OpenAI’s GPT-6 Cyber joins Google’s Gemini 4 Carbon in a restricted rollout, with stronger reported security capabilities and access conditions that prevent treating either as an ordinary public model launch.
Anthropic’s amended IPO filing reports $16.956 billion of Q3 revenue and a $567 million operating loss, with future compute commitments carrying a different scale and timetable.
Anthropic Q2 and Q3 2026, respectively: revenue $11.555B and $16.956B; GAAP operating income +$0.974B and loss −$0.567B. Each is a three-month period. Revenue and operating result use separate scales. Unaudited; neither operating result nor commitments is cash burn.
Revenue rose 47% from Q2; the operating result turned negative.
Unaudited three-month results. Operating result is not cash flow; future commitments are outside this comparison.
New tools, considered
METR estimates that growth in Anthropic’s internal capability index accelerated 1.9 times, but the cause and the most informative supporting evidence remain partly uncertain or private.
A 1.9× point estimate and 1.5–2.8 interval on a zero-to-three scale, with a separate 1× historical reference.
multiple of earlier growth rate
METR’s observed index-growth comparison
Bounds: 1.5–2.8 × earlier growth rate
Point estimate; no bounds shown.
95% confidence interval for index growth; not a causal AI effect or individual productivity.
AWS, Azure and Google Cloud reported strong September-quarter revenue growth, but company-wide cash flow and future capacity obligations show why cloud sales alone cannot establish the return on AI spending.
AWS revenue grew 39.3% to $46.0B, Azure 48% to $33.165B, and Google Cloud 103% to about $30.7B in the three months ended September 30, 2026. Bars share a zero-to-110-percent scale. Business scopes differ and include non-AI; Google Cloud includes TPU-system sales. These are not market shares.
September-quarter growth within each disclosed business.
$46.0B recognized revenue
$33.165B recognized revenue
About $30.7B recognized revenue
Year-on-year reported growth. Scopes differ and include non-AI; Google Cloud includes TPU-system sales. Not market shares.
Mistral Large 4 and Kimi K3.1 Pro now have inspectable weight repositories, but their licences draw materially different boundaries around model-serving businesses.
Conductor says its cloud service will become read-only November 13 before shutdown and deletion begin November 20, while OpenAI separately announced an agreement to acquire Pierre’s developer tools.
A proportional November 4–20 timeline marks November 9, November 13 read-only and November 20 shutdown/deletion. An export window spans the period before shutdown.
Cloud enters read-only mode November 13, then shuts down and begins deletion November 20.
Calendar dates only; no cutoff time or timezone was published. Finish before November 20.
Cursor can now provision an app, database, authentication and hosting, but publication still requires passing checks and an explicit publish action; a separate voice release lets users supervise agents conversationally.
Cognition’s SWE-2 Flash and Devin Review 2.0 reduce the advertised cost of everyday engineering work, while public model weights and independently verified review savings remain unavailable.
Thinking Machines launched Inkling 2 Small through Tinker with weights still planned, while Ricursive’s Apache-2.0 Ripple-65B Preview offers an immediately inspectable alternative for self-hosting experiments.
Images 2.6 adds markup-guided edits, SVG and GIF output, while Gemini Omni 1.2 Flash extends video to fifteen seconds and conversational revision; the useful test is how much work survives into a finished asset.
Perplexity launched ongoing Teammates for Max subscribers and a separate fully local Computer mode for supported Macs, creating different choices about delegation and where work runs.
Claude Enterprise’s inference hooks are generally available, while its security center and centralized management remain beta and sign-out does not revoke running cloud work or API keys.
Claude message threads reduce repeated context uploads without reducing token accounting, while OpenAI’s Agents API adds mutable tools, background compaction and separately billed automatic review.
Ai2’s StateBench finds that four tested frontier models can retrieve simple facts yet fall to 13% accuracy or less on the longest dependency-heavy state-tracking sequences.
Artificial Analysis’s Decision Index v2 adds document-level completion and calibrated confidence to the buying question, showing why per-answer accuracy and token price alone are incomplete measures.
Four bars on a 0–100 scale: wrong at 90% confidence 9.5, wrong at 60% 32, right at 60% 92, right at 90% 99.5.
The Decision Score combines correctness and confidence.
Calculated examples using 50 × (1 + c − (p − c)²); no measured model results.
DeepSeek’s public super-agent prototype turns authorized GitHub or Feishu requests into separate work items, with host-enforced permissions and an explicit warning that it is research code.
ERNIE 5.5 Flash Base gives developers a substantial pretrained starting point, but it is not an instruction-tuned assistant and its licence excludes the EU, UK and South Korea.
Preferred Networks released PLaMo 3 Translate in 0.8B, 2B, 8B and 31B sizes, offering a practical local translation option with conditional commercial licensing.
Four blue horizontal bars on a common zero-to-31-billion-parameter scale: 0.8, 2, 8 and 31 billion.
PLaMo 3 Translate spans 0.8B to 31B parameters.
Parameter counts are not translation-quality scores or memory requirements.
Nemotron 3.5 Transcribe offers downloadable speech-model weights and per-request control over transcript conventions, including speaker labels and timestamps that do not establish a speaker’s real identity.
PSI reports a Lean-checked proof of the Aaronson–Ambainis conjecture, a major result about quantum queries on typical inputs whose scope does not extend to all quantum algorithms.
Anthