Corrections and updates

Added Microsoft's out-of-band .NET updates for CVE-2026-111497, the GitHub-reviewed advisory GHSA-9v8h-qqgm-4m35 and CyberScoop's report that no exploitation has been observed.

Anthropic made its S-1 public on EDGAR on October 26, for a planned Nasdaq listing under the symbol ANT.

Added the October 26 release describing a binding commitment for an approximately $40 billion GPU lease with an entity managed by Apollo Global Management, and how its wording compares with Elon Musk's October 25 post.

Added Apple's October 26 announcement that Siri AI is now in beta in five more languages with the 27.2 updates, and what Apple says remains unavailable or limited.

Added OpenAI's October 20 update to its election post, which includes a new rule for election-related images, and CNET's October 26 test of it.

Added Moonshot's apology and quota reset for subscribers, independent rankings from Artificial Analysis and Arena, and unverified reports of a creator brief.

Added Jensen Huang's remarks at GTC Berlin on what Nvidia charges for its agent software, as reported by Tom's Hardware.

The PHP Foundation describes how small its staff is and says it has no evidence of exploitation before disclosure or on patched sites, and OpenAI's chief executive says patching alone can no longer keep up.

Added Elon Musk's statement that SpaceX has closed the $40 billion debt financing, and what Reuters and CNBC say remains unclear.

Researchers now describe botnet attempts against PHP2Shell and public exploit code that mostly fails, and Mandiant's CTO says there is no current evidence of successful widescale exploitation but that limited exploitation may have begun.

A state-media-affiliated account says DeepSeek and MiniMax have completed registration under the draft rules and been cleared for overseas use.

Reuters reports that Anthropic is pressing ahead with a mid-November IPO, with marketing to investors expected to begin the week of November 9.

Security firms report exploitation attempts but no confirmed compromise from PHP2Shell, and Sophos has corrected an earlier report of a successful exploitation.

MiniMax has published M3.1 Flash's weights on Hugging Face under its own community licence.

The Nasdaq fell 2.6% and the Philadelphia Semiconductor index 6.7% on Thursday, October 22, and SK hynix said on Friday that Solidigm has halted its IPO review.

CNBC reports that OpenAI is close to finalizing a round of up to $30 billion at a $1.25 trillion pre-money valuation, led by investors from the UAE and Coatue.

Meta's preliminary information statement, filed with the SEC on October 21, sets out how the Class C shares were approved and how Zuckerberg's own shares are restricted.

Wiz says credentials stolen through the malicious release were used within hours to check for access to Amazon Bedrock.

LiteLLM's own incident post says its Docker Hub and GHCR images were also poisoned and sets out how the release was published.

Japanese banks paused some online services as a precaution, South Korea issued a cyber "Caution" alert, and the PHP Foundation and Chainguard gave conflicting accounts of how the embargo broke.

Amazon says PHP2Shell is now being broadly exploited (Amazon's claim, unverified as of October 23; see the dated note at the end of this section) and that its honeypots logged an attempt before disclosure; CISA has told organisations that…

The KOSPI fell 3.52% on Thursday as chipmakers gave up the previous day's gains.

Yahoo Finance reported that Anthropic could file its S-1 publicly as soon as Friday, October 23.

The KOSPI closed 3.29% higher and Elice ended at its daily upper limit.

The Nasdaq ended Tuesday flat; Seoul rebounded nearly 3% in early Wednesday trade.

The week closed with a chip bounce on October 19, then Asia resumed selling on October 20.

Reuters reports OpenAI is racing to lock in the round before Anthropic's mid-November IPO disclosures.

Bloomberg reports talks on a second roughly $45 billion Nscale deal, and Microsoft has exited a UK Nscale site over power delays.

The declaration Handelsblatt reported was signed at the Invest in Germany summit.

Added four reports first posted by OpenAI on October 15 about June activity.

CoreWeave rebounded on October 15 while chip stocks slipped again, and the KOSPI opened higher on October 16.

Apple executive chairman Tim Cook said Apple Intelligence is approved in China and a test version is coming later in October.

Anthropic said on X that it has committed $5 million.

Anthropic described a METR arrangement in which it says it cannot veto publication.

The KOSPI closed down 7.57% after its first circuit breaker since late July.

Further investor-day figures, attributed to the Financial Times and The Wall Street Journal, were relayed by CNBC and Business Standard.

The FTC has sent civil investigative demands to OpenAI and Anthropic.

OpenAI agreed in court not to use the trade secrets Apple says were taken while the case proceeds, according to Bloomberg's reporting.

OpenAI has filed counter-allegations in Apple's lawsuit.

Archive review — October 14, 2026

The archive was revised to align with our editorial standard. The articles listed here changed in that review only in ways that did not correct a stated error, such as removing editing notes left in the text, adding internal links, or adding a source. Factual corrections from the same review are listed separately.

An earlier version made overly broad statements that no AI bill had been tabled or legislation proposed.

The KOSPI rebounded more than 2% in early trade on Wednesday after five sessions of losses.

Removed an unverified 'first structured outside test' claim and an unsupported description of parental consent; ChatGPT for Teens is described as OpenAI's under-18 product with parental controls.

Corrected active parameters to 52B and replaced an unattributed October 27 weights date with TechCrunch's 'three weeks'.

Attributed the OpenAI link to Wikimedia's belief and corrected when OpenAI's rogue-agent disclosures began (August, not this month).

The opening comparison is now attributed to AMI, rather than presented as an independent ranking.

Removed an unsourced $1B ad run-rate reference and 'ARR' wording.

Sourced Intel's share move to Business Insider and removed unverified premarket and project-cost figures.

Changed 'OpenAI's largest backer' to 'one of OpenAI's largest backers'.

Removed claims that Beam was not on Artificial Analysis and an unsupported 'most serious U.S. attempt yet' superlative; noted that the comparison table is vendor-compiled.

Added the CNBC report (citing the Journal and Trump's Truth Social post on an “AI Force”) and ODNI context for Clayton's role.

Corrected Robinson's role: three and a half years at OpenAI, where he led launch safety reports.

Narrowed a Limits note to the Muse Gadgets product page actually checked, rather than all of Meta's published pages.

We removed an unsupported description of the 64GB configuration as a cheaper tier.

We clarified that the $189.9 million announced at Armadin’s March launch was combined seed and Series A funding.

Removed an unverified 'first clear' characterization of the subpoena.

Removed an unverified ~$915 million purchase-price figure.

Replaced 'ARR' with the company's 'revenue run rate' and removed an unsourced task-count/score clause.

Narrowed the Artificial Analysis #1 claim to streaming transcription accuracy (final and first-partial transcripts), as Microsoft stated it.

Corrected attribution: the January 2026 founding date and three-month prototype come from the Nebius/EQS release; Ctech supplied headcount, seed lead and leadership context.

Removed an unverifiable response deadline and an unsupported 'second state AG' framing.

Dated and scoped the statement that the safety bodies had not responded.

Described The Next Web as relaying SoftBank's release rather than independently confirming it.

The summary now clarifies that independent benchmark performance has not been established for this article.

Removed an unverified claim that the package sets a de facto national compliance floor for employers and deployers with a California nexus, from the summary and the compliance section.

Clarified that deployment and first-expansion claims are Clio’s account, and that the financial-terms statement describes what its announcement discloses.

Stated the investor's name plainly (Valor Equity Partners; TechCrunch spelled it 'Valar') and removed editing notes left in the text.

Separated the Post's 'a few weeks ago' timing from Reuters' report that Ferguson had concerns before the Hugging Face hack; removed an unsourced description of the White House accord and its 'morally binding' quote; described Reuters as…

Limited 'without cyber guardrails' to trusted Fairwind defenders and removed claims that no Arena ranking existed.

Attributed the Art.

Removed an unsupported 'first public attribution' and 'industrial-scale' framing; OpenAI's attribution to Moonshot-associated individuals is described as qualified.

Described GPT-Synopsys as a model to be trained, not an existing one, and described Converge Digest as reporting that cites the companies rather than independent corroboration.

Removed an unsourced Senate committee status and Commerce draft-rule claim, and a mismatched cost comparison.

Added the directly inspected Business Wire release, corrected the access disclosure, and attributed reliability and deployment claims to ZoomInfo.

Corrected the read-only research description (follow-up actions go through the dot's normal action rules), removed an unverified claim about EU and UK rules, and corrected two internal links.

Removed unverifiable claims that Arena had not ranked the model and an unsourced line about Astra Ultrafast availability; fixed internal links.

Removed an unsourced reference to a Salesforce–Fin deal and an internal drafting note left in the text.

Replaced the inaccessible full-text reprint with the archived accord itself, removed unsupported political-causality and implementation claims, and narrowed the legal characterization to the document and attributed reporting.

Corrected “matches” in the headline to “nears”: AA measured 56 for Sonnet versus 58 for Opus.

Corrected the reference to a former employee and a current board member, removed unsupported product-week comparisons, and disclosed that the motion itself was inaccessible.

This corrected account relies on the accessible September 29 TNW report.

Corrected when the prime minister was informed, distinguished possible other government sites from health sites, and removed the unsupported first-ever framing.

Removed unverified comparative context, identified the demo report as a September 25 update, and replaced an unsupported refusal-to-disclose characterization with what the announcement actually omits.

The summary, opening and SEO description now say the White House, OpenAI and Anthropic did not immediately respond to Reuters, rather than that none confirmed on the record.

The full OpenAI announcement identifies Airbnb CTO Ahmad Al-Dahle as the speaker behind the 80% feature-shipping statement.

Removed the unsupported headline claim that the Council 'left with none' and an unverified claim about a standard; noted that Amodei spoke by video call, per the Guardian; attributed the slowdown remark to Altman alone; removed unverified…

The summary now attributes to Anthropic the statement that no one had previously tied the enzyme, repeats and accessory protein together.

Corrected the OSWorld version to 2.1, separated later independent checks from the September 22 launch, and removed an unverified historical Arena listing and mixed-effort speed/cost comparison.

Corrected the Limits note: OpenAI's post claims an internal model resolved Navier–Stokes and more than 100 open problems without publishing proofs.

Removed an unpreserved historical Arena snapshot and an unverified secondary AA score.

Clarified that the arrangement and investments were announced plans, and that Anthropic was also discussing evaluator-funded nonprofit pilots.

The launch post announced a future default-model change.

Anthropic clarified on September 30 that the linked form records an expression of interest rather than an application; this article now reflects that clarification.

Attributed the 'first Australia lease' headline to Reuters reporting.

Write approval is the documented default, not an unconditional guarantee.

Attributed the Artificial Analysis ranking claims to Google in the summary, SEO and body, and renamed the 'Independent checks' heading.

Corrected the $1.5 trillion figure: per NYT (via The Next Web) it is OpenAI's own ask against investors' ~$1.2 trillion proposals, not a conflicting report; sourced the WSJ revenue figures through StockAnalysis; changed 'operating margin…

Removed the claim of independent corroboration and additional legislative and executive-reaction details that could not be verified in the accessible sources.

Salesforce’s no-customer-data statement concerns training.

Removed unsupported comparative “last” and waitlist claims.

Sourced the FT profit details to its Irish Times syndication; removed an unsourced claim that Anthropic had a confidential S-1 on file since June, plus the related 15-day rule line; corrected the source for Altman's no-2026-IPO remark…

Corrected the summary of Amodei's three steps to match the essay: export controls and anti-distillation belong to step 2 (pacing within democracies); step 3 is global pacing, which he says requires cooperation with China.

Strengthened attribution for company allegations and clarified that disrupted biology cases were not all blocked before model assistance; preserved the explicit no-harmful-intent caveat.

Privacy and residency details are attributed to documentation inspected October 14; removed an unsupported SDK launch-month comparison.

Clarified that Christiano joins the Foundation Board; his non-voting observer status applies to the separate Group PBC board.

The company’s announcement itself confirms the valuation and lead investors; the previous secondary-only attribution was incorrect.

The opening no longer says two labs claimed solutions to millennium-class problems (Fermat's Last Theorem is not a Millennium Prize Problem).

Clarified that some reinforcement-learning work resumed, that 3.1× describes the evaluated agent runtime rather than production-wide researcher output, and that March 2028 is an automated-researcher target rather than a product launch date.

The primary announcement describes planned investment of up to ₹70,000 crore and capacity of up to 1 GW, developed in phases according to demand; it does not promise 1 GW in the first phase.

