-anthropics-ai-model/; synthesized with references at https://en.wikipedia.org/wiki/Claude_Mythos

2. The UK AISI evaluation exists but is dated April 13, 2026 — not May 5 — and it does NOT say GPT-5.5 matched Mythos.

…(truncated — the summary above captures the substance)

Round 1 · Finding 4

Business Developments for May 3–9, 2026: Confirmation Assessment

Summary verdict

1. Pentagon AI procurement contracts excluding Anthropic — CONFIRMED (date corrected to May 1)

This is the week's real, verifiable business/defense story — and the only one of the five that survives contact with primary-adjacent reporting.

• SiliconANGLE, May 1, 2026 (updated May 3, 2026): "Pentagon inks AI procurement deals with seven companies, leaves out Anthropic." Full text fetched confirms: the U.S. Defense Department announced AI procurement contracts with seven firms — Amazon Web Services, Google, Microsoft, Nvidia, OpenAI, SpaceX, and startup Reflection AI — while Anthropic was "notably absent." Contracts cover Impact Level 6 and 7 (classified) environments via the GenAI.mil portal, which 1.3M+ DoD personnel have adopted. The article attributes Anthropic's exclusion to Defense Secretary Pete Hegseth's supply-chain-risk designation (March), which prohibits Pentagon use of Claude; Anthropic filed suit in March. https://siliconangle.com/2026/05/01/pentagon-inks-ai-procurement-deals-seven-companies-leaves-anthropic/
• Baltimore Sun, May 1, 2026: "US military reaches deals with 7 tech companies to use their AI on classified systems" (indexed via Google News RSS, fetched).
• TechCrunch (Feb 24, 2026), "Anthropic won't budge as Pentagon escalates AI dispute" (indexed via Google News RSS) — establishes the longer-running dispute that contextualizes the exclusion.

Caveats: the DoD announcement is dated May 1, 2026 (SiliconANGLE's URL and og:updated_time May 3), so the "May 5" date in the prior finding is off by days. I verified via reputable tech press, not the primary defense.gov release itself (not reachable in this session), so the primary record was not independently checked. The substantive claim — seven companies contracted, Anthropic excluded — is confirmed.

…(truncated — the summary above captures the substance)

Round 2 · Finding 1

Research, Open-Source, and Benchmark Developments — Week of May 3–9, 2026

Important caveat on primary sourcing

All items below are drawn from secondary/aggregator and practitioner sources; I could not reach a primary arXiv entry or vendor press release dated May 3–9, 2026 for most of these (the live arXiv API queried in real time returns August 2026 submissions, not May, and could not be date-restricted to the window in the session). Where an item appears week-dated only through aggregation, it is flagged accordingly.

1. DeepSeek "Thinking with Visual Primitives" — the standout research item of the week (CONFIRMED as a real DeepSeek work; week-date partially corroborated)

• DeepSeek-AI's "Thinking with Visual Primitives" paper is dated April 30, 2026 on arXiv, but it "gained massive traction on May 4, 2026", with multiple week-dated analyses published May 4–5 (https://www.mindstudio.ai/blog/deepseek-thinking-visual-primitives-5-technical-breakthroughs-paper; https://www.devflokers.com/blog/ai-news-may-2026-models-papers-open-source).
• The paper introduces a framework making spatial markers (points, bounding boxes) "minimal units of thought" interleaved into the Chain-of-Thought, to close the "Reference Gap" in dense-scene visual reasoning. Reported results: 66.9% maze navigation, 56.7% path tracing, 98.7% spatial reasoning vs. ~96–97% for frontier models; ~7,000× visual-token compression via Compressed Sparse Attention (CSA) (https://www.devflokers.com/blog/ai-news-may-2026-models-papers-open-source).
• Built on the DeepSeek-V4-Flash backbone (284B-parameter MoE, ~13B active) plus an in-house vision encoder; reportedly matches/beats GPT-5.4, Claude Sonnet 4.6, and Gemini 3 Flash on counting/spatial/topological benchmarks (https://www.jocheojeda.com/2026/06/24/deepseek-v4-vision-thinking-with-visual-primitives/).
• Pulled-paper confirmation: A June 2026 practitioner investigation confirms the official DeepSeek repository for this paper was "published, cited, and then quietly pulled" — the original repo currently 404s, with an archived community mirror retaining the technical report (MIT license) but no model weights; the paper text states the work is "to be integrated into the foundation model in the future" (https://www.jocheojeda.com/2026/06/24/deepseek-v4-vision-thinking-with-visual-primitives/; archived mirror at https://huggingface.co/datasets/NodeLinker/deepseek-ai-Thinking-with-Visual-Primitives-deleted-repo). Confidence: HIGH that the paper is real and was pulled; MEDIUM on exact pull timing relative to May 3–9.

…(truncated — the summary above captures the substance)

Round 2 · Finding 2

AI Model / Product / API Releases, May 3–9, 2026 — Findings by Vendor

Scope and method

I searched official vendor newsrooms and release notes (OpenAI, Anthropic, Google DeepMind, Meta AI, xAI, Mistral, Cohere) and cross-checked with dated aggregator roundups. Primary-source confirmed items are separated from secondary-source-reported items below. All figures and dates are cited; where I could not reach a primary record, I say so explicitly.