Removed an unverified quotation and later-event comparison.

Attributed the ~18,000-post count and the May 11–July 2 window to the researchers' report; corrected who said OpenAI knew earlier (two people familiar, via Reuters); quoted OpenAI's statement verbatim; removed an unsupported claim about…

Only selected surface variables use a 5 km grid.

Cited NVIDIA's 8-K for the deal terms and the first-half-2027 close, changed 'Buys' to 'to Buy', and removed unverified claims about other bidders and past valuations.

Removed an unverified letter analogy and personal attribution, narrowed the access statement to documented cyber gates, and avoided claiming unverified leaderboard absence.

Removed the unsupported 'third Flash in six weeks' and 'Critical-tier' framing, the comparison with 3.7 pricing, and an unverified named-partner list.

Removed unsupported launch-day executive quotes and the unverified September 4 max-availability date; distinguished the current company page from the dated launch evaluation.

EFS was announced for a phased fall rollout, not opened generally on launch day.

Described Nvidia as holding the lease (per The Wall Street Journal, via DCD), not as landlord; corrected DCD's attribution to the WSJ; removed an unverified $19.6 billion figure and timestamp; turned an internal reference into a link.

Corrected the claim about production models: Anthropic says they are 'unlikely' to engage in the more extreme behavior, not that they did not.

Corrected the Mac revenue line to 'nearly 29% to $10.4 billion', per The Decoder.

Removed a like-for-like comparison between annualized run rate and an unverified full-year target, along with unsourced quotes and regulatory context.

Removed unsupported posting and distribution-agent claims.

Removed unsourced statements attributed to Cursor's Michael Truell (OpenAI ~5% of traffic) and Anthropic's Tom Brown.

Clarified that the 10,000-seat allocation covers both free Standard and discounted Premium seats; removed an unsourced IPO-timing inference.

Clarified that QuEra’s 99.3% result came from a finished deterministic script in a later blind test, not a live agent recovery rate; corrected the roadmap/publication distinction.

Specified that the 60% speed claim measures system throughput for 360p versus 720p, as Google’s footnote states.

Attributed the absence of deadlines, pledges and metrics to Axios and removed an unverified timing link to the Hugging Face post-mortem.

Added the official model card, clarified additional embedding/MTP parameters and open-weight licensing, and removed an unverified training-cost ratio and historical API price.

Removed unverified campus totals, revenue forecasts and later deal comparisons.

Clarified that the additional GPUs and Trainium connectivity are planned deployments, not completed installations.

The launch announced public-preview APIs, not a Chirp 3 retirement.

The earnings-call transcript places Rubin’s approximately 20% Data Center share in the Q3 outlook, not Q2 results.

Limited the no-customer-impact statement to OpenAI, corrected the denominator of the 93% figure, and distinguished the independent review’s scope.

Sourced Grok 4.6's Foundry price to Microsoft's blog and removed an unverified claim that it matches the xAI API and Bedrock prices.

Removed an unsupported Broadcom co-development/June unveiling claim and an unsupported claim that SemiAnalysis verified runs on-site.

Restored OpenAI’s “very likely” qualification to geographic attribution in the headline, summary and table.

The subpoena was dated and served August 20, before the August 24 announcement.

Attributed Groq's early-adopter and Dell plans to Groq's blog, and corrected the CNBC 'later this year' timeline, which refers to the racks at Nebius.

Narrowed the headline to the verified launch and attributed cost result; the vendor post does not specify the comparison baseline for its 82% claim.

Preserved the SpaceXAI name used in Nvidia’s dated announcement and clarified planned deployment; removed unsupported earlier launch and LPX comparisons.

Removed an unverified name/role detail, specific charges and a 'first known indictment' claim; the indictment is described as being over alleged illegal export.

Removed the unverified staff-hiring claim and Poolside comparison; retained the reported equity talks and distinguished previously considered licensing from an agreed structure.

Narrowed the headline and analysis to the reported configuration-dependent increases; removed an unsupported blanket claim about all rack budgets and the sole shipment bottleneck.

Corrected the grant description ('a small number of larger pilot grants') and narrowed the partner claim to integrations Anthropic says are in progress.

Corrected the test-settings line and removed an unsourced image-format list and claims that no independent listing existed.

Linked and checked the original policy post; removed an unsupported attribution of a specific incident and admission of failure.

Corrected the platform name and nine-day launch interval; removed unsourced Workspace, competing-model and later Foundry claims.

Corrected the beta's plan list (SuperGrok Heavy, Cursor Ultra and Cursor Teams Premium), removed an unverified app version number and a reference to a later event, and turned an internal reference into a link.

Removed unsupported dated comparisons with separate products; retained the verified Platform availability and HIPAA eligibility limits.

Removed unverified competitor/app-usage comparisons and attributed deployment examples and download totals to Google.

Removed an unverified end-2027 distribution claim, an unconfirmed Information/WSJ corroboration and the Groq date; said the Groq deal ($20B) was larger; described the 109 staff as receiving job offers.

Limited the Pixel AI Pro bundle to Pixel 11 Pro phones and removed an unsourced comparison with OpenAI's teen and education push.

Corrected Marvell's share move to Reuters' 'nearly 8%' intraday and removed unverified claims about a Broadcom selloff and Alphabet's close.

Distinguished the August preview from the source’s September rollout update and removed unverified spokesperson/comparator details.

Removed unverified contract and historical platform claims; the announcement supports model availability, supported regions, context, effort and pricing.

Corrected the launch date of the terminal Grok Build to May 25, 2026.

Removed unsupported claims about the trust’s intended purpose and unrelated financial context.

Removed references to an August 22 TechCrunch piece that postdates this August 18 article.

Corrected the description of 31 European markets: the list includes non-EU countries.

Clarified that Study Hours is configurable rather than universally on by default; removed unsourced litigation and unrelated product context.

Removed an unverified later billing statement; the 20% figure describes OpenAI’s internal monitoring estimate.

Added the direct SEC filing and its guarantee conditions; removed unverified negotiation and revenue estimates.

Changed the headline to describe an agreed acquisition rather than a completed one, corrected the Series B lead investor (CapitalG), and removed a DealBook quote that could not be found and an unverified newsroom-confirmation claim.

Removed unverified capacity, payment-split and later-close claims; corrected the credit-support scale and restored Reuters’ verification caveat.

Marked SWE-bench Pro and OSWorld figures as vendor-run with harness differences, removed an unsupported characterization of the Max preview, and turned an internal reference into a link.

Corrected the reporting period (from the February 2026 prior report) and added Anthropic's statement that the biological-classifier gap was remediated with no evidence of concerning misuse; removed an unverified 'first time' claim.

Clarified provider-versus-deployer obligations, removed the unsupported “first obligation/every sentence” claim, and added the dated September 1 private-preview detector update.

Attributed the May $10.9B/$559M projection to The Wall Street Journal (via Engadget), and removed an unsupported comparison with OpenAI's run rate.

Corrected the toggle's status to rolling out, removed an unverified claim about OpenAI and Meta watermark practice, replaced a paraphrase that was presented as a quotation, and turned an internal reference into a link.

Removed a ~$100B figure attributed to The Information, whose report was not read; the headline and body now use the WSJ figure of under $120B for a first phase, confirmed through Reuters' syndicated report.

Replaced unsupported regulatory and newsroom-absence assertions with the reporting’s explicit deal and comment limits.

Removed an unsupported characterization from the opening.

Removed an unverified June 22 date for IBM's Daybreak Cyber Partner Program participation.

Removed the generic forum citation and an unsourced training-hardware comparison.

Clarified that the employee rollout and AI workflows are announced plans under the five-year partnership, not verified completed deployments.

Kept AA comparisons explicitly attributed to the vendor table and removed the unsupported claim that this review independently verified a dated ranking.

Used the primary release’s rounded backlog, clarified non-GAAP EBITDA and delivery conditions, and removed unverified consensus, duration-mix and cross-deal figures.

Removed separately unverified historical user counts, product-order numbering and cross-company comparisons.

Removed undated Artificial Analysis scores from the summary, SEO description and checks section; corrected the LangChain partner figures (74% lower cost on 145 tasks, 7% of calls routed to a frontier model, 6% accuracy tradeoff); turned…

Attributed the 'shake-up' framing to CNBC.

Clarified that this was an additional international expansion, after earlier announced Canada/Australia/New Zealand pilots; Go is paid but ad-eligible.