1. OpenAI — the only major vendor with confirmed week-dated product releases

GPT‑5.5‑Cyber (security-focused variant) — May 7, 2026 (PRIMARY-SOURCE CONFIRMED)

• OpenAI's own announcement page — "Scaling Trusted Access for Cyber with GPT‑5.5 and GPT‑5.5…" (https://openai.com/index/gpt-5-5-with-trusted-access-for-cyber/) and the related policy post "Trusted access for the next era of cyber defense" (https://openai.com/index/scaling-trusted-access-for-cyber-defense/) describe GPT‑5.5‑Cyber as a limited-preview cybersecurity variant under the Trusted Access for Cyber program.
• The May 7, 2026 date is independently confirmed by CNBC: "OpenAI rolls out new GPT-5.5-Cyber to vetted cybersecurity teams" (https://www.cnbc.com/2026/05/07/openai-rolls-out-new-gpt-5point5-cyber-to-vetted-cybersecurity-teams.html). Also corroborated by The Hacker News (https://thehackernews.com/2026/05/openai-launches-daybreak-for-ai-powered.html) and cybersecuritynews.com (https://cybersecuritynews.com/gpt-5-5-cyber/).
• Assessment: This is the single most significant, securely-dated OpenAI product event of the week — a new restrictive-access cyber model.

GPT‑5.5 Instant becomes ChatGPT's default — May 5, 2026 (SECONDARY-SOURCE REPORTED; not independently confirmed from OpenAI's primary announcement)

• The codersera changelog "OpenAI May 2026 Updates: GPT-5.5, Codex, GPT-5.6" (https://codersera.com/blog/openai-may-2026-updates-roundup/) states that on May 5, 2026 GPT‑5.5 Instant replaced GPT‑5.3 Instant as the default for all ChatGPT tiers, and that "ChatGPT for Excel and Google Sheets" reached general availability the same day. It reports a 52.5% relative hallucination reduction (18.7%→8.9% on OpenAI's "HallucinationBench") for the Instant variant.
• Caveat: The OpenAI newsroom page I fetched (https://openai.com/news/) only surfaced later (August 2026) entries in the returned content plus a "Load more" pattern, so I could not independently confirm the May 5 default-swap from OpenAI's own announcement page. Treat the exact date as unverified-at-primary-source, consistent with the secondary record.

Codex CLI v0.129.0 — May 7, 2026 (SECONDARY-SOURCE REPORTED)

• Same codersera changelog dates Codex CLI v0.129.0 (modal Vim editing) to May 7, 2026. Not confirmed from a primary OpenAI release-note fetch.

…(truncated — the summary above captures the substance)

Round 2 · Finding 3

Week-dated AI Business, Chip, and Policy Developments — May 1–9, 2026 (Primary-Source Verification)

Verification focus: (a) ElevenLabs $500M raise / annualized revenue, (b) Pentagon AI contract list, (c) chip/export-control/regulatory items dated that week. Note: several prior-round "week" claims are real events but misdated — I flag each with its verified date.

1. ElevenLabs ("$500M raise," May 8) — CONFIRMED as a May 5, 2026 event, but the figures differ from the claim

Confirmed (primary source). ElevenLabs' own blog post, "ElevenLabs crosses $500M ARR and welcomes new investors," is datePublished 2026-05-05 (https://elevenlabs.io/blog/500m-arr-and-new-investors). It states the company ended 2025 at $350M ARR and "in the first four months of 2026, we have already surpassed $500 million ARR." It announces the third close of the Series D with new investors BlackRock, Wellington, D.E. Shaw, Schroders, NVIDIA (via NVentures), Santander, Jamie Foxx, Eva Longoria, and Hwang Dong-hyuk, plus a $100M tender offer. So:

• The $500M figure is annualized run-rate revenue (ARR), NOT a "$25B annualized revenue" figure — the prior-round "$25B annualized revenue" appears to be a conflation (that $25B figure belongs to OpenAI, see §2). ElevenLabs' own confirmed metric is $500M ARR.
• No fresh "$500M raise on May 8" is confirmed. The large $500M Series D raise led by Sequoia at an $11B valuation was announced in early February 2026 (TechCrunch dated 2026/02/04, https://techcrunch.com/2026/02/04/elevenlabs-raises-500m-from-sequioia-at-a-11-billion-valuation/; CNBC 2026/02/04, https://www.cnbc.com/2026/02/04/nvidia-backed-ai-startup-elevenlabs-11-billion-valuation.html). The May 5 item is the third close of that existing Series D plus the ARR announcement. A TechCrunch item dated May 5, 2026 ("ElevenLabs lists BlackRock, Jamie Foxx and Eva Longoria as new investors," https://techcrunch.com/2026/05/05/elevenlabs-lists-blackrock-jamie-foxx-and-eva-longoria-as-new-investors/) corroborates the same day.

Verdict: The week-dated ElevenLabs development is real and primary-confirmed, but as $500M ARR + third Series-D close (May 5, 2026), not "$25B annualized revenue" and not "May 8."

2. OpenAI "$25B annualized revenue" (May 9) — NOT a this-week event; it dates to early March 2026

Confirmed date is March, not May. Reuters, "OpenAI tops $25 billion in annualized revenue … The Information reports," is dated 2026-03-05 (https://www.reuters.com/technology/openai-tops-25-billion-annualized-revenue-last-month-information-reports-2026-03-05/). US News (2026-03-04) carries the same "topped $25B annualized revenue last month" report (https://money.usnews.com/investing/news/articles/2026-03-04/openai-tops-25-billion-in-annualized-revenue-last-month-the-information-reports). No May 3–9 OpenAI revenue announcement was located. The prior-round flag of this as a May 9 item is a misdating.