Removed unsourced cross-launch chronology.

Linked Anthropic’s existing February electricity pledge and distinguished it from a newly specified Theseus contract term.

Attributed the open-weight policy framing to Reuters and the Aug 5 cyber-eval breach to The Information; turned an internal reference into a link.

The August 10 news was an announcement and waitlist.

Removed later distribution and unrelated Astra comparisons that were not supported by this launch source.

Corrected the migration guidance: passwords can be imported into Chrome from an exported CSV, and cookies and sessions do not transfer.

Replaced the unsupported 15–18 MW range with Firebird’s 15 MW facility figure and described inauguration remarks without implying every speaker attended in person.

Clarified that the surface-level total fallback reductions are Anthropic’s expected effects in a footnote, while the 85% biology figure comes from its testing.

Clarified that Free/Go unlimited text and the Think button were scheduled for the following week, and removed unsupported claims about historical rankings carrying over.

Removed an undated Arena Elo figure and unverified Artificial Analysis scores and comparisons; the summary now says lab coding benchmarks lead the post.

Corrected the comparison with OpenAI's Hugging Face incident (its agent exploited a novel vulnerability; Meta's and Anthropic's cases came from testing-partner mistakes), replaced unverified exact dates with the wire timing, and removed…

The order’s 30-day access window precedes release to other trusted partners, not necessarily public launch.

Attributed the on-device adaptation figures (a few hours of data, or ~200 examples) to Ars Technica's report, and attributed full-body-control claims to Google as reported.

Replaced 'speculative decoding' with OpenAI's stated 15%+ token-generation efficiency gain; removed an unverified Auto-review migration and ~10× claim, unsupported corroboration language, and unattributed competitive-pressure claims.

Removed unverified input pricing, launch-infrastructure and historical ranking claims; retained vendor benchmarks with their source and caveats.

Removed unsupported watermarking, backend-replacement and competitor-comparison claims.

Replaced the unverified historical count of 1,224 with Reuters’s dated “more than 1,100” figure, separated personal signatures from company responses, and removed unsupported roadmap assertions.

Corrected the timing of the House kill-switch bill relative to the letter, and replaced general context with the Lieu–Moran bill's DHS graduated emergency powers; corrected an internal link.

Described DHS's authority as graduated orders after a covered incident, as POLITICO reported the text; corrected the related-provisions row and an internal link.

Distinguished the planned Cyber pilot from immediate access and removed unsupported historical leaderboard comparisons.

Removed the unsupported historical-first assertion.

Removed unsupported historical firstness; the announcement describes a focused rare-disease call.

No independent leaderboard snapshot was verified, so the previous empty-board assertion is withdrawn.

Corrected the timing to days (not weeks) after Kimi K3, softened a coding-board superiority claim, and removed unverified groundwork and corroboration claims; same-day outlets syndicated Bloomberg rather than independently confirming it.

Re-dated from July 16 to July 17, the date of the Artificial Analysis post and the WAIC opening this article relies on.

Described SynthID as being adapted for DNA synthesis screening, not as a deployed screen; removed an unverified comparison with OpenAI's safety work and a reference to a PDF that had not been checked.

Removed unrecovered Text Overall, peer Elo and vote-count snapshots.

Removed unverified publication-week comparisons and secondary-source placeholders; the robustness findings remain attributed to OpenAI.

Replaced an unsourced industry-wide ownership comparison with Anthropic’s documented governance structure.

Removed the default-authentication description, earlier-preview date and an unverified same-day model-launch comparison.

Removed an unverified 'billions of users' figure and unverified prior-partner claims; dated Meta's Arena ranking claims to July 5, before public launch.

This article was originally dated March 6, 2026, and described court documents as unsealed in early March.

Removed an unsupported historical-first claim and later plan-change context; access details are attributed to Anthropic’s June 30 post and July 1 update.

Clarified that analysis context is sent to Claude even when datasets remain on lab infrastructure; removed unrelated launch-day context.

Narrowed the access loss to parts of the NSA, per Nextgov, which reported more limited use may remain; removed an unverified description of analysts' reactions and added the Nextgov report to Sources.

Removed an unverified 'first congressional oversight' claim from the headline and an unverified Washington Post characterization.