…(truncated — the summary above captures the substance)

Round 2 · Finding 4

Resolution: AISI evaluation of Claude Mythos Preview — publication date and comparison model

Finding 1: Correct publication date is April 13, 2026 (not May 5, 2026)

The Wayback Machine's CDX index for https://www.aisi.gov.uk/blog/our-evaluation-of-claude-mythos-previews-cyber-capabilities shows the earliest capture timestamp is 20260413160726 (April 13, 2026, 16:07 UTC), followed by a continuous run of captures on April 13, 14, 15, 16, 22, 25, 27, and 28, 2026 (http://web.archive.org/cdx/search/cdx?url=www.aisi.gov.uk%2Fblog%2Four-evaluation-of-claude-mythos-previews-cyber-capabilities&output=json&limit=100). There is no Wayback capture on or near May 5, 2026 — the capture record jumps directly from 2026-04-28 to 2026-05-14. A page captured within hours of publication (April 13, 16:07 UTC) with an unbroken April archive trail cannot have been first published on May 5.

The May 14, 2026 snapshot, retrieved in full, displays the byline date "Apr 13, 2026" directly beneath the post's summary line (https://web.archive.org/web/20260514124354/https://www.aisi.gov.uk/blog/our-evaluation-of-claude-mythos-previews-cyber-capabilities). The current live page, fetched today (server date Aug 14, 2026), shows the identical byline date "Apr 13, 2026" (https://www.aisi.gov.uk/blog/our-evaluation-of-claude-mythos-previews-cyber-capabilities). The live page carries a last-modified: Mon, 10 Aug 2026 header indicating later edits, but the byline date and body content match the May 14 snapshot, so the later edit did not alter the publication date.

Conclusion: April 13, 2026 is confirmed. The "May 5, 2026" date is not supported by the primary record — no AISI blog capture exists for May 5, and no AISI publication of that date was found.

Finding 2: Correct comparison model is GPT-5.4 / GPT-5.3-Codex (not GPT-5.5)

Both the May 14, 2026 archived snapshot and the current live page name the comparison models in the Figure 3 caption of the "cyber range" results:

"Mythos Preview, Opus 4.6, and GPT-5.4 average 10 runs up to 100M tokens. Opus 4.5, GPT-5.1 Codex, and Sonnet 4.5 each average 15 runs up to 10M and 5 runs up to 100M tokens. GPT-5.3-Codex averages 10 runs up to 10M and 5 runs up to 100M tokens. Sonnet 3.7 and GPT-4o average 10 runs up to 10M tokens only."

The Figure 1 caption additionally names GPT-3.5 Turbo through Claude 4 Opus and "GPT-5 through to Mythos Preview." The string "GPT-5.5" appears nowhere in either the archived May 14 version or the live version of the post. The closest GPT-5-family comparators in the evaluation are GPT-5.4 and GPT-5.3-Codex, plus GPT-5.1 Codex in the CTF/range charts.

Conclusion: The comparison models in the AISI Mythos evaluation are GPT-5.4 and GPT-5.3-Codex (with GPT-5.1 Codex also charted). The "GPT-5.5" attribution is refuted for this document — it does not occur in the primary source in either its May 2026 archived form or its current form.

…(truncated — the summary above captures the substance)

Round 3 · Finding 1

AI Policy, Safety, and Legal Developments: May 3–9, 2026

Executive Summary

A targeted sweep of regulator sites, safety-body publications, and court/legislative reporting found no confirmed AI policy, safety, or legal events dated within May 3–9, 2026 beyond the May 1 DoD contract agreement (previously documented in the prior round). The White House AI executive order that press coverage references was announced and then postponed May 20–21, 2026, outside the search window. UK AISI published no evaluation during the target week. The Anthropic copyright settlement was finalized in July 2026, not in-window.

Key Findings by Category

1. US White House / Executive Branch — No in-window events CONFIRMED

The prior round's finding that no in-window US AI executive action is supported by this sweep. The most relevant proximate event came weeks later:

• An executive order on advanced AI innovation and security was ultimately signed in June 2026 (https://www.whitehouse.gov/presidential-actions/2026/06/promoting-advanced-artificial-intelligence-innovation-and-security/), and Trump postponed a separate AI security executive order on May 21, 2026 (https://www.cnn.com/2026/05/20/tech/ai-executive-order-trump-white-house, https://www.politico.com/news/2026/05/21/white-house-postpones-ai-executive-order-signing-00931904, https://techcrunch.com/2026/05/21/trump-delays-ai-security-executive-order-i-dont-want-to-get-in-the-way-of-that-leading/). Both fall outside the May 3–9 window.

2. EU AI Act / European Commission — No in-window enforcement actions CONFIRMED

Search results surfaced the EU's Code of Practice on Transparency of AI-generated Content and the General-Purpose AI Code of Practice (https://digital-strategy.ec.europa.eu/en/policies/code-practice-ai-generated-content, https://digital-strategy.ec.europa.eu/en/policies/contents-code-gpai), but these predate/postdate May 2026. Law-firm analyses (e.g., Jones Day on the "final code of practice on AI labelling" dated June 2026, https://www.jonesday.com/en/insights/2026/06/european-commission-publishes-final-code-of-practice-on-marking-and-labelling-aigenerated-content; Cooley noting transparency obligations take effect August 2, 2026, https://www.cooley.com/news/insight/2026/2026-08-03-eu-ai-act-transparency-obligations-take-effect-2-august-2026) place the operative events after the target week. No EU AI Act enforcement action dated May 3–9, 2026 was found.