Narrowed attribution to OpenAI’s assessment of likely Chinese origin and added its limited-impact finding; removed unverified wire corroboration.

Added AA’s evaluation effort and measured fallback rates; removed unverified retrospective Arena rankings and export-control chronology from the launch account.

Corrected the source access date and removed unverified market commentary about S-1 timing.

Added inspected Microsoft source links and qualified inconsistent Voice 2 language/Flash availability descriptions across Microsoft posts.

Verified the July 9 expansion against its dated release-note entry, added its regional limits, and removed an unsupported comparison of preview and later service-level agreements.

Added the primary source for OpenAI’s $852 billion comparison and removed unsupported subsequent IPO context.

Removed an unsourced reference to earlier May previews; the verified announcement establishes the May 25 beta and subscriber access.

Removed an unverified Big Four alignment claim and unverified wire-confirmation language; reframed the pilots and lock-in points as EY and Microsoft's pitch.

Corrected the Spark description: Google says it rolls out to trusted testers first, with email and chat access coming later.

Changed availability language to the announcement’s planned Data Platform connection and exploratory AI Factory integration; fully local inference was not established.

Corrected the Tomoro headcount to about 150 forward-deployed engineers and deployment specialists, the investment to more than $4 billion per OpenAI, the syndicate roles, Tomoro's cited clients, and the acquisition status (agreed, not…

Corrected the default-tier description and partner list to OpenAI's Daybreak page, and replaced unverified Codex Security plugin details with OpenAI's description of Codex as the agentic harness.

Removed unsourced cross-launch context and secondary corroboration claims; pricing and model roles are checked against OpenAI’s announcement.

Replaced unsupported traffic/session assertions with OpenAI’s stated user reach and clarified test-versus-deployment reasoning effort.

Replaced an editing note left in the text with a limit: the exact test conditions have not been checked against the labs' system cards.

Qualified the export window and removed unsupported causal and partnership claims.

Removed unsupported base-retraining conclusions and retrospective Arena figures; distinguished OpenAI’s Terminal-Bench 2.0 result from AA’s Terminal-Bench Hard.

Clarified that the incident account is Bloomberg's report, relayed by TechCrunch and Fortune with Anthropic's statement.

Clarified that case postures come from McKool Smith's April 20 tracker and that the underlying filings were not reviewed for this article.

Clarified that Claude Platform on AWS was coming soon, matching Anthropic’s April 21 clarification; the agreement did not make it generally available on April 20.

Corrected the false independent-confirmation claim and the inference that the NSA was using Mythos for vulnerability scanning or exploitation.

Corrected the reward-hack descriptions to the paper's wording (the most-frequent-answer shortcut on math; writing and executing unit tests on coding; exfiltrating labels via repeated scoring) and removed the unsupported 'disqualified'…

Rewrote the China, U.S. and Russia rows to match the New York Times report (read via the Indian Express syndicated copy), and removed unsupported references to Anthropic's Mythos Preview and lab cyber warnings from the summary and opening.

The meeting account is Washington Post reporting, not a company primary record or independently corroborated account.

Qualified the archive-bias statement to describe a possible limitation rather than an established English-language bias.

Corrected the reward-hacking steering figures to Anthropic's reported range (−0.1 to +0.1 steering; ~5% to ~70%; calm suppression ~65%, strong calm ~10%).

Added the primary on-device/edge source and removed unverified licensing-firstness wording.

The four disclosed rounds total $188B, approximately 62.7% of the stated $300B total.

Made company attribution explicit and clarified that clinician training remains necessary; the release does not establish mortality benefits.

Clarified the proposed construction/upgrading scope and the bill’s alternative facility-definition tests; the proposal is not an enacted moratorium.

Eligibility and expiry details are checked against current help documentation.

Added the framework’s stated exceptions to federal preemption.

Clarified that nano was API-only, specified xhigh for the cited mini results, and distinguished the March 17 announcement from March 18 release notes.

Corrected the complex count to EMBL's 30 million and the EMBL announcement date to March 16.

Specified which output-length comparison applies to Qwen on AA LCR and which applies to GPT-OSS 120B on LiveCodeBench.

Rewrote the article around Google Research's Groundsource announcement.