3. UK AI Safety Institute (AISI) — No in-window publications CONFIRMED

The AISI research/publications index (https://www.aisi.gov.uk/research) lists publications chronologically. The items bracketing the target week are:

• Apr 29, 2026 — A Decision-Theoretic Formalisation of Steganography With Applications to LLM Monitoring
• May 14, 2026 — Automated alignment is harder than you think
• May 21, 2026 — Loss of Oversight: How AI systems may become harder to audit

…(truncated — the summary above captures the substance)

Round 3 · Finding 2

Chinese Lab and Open-Source AI Developments — Week of May 3–9, 2026

Scope and Exclusions

This investigation focused on (a) Chinese labs (Alibaba/Qwen, Moonshot, Zhipu, MiniMax, ByteDance/Seed, Baidu/ERNIE) and (b) open-source ecosystem items (Hugging Face trending, OpenClaw, A2A, MCP, open-weight model releases), dated May 3–9, 2026. Notably, the dominant Chinese release events of May 2026 — Alibaba's Qwen 3.7-Max (May 20) — fell outside this window. Only one flagship Chinese model release, ERNIE 5.1, falls squarely inside it.

1. CONFIRMED IN-WINDOW: Baidu ERNIE 5.1 (released May 8, 2026)

Status: Confirmed. Primary source is Baidu's own ERNIE blog at https://ernie.baidu.com/blog/posts/ernie-5.1-0508-release/ (the URL itself carries the 05-08 date marker), and multiple independent secondary sources corroborate the May 8 date.

What it is: ERNIE 5.1 is Baidu's flagship model, the first Chinese model to crack the global top 10 on LMArena's Search Arena leaderboard — landing at #4 globally with an Elo of 1223, and #14 on the standard text leaderboard (the highest position any Chinese lab has held there) (https://codersera.com/blog/ai-model-releases-may-2026-roundup/).

Efficiency profile: ERNIE 5.1 compresses total parameters to roughly one-third of ERNIE 5.0 and active parameters to one-half, spending only 6% of the pre-training compute of comparable frontier models (https://chatforest.com/reviews/baidu-ernie-5-1-moe-frontier-llm-review/; https://www.baidumarketing.com/blog/baidu-ernie-5-1-release). On AIME26 with tool use it scores 99.6 — second only to Gemini 3.1 Pro.

Availability: 128K context window; API access and direct chat at ernie.baidu.com, priced via Baidu's Qianfan platform. The preview had been quietly live since April 30 (https://codersera.com/blog/ai-model-releases-may-2026-roundup/).

Confidence: High (confirmed by primary Baidu source and multiple independent trackers).

2. CONFIRMED IN-WINDOW: Subquadratic SubQ 1M-Preview (launched May 5, 2026)

Status: Confirmed for date; model architecture claims UNVERIFIED.

Subquadratic, a Miami startup (founded by ex-SoftBank Vision Fund and Tinder alumni, $29M raised at a $500M valuation), launched SubQ on May 5, 2026 with a claimed 12-million-token context window (~10× larger than any frontier model in production) (https://codersera.com/blog/ai-model-releases-may-2026-roundup/). This is the first "fully subquadratic" context model — selecting a small subset of positions per query token. Claimed performance includes a 7.2× speedup at 128K tokens and 92.1% on needle-in-a-haystack at 12M tokens.

…(truncated — the summary above captures the substance)

Round 3 · Finding 3

Notable AI papers, model releases, and acceptance announcements — week of May 3–9, 2026

Bottom line: The week's research output is now documented from primary and archival routes. The arXiv 2605.0xxxx ID range is confirmed real and populated; Hugging Face's daily paper feeds for May 3–8, 2026 are archived and recoverable; and — importantly — ICML 2026 acceptances were not announced in this window (notifications went out April 30, 2026, per ICML's own dates page). No Nature- or NeurIPS-dated item for May 3–9 was found. The most significant items of the week are open-weight model reports and agent/search-agent papers, led by NVIDIA's Nemotron 3 Nano Omni and the academic search-agent paper OpenSeeker-v2.

1. ICML 2026 acceptance announcements — REFUTED for this window (primary record)

The ICML 2026 official Dates and Deadlines page states "Author Notification: Apr 30 '26 (Anywhere on Earth)" — i.e., acceptances were released the week before the target window, not during May 3–9. Main conference dates are July 7–9, 2026 (https://icml.cc/Conferences/2026/Dates). The only ICML-adjacent in-window deadline is the workshop-level "Universal notification deadline for all submissions to individual ICML workshops: May 14 '26," also outside the window. The accepted-papers list itself is live at https://icml.cc/virtual/2026/papers.html (an aggregator, https://aiconfpaper.com/conferences/icml-2026, counts 6,634 accepted papers), and the ICML blog's awards post is dated July 5, 2026 (https://blog.icml.cc/2026/07/05/announcing-the-icml-2026-awards/) — both post-window artifacts. Verdict: no ICML 2026 acceptance announcement occurred May 3–9; the week's ICML-relevant news was limited to community coverage of the April 30 notification.

…(truncated — the summary above captures the substance)

Round 3 · Finding 4

Verification Report: OpenAI May 5–7, 2026 changelog items

Verdict: Both items are CONFIRMED by primary sources.

1. "GPT-5.5 Instant" became the default ChatGPT model on May 5, 2026 — CONFIRMED

Primary source (OpenAI's own announcement): The official page https://openai.com/index/gpt-5-5-instant/ is dated May 5, 2026 (the dateline appears directly on the page) and titled "GPT‑5.5 Instant: smarter, clearer, and more personalized." Its Availability section states verbatim: "GPT‑5.5 Instant is rolling out starting today to all ChatGPT users, replacing GPT‑5.3 Instant as the default model, and in the API as chat-latest. For paid users, GPT‑5.3 Instant will remain available for three months, accessible through model configuration settings, before being retired." The page also contains a later "Update on June 9, 2026" note about personalization expanding to Free/Go tiers — but the May 5 default-model switch is unambiguously the original, dated event. (Verified by fetching the live page; metadata/canonical confirm the URL.)

Reputable secondary corroboration: TechCrunch published https://techcrunch.com/2026/05/05/openai-releases-gpt-5-5-instant-a-new-default-model-for-chatgpt/ with datePublished: 2026-05-05T17:00:00+00:00 (byline "10:00 AM PDT · May 5, 2026"), opening: "On Tuesday, OpenAI released a new foundation model called GPT-5.5 Instant, which will replace GPT-5.3 Instant as the default ChatGPT model." (May 5, 2026 was indeed a Tuesday, consistent.) It reports the same API detail ("chat-latest," 5.3 kept for paid users for three months) and the AIME 2025 score of 81.2 vs. 65.4.

Help-center changelog / Wayback angle: The official ChatGPT release-notes page (https://help.openai.com/en/articles/6825453-chatgpt-release-notes) has Wayback captures bracketing the release — May 3, 2026 (capture 20260503135946, HTTP 200) and May 10, 2026 (capture 20260510132637, HTTP 200), with a bot-blocked capture in between on May 8 (20260508053536, HTTP 403). This confirms the changelog page existed and was archived both before and after the May 5 date. Caveat: I could not directly quote the changelog entry's text — the May 10 raw replay (via id_ format) returned gzip-compressed bytes that did not decode inline, so the help-center entry itself was not readable in this session. The item is nonetheless confirmed on the strength of OpenAI's own dated announcement page, which is the stronger primary record.

2. "Codex CLI v0.129.0" released on May 7, 2026 — CONFIRMED

…(truncated — the summary above captures the substance)

Investigation Trail

Round 0

• Which major AI models or model releases (e.g., from OpenAI, Anthropic, Google DeepMind, Meta, xAI, Mistral, or Alibaba) were announced or launched during the week of May 3-9, 2026, and what are their headline capabilities and benchmark results?
• What were the most significant AI business developments during the week of May 3-9, 2026 — including major funding rounds, acquisitions, partnerships, compute/data-center deals, or executive changes at leading AI companies?
• What major AI policy, regulatory, or legal developments emerged during the week of May 3-9, 2026 — e.g., actions by the U.S. government, EU AI Act enforcement, Chinese AI rules, court rulings, or international agreements affecting AI development and deployment?
• What notable AI research breakthroughs or benchmark results were published or promoted during the week of May 3-9, 2026 — e.g., in agentic AI, reasoning, multimodal models, model efficiency, or safety — and what do the authors claim about their significance?

Round 1

• Did Anthropic release Claude Opus 4.7 on May 5, 2026 at a Wall Street finance event, and does a restricted cybersecurity model called 'Claude Mythos' exist? Check Anthropic's newsroom/social accounts, the May 5, 2026 UK AISI report, and reputable tech press for both.
• Which of these reported business developments for May 3-9, 2026 can be confirmed by primary sources or reputable tech press: Pentagon AI contract list excluding Anthropic (May 5), ElevenLabs' $500M raise (May 8), OpenAI's $25B annualized revenue (May 9), Anthropic surpassing OpenAI in revenue (May 4), and OpenAI-Anthropic acquisition talks?
• What AI policy, regulatory, safety, export-control, and chip infrastructure developments occurred during May 3-9, 2026 (e.g., EU AI Act enforcement, US chip export controls, NVIDIA/TSMC announcements, model safety incidents)?
• What notable AI research papers and model announcements appeared during May 3-9, 2026 from arXiv (cs.AI/cs.LG/cs.CL/cs.CV), Google DeepMind, Meta FAIR, NVIDIA, and ICML 2026 accepted-paper announcements?

Round 2

• Which AI models, products, or API features did OpenAI, Google DeepMind, Meta AI, xAI, Mistral, and Cohere announce or release between May 3 and May 9, 2026? Search official blog/newsroom pages with date filters (e.g., OpenAI newsroom May 2026, Google DeepMind blog May 5 2026).
• What is the correct publication date — April 13 or May 5, 2026 — and the correct comparison model — GPT-5.5 or GPT-5.4/GPT-5.3 Codex — for the UK AISI evaluation of Claude Mythos previewing cyber capabilities? Use the archived AISI blog and publications index via the Wayback Machine.
• What notable AI research papers, open-source model releases, and benchmark launches appeared in the week of May 3–9, 2026 — e.g., on arXiv, Nature, Hugging Face trending models, or major conference proceedings?
• From primary sources, which week-dated AI business, chip, and policy developments for May 1–9, 2026 are confirmed: (a) ElevenLabs' reported $500M raise and $25B annualized revenue; (b) the Pentagon/DoD contract naming seven AI companies without Anthropic; (c) any NVIDIA, TSMC, AMD, Intel, EU AI Office, U.S. Commerce-BIS, or NIST chip/export-control/regulatory announcements dated that week?