Restored pre-event chronology and removed retrospective NemoClaw and record-attendance claims from this March 10 preview.

Corrected the memo's terms to CBS News's report (removal within 180 days; exemptions only for mission-critical work where no viable alternative exists), attributed the Iran detail to CBS's sources, and removed sources that had not been…

Corrected the scope: NVIDIA's COUPE-based co-packaged optics are for its Quantum-X and Spectrum-X Photonics network switches, not joint AI chip designs.

Rewrote the Politico summary to its text: Congress has largely skirted the issue, Republicans are reluctant to buck Trump or hand China an edge, and most bills focus on training and layoff data.

Removed unsupported claims about one-week mastery and comparable-task power savings.

Corrected the designation timeline: Hegseth announced it on X, and Anthropic says it received a confirming letter on March 4.

Replaced blanket Pro-performance claims with the launch post’s qualified positioning and exact availability and prices.

Removed unverified historical Arena scores and later mini/nano launch context.

Re-dated from March 1 to March 2, the date Qwen's GitHub release log gives for the availability of the four Qwen 3.5 Small models.

Clarified launch access restrictions and separated Google’s benchmark claim from unverified historical rankings.

Declaration confirmation comes from a February 21 government update, after this article’s original date.

Attributed the 8.6×–19× throughput figures to Alibaba and named the Qwen3-Max baseline and context lengths; removed an unverified blog timestamp and an unsupported Reuters attribution; corrected the pretraining description and removed an…

Qualified the license claim against the current official model-license text, which is dated February 13; the February 12 launch date remains unchanged.

Linked the official public memo and replaced the inaccurate secondary-only/private-memo framing.

The inspected sources support Ascend deployment and training reproduction, not exclusive original training.

Removed unsupported historical Arena rankings, qualified the GDPval-AA result as reported by Anthropic, and added preview and long-context pricing limits.

Specified that the High designation concerns cybersecurity and added the vendor benchmark reasoning-effort setting.

Added the reported exchange ratio and executive cash option; removed the unsupported claim that exchange ratios were undisclosed.

Added the dated primary source for the later February 9 test; the original January 16 announcement preceded live advertising.

Corrected the date and attribution of the InfoQ coverage.

Added the primary acquisition announcement and removed an unverified claim of three-outlet corroboration.

Calcalist reports a lower valuation for the proposed deal.

Korean chip selling continued for a fifth day, with the KOSPI down more than 6% on Tuesday, and the IMF warned about an AI correction.

The opening paragraphs were reworded: the first now attributes the comparison with CAISI and Anthropic’s case for wider defender access to Anthropic’s own post, and a later sentence states plainly that Anthropic’s figures come from its…

The October 9 afternoon section previously credited Maas Group's fall of nearly 11% to Reuters' report on the withdrawal.

CrowdStrike has published a forensic account of the attacker's tooling, ARTEX's developer has withdrawn the source code, and Korea's financial regulator has given a victim count.

Travel booking site Skyticket has disclosed that up to 14.64 million customer records may have been exposed.

Bloomberg reports that OpenAI expects its annualized revenue to reach or exceed $70 billion by the end of the year.

Bloomberg reports that Firmus is exploring a private round of up to US$3 billion, and that its Batam partner DayOne is still working on a bond.

Firmus has withdrawn its ASX listing application and says it will pursue private capital.

The three researchers have now named themselves and set out their side in a letter, and OpenAI has answered on the record.

A new preprint argues that OpenAI’s Lean formalisation of its September Navier–Stokes blow-up claim does not match the written proof, and a new group calling itself the Association for Human Mathematics has urged mathematicians to stop…

Firmus’s IPO documents were reportedly withdrawn from the data room used by fund managers, and bankers were reported to be considering pulling the deal.

Firmus’s IPO book-build has reportedly closed without enough demand at A$11, and bankers were reportedly working to save the float at about A$8.25 a share, according to reports citing unnamed sources.

Two October 6 reports pick up the liability thread this piece left open.

Firmus and CDC Data Centres have confirmed that their Project Southgate partnership, the Australian rollout that was to reach 1.6 GW, is over at about 42 MW.

Reflection showed its cards.

The September 29 report now has names attached.

Sunday’s reported ~$20 billion talks became a signed deal.