Round 3

• Do OpenAI's official help-center changelog, release notes, or archived snapshots (Wayback Machine captures from around May 5–7, 2026), plus official OpenAI X/Twitter posts, confirm that 'GPT-5.5 Instant' became the default ChatGPT model on May 5, 2026 and that 'Codex CLI v0.129.0' was released on May 7, 2026? Provide direct URLs or archived captures for confirmation, or state clearly if the items cannot be confirmed.
• Which notable AI papers, model releases, or acceptance announcements were published the week of May 3–9, 2026, as found through arXiv ID ranges around 2605.0xxxx, Semantic Scholar or OpenAlex date-filtered searches, Google Scholar with date bounds, alphaxiv, or Hugging Face paper listings — including any ICML 2026 acceptance announcements or Nature/NeurIPS-adjacent items dated in that window?
• What AI model releases or open-source developments came from Chinese labs (Alibaba/Qwen, Moonshot AI, Zhipu AI, MiniMax, ByteDance/Seed, Baidu/ERNIE) and from open-source ecosystems (Hugging Face trending models, OpenClaw, the A2A agent standard, MCP) during the week of May 3–9, 2026?
• Which AI policy, safety, and legal developments occurred during May 3–9, 2026 — e.g., EU AI Act code-of-practice updates, UK AISI evaluation publications, White House AI executive orders or actions, court rulings in AI-related lawsuits (such as those involving Anthropic), or regulatory announcements reported around May 6, 2026?

Sources

• https://theaiforest.com/ai-regulation-news-2026-us-eu-global-updates/
• https://www.themodernblog.com/ai-regulation-news/
• https://www.gunder.com/en/news-insights/insights/2026-ai-laws-update-key-regulations-and-practical-guidance
• https://www.morganlewis.com/pubs/2026/05/the-new-rules-of-ai-a-global-legal-overview-may-2026
• https://techinsightedge.com/insight/ai-regulation-news/
• https://astraea.law/insights/federal-ai-regulation-landscape
• https://www.hungyichen.com/en/insights/ai-governance-regulatory-landscape-2026
• https://cubbbix.com/blog/ai-regulation-news-2026-global-update/
• https://beyondtmrw.org/article/ai-regulation-update-2026-eu-ai-act-enforcement-and-us-state-rules
• https://aiunpacking.com/guides/ai-regulations-worldwide-global-overview-2026/
• https://axis-intelligence.com/eu-ai-act-news/
• https://axis-intelligence.com/eu-ai-act-enforcement-guide/
• https://techjacksolutions.com/eu-ai-act-news/may-2026/
• https://www.mayerbrown.com/en/insights/publications/2026/07/eu-ai-act-news-digital-omnibus-on-ai-new-guidance-on-risk-classification-gpai-and-transparency-obligations
• https://www.technology.org/2026/07/17/eu-ai-act-what-actually-applies-on-2-august-2026/
• https://presenc.ai/research/eu-ai-act-enforcement-tracker-2026
• https://www.cnbc.com/2026/08/03/eu-ai-act-enforcement-powers.html
• https://www.cooley.com/news/insight/2026/2026-08-03-eu-ai-act-transparency-obligations-take-effect-2-august-2026
• https://smartaiforbiz.com/eu-ai-act-news-2026-high-risk-rules-delayed-what-every-business-must-know-now/
• https://commission.europa.eu/news-and-media/news/safer-and-more-transparent-ai-2026-08-02_en
• https://en.m.wikipedia.org/wiki/May
• https://simple.m.wikipedia.org/wiki/May
• https://www.almanac.com/content/month-may-holidays-fun-facts-folklore
• https://www.calendarr.com/united-states/the-meaning-of-may-everything-you-need-to-know-about-the-fifth-month/
• https://www.timeanddate.com/calendar/months/may.html
• https://funworldfacts.com/facts-about-may/
• https://www.holidaysmart.com/calendars/month/may
• https://www.britannica.com/topic/May-month
• https://facts.net/may-facts/
• https://openai.com/
• https://gemini.google.com/
• https://chatgpt.com/
• https://ai.google/
• https://deepai.org/
• https://www.perplexity.ai/
• https://duck.ai/
• https://copilot.microsoft.com/?msockid=276691dd38c763b42ea3866839116237
• https://gemini.google/us/about/?hl=en
• https://en.wikipedia.org/wiki/Artificial_intelligence
• https://copilot.microsoft.com/?msockid=33f4cac4137c603214f3dd71128e61f6
• https://en.wikipedia.org/wiki/China
• https://www.geocountries.com/china
• https://www.britannica.com/place/China
• https://theworldfactbook.org/country/china.html
• https://www.scmp.com/news/china
• https://www.britannica.com/place/China/State-and-society
• https://www.reuters.com/world/china/
• https://www.worldatlas.com/maps/china
• https://www.chinahighlights.com/travelguide/china-facts.htm
• https://www.cnn.com/world/china
• https://www.devflokers.com/blog/latest-ai-models-open-source-projects-may-2026
• https://aitoolsrecap.com/Blog/ai-news-may-2026
• https://web.archive.org/web/20260719210226/https://www.aicritique.org/us/2026/06/01/ai-developments-in-may-2026/
• https://codersera.com/blog/ai-models-released-may-2026-monthly-roundup/
• https://codersera.com/blog/ai-model-releases-may-2026-roundup/
• https://www.digitalapplied.com/blog/ai-model-releases-may-2026-complete-tracker
• https://www.aicritique.org/us/2026/06/01/ai-developments-in-may-2026/
• https://models.evertune.ai/
• https://www.vtnetzwelt.com/ai-development/latest-ai-technology-news-roundup-may-2026/
• https://aitoolsrecap.com/blog/ai-model-releases-may-2026-what-to-expect
• https://blog.google/innovation-and-ai/technology/ai/google-ai-updates-may-2026/
• https://frontierwisdom.com/ai-updates-2026-agentic-open-models/
• https://www.iteache.com/resources/new-ai-tools-launched-may-2026
• https://best-ai.news/ai-model-releases-2026
• https://blog.google/innovation-and-ai/technology/ai/google-io-2026-all-our-announcements/
• https://whatllm.org/blog/new-ai-models-may-2026
• https://copilot.microsoft.com/?msockid=2b33b76df5b96be405aca0d8f4696a0e
• https://arxiv.org/abs/2605.05365
• https://magazine.sebastianraschka.com/p/llm-research-papers-2026-part1
• http://export.arxiv.org/api/query?search_query=all:%22Bayes-consistent%22&max_results=15
• https://www.devflokers.com/blog/ai-news-may-2026-models-papers-open-source
• http://export.arxiv.org/api/query?search_query=ti:%22Thinking+with+Visual+Primitives%22&max_results=5
• https://www.news.tunx.ai/ai-research-this-week-the-most-important-papers-breakthroughs-and-benchmark-results-explained-2026/
• https://www.devflokers.com/blog/ai-tech-breakthroughs-may-2026-developments
• https://deepmind.google/research/publications/
• https://arxivlens.com/research/weekly-summaries
• https://www.nature.com/natmachintell/articles?year=2026
• https://realifeai.com/ai-breakthroughs-in-2026/
• https://www.cognizant.com/us/en/ai-lab/blog/ai-research-updates-may-2026
• https://sakana.ai/ai-scientist-nature/
• https://arxiv.org/abs/2601.12538
• https://arxiv.org/pdf/2601.12538
• https://www.aimodels.fyi/papers/arxiv/agentic-reasoning-large-language-models
• https://www.researchgate.net/publication/399930413_Agentic_Reasoning_for_Large_Language_Models
• https://sungsoo.github.io/2026/02/05/agentic-reasoning.html
• https://github.com/VoltAgent/awesome-ai-agent-papers
• https://github.com/weitianxin/Awesome-Agentic-Reasoning
• https://papers.cool/arxiv/cs.AI
• https://huggingface.co/papers/2601.12538
• https://deepseek.com/en/index.html
• https://www.deepseek.com/
• https://deep-seek.com/chat
• https://chat.deepseek.com/sign_in
• https://platform.deepseek.com/
• https://deep-seek.ai/chat
• https://en.wikipedia.org/wiki/DeepSeek
• https://deep-seek.online/
• https://deepseek.net/
• https://download.deepseek.com/
• https://aitoolsrecap.com/Blog/ai-news-may-6-2026
• https://aitoolsrecap.com/Blog/ai-news-may-5-2026
• https://aitoolsrecap.com/Blog/ai-news-may-4-2026
• https://aitoolsrecap.com/Blog/ai-news-may-11-2026
• https://aitoolsrecap.com/Blog/ai-news-may-9-2026
• https://aitoolsrecap.com/Blog/ai-news-may-8-2026
• https://aitoolsrecap.com/Blog/ai-news-may-7-2026
• https://inforcapital.com/blog/2026-05-09-ai-startup-funding-surges-in-may-37-deals-and-25-billion-as-investors-double-down-on-machine-learning/
• https://aitoolsrecap.com/Blog/MayNews2026.aspx
• https://news.crunchbase.com/venture/biggest-funding-rounds-june-5-2026/
• https://money.usnews.com/investing/news/articles/2026-05-05/openai-anthropic-ventures-in-talks-to-buy-ai-services-firms-sources-say
• https://www.ai-market-watch.com/news/category/acquisition
• https://techfundingnews.com/category/acquisition/
• https://blog.mean.ceo/ai-startup-funding-news-may-2026/
• https://af.net/realtime/ai-funding-rounds-2026-live-deal-tracker-updated-daily/
• https://techfundingnews.com/
• https://aifundingtracker.com/ai-startup-funding-news-today/
• https://aifunding.me/insights/ai-startup-funding-explained-complete-2026-guide
• https://www.joineta.org/blog/ai-investment-roundup-august-2026
• https://aifundingtracker.com/
• https://www.blockchain-council.org/ai/ai-funding-news-2026-records-mega-rounds-trends/
• https://aifunding.me/deals
• https://www.fundedstartupsdaily.com/raises/ai-ml/
• https://www.ai-market-watch.com/news/category/funding
• https://awaira.com/funding-rounds
• https://copilot.microsoft.com/?msockid=08e49cc1db83699122468b74dab268fd
• https://arxiv.org/
• https://arxiv.org/login
• https://en.m.wikipedia.org/wiki/ArXiv
• https://info.arxiv.org/about/index.html
• https://info.arxiv.org/about/spinout_faq.html
• https://arxiv.gg/
• https://blog.arxiv.org/
• https://www.alphaxiv.org/
• https://tech.cornell.edu/arxiv/
• https://www.reddit.com/r/PhD/comments/10k2w7f/arxiv_what_is_it/
• https://www.google.com/
• https://about.google/
• https://www.google.com.nf/webhp?gl=nf&hl=en&gws_rd=cr&pws=0
• https://about.google/products/
• https://maps.google.com/
• https://www.youtube.com/google
• https://ogs.google.com/widget/empty
• https://blog.google/
• https://accounts.google.com/
• https://icml.cc/
• https://www.icmlonline.com/
• https://icml.cc/virtual/2026/papers.html
• https://www.icmlonline.com/exams/Default.aspx?p=MLE1
• https://info.lubecouncil.org/
• https://en.wikipedia.org/wiki/International_Conference_on_Machine_Learning
• https://dl.acm.org/conference/icml
• https://openreview.net/group?id=ICML.cc/2026
• https://openreview.net/group?id=ICML.cc/2026/Conference
• https://info.lubecouncil.org/membership-4/
• https://www.linkedin.com/company/openai
• https://www.reddit.com/r/OpenAI/
• https://www.youtube.com/@OpenAI
• https://openai.com/index/chatgpt/
• https://www.cnbc.com/2026/08/11/longtime-openai-executive-brad-lightcap-leaves-as-shakeup-at-ai-lab-continues.html?msockid=009bfe5fab286e3d05b0e9eaaa4f6fe6
• https://www.coursera.org/articles/what-is-openai?msockid=009bfe5fab286e3d05b0e9eaaa4f6fe6
• https://chatgpt.com/overview/
• https://platform.openai.com/
• https://nvidianews.nvidia.com/news
• https://nvidianews.nvidia.com/news/nvidia-partners-with-apollo-blackrock-blackstone-brookfield-goldman-sachs-and-kkr-to-establish-ai-compute-infrastructure-financing-platforms-to-mobilize-over-500-billion-of-third-party-capital
• https://blogs.nvidia.com/blog/800-vdc-power-architecture-ai-factory/
• https://copilot.microsoft.com/?msockid=0e197417cb2b6f78377e63a2ca736e2d
• http://www.nvidia.com/page/home.html
• https://www.nvidia.com/en-us/
• https://en.wikipedia.org/wiki/Nvidia
• https://play.geforcenow.com/
• https://marketplace.nvidia.com/en-us/
• https://www.nvidia.co.uk/Download/indexsg.aspx?lang=en-us
• https://nvidianews.nvidia.com/
• https://investor.nvidia.com/home/default.aspx
• https://www.investing.com/equities/nvidia-corp
• https://copilot.microsoft.com/?msockid=29e1dc2c5a9c643c3c60cb995b7a65ce
• https://en.m.wikipedia.org/wiki/European_Union
• https://european-union.europa.eu/index_en
• https://european-union.europa.eu/principles-countries-history/eu-countries_en
• https://en.m.wikipedia.org/wiki/Member_state_of_the_European_Union
• https://www.britannica.com/topic/European-Union
• https://commission.europa.eu/index_en
• https://worldpopulationreview.com/country-rankings/european-union-countries
• https://de.m.wikipedia.org/wiki/Europ%C3%A4ische_Union
• https://www.europarl.europa.eu/portal/en
• https://www.investopedia.com/terms/e/europeanunion.asp
• https://www.cnbc.com/2026/08/11/longtime-openai-executive-brad-lightcap-leaves-as-shakeup-at-ai-lab-continues.html?msockid=388473fb9daf68d410f8644e9cca695d
• https://www.coursera.org/articles/what-is-openai?msockid=388473fb9daf68d410f8644e9cca695d
• https://www.anthropic.com/claude/mythos
• https://www.anthropic.com/system-cards
• https://fortune.com/2026/03/26/anthropic-leaked-unreleased-model-exclusive-event-security-issues-cybersecurity-unsecured-data-store/
• https://www.nytimes.com/2026/04/07/technology/anthropic-claims-its-new-ai-model-mythos-is-a-cybersecurity-reckoning.html
• https://www.wsj.com/tech/ai/anthropic-set-to-preview-powerful-mythos-model-to-ward-off-ai-cyberthreats-75683cf5
• https://www.wired.com/story/anthropic-mythos-preview-project-glasswing/
• https://www.bbc.com/news/articles/crk1py1jgzko
• https://www.theguardian.com/technology/2026/apr/22/what-is-anthropic-mythos-ai-threat-global-cybersecurity
• https://www.forbes.com/sites/jonmarkman/2026/04/08/what-is-claude-mythos-and-why-anthropic-wont-let-anyone-use-it/
• https://www.scientificamerican.com/article/what-is-mythos-and-why-are-experts-worried-about-anthropics-ai-model/
• https://en.wikipedia.org/wiki/Claude_Mythos
• https://www.aisi.gov.uk/blog/our-evaluation-of-claude-mythos-previews-cyber-capabilities

Trace Index

Tool-call traces are persisted under /srv/swarm_web_runs/run-1786712828394-0003/traces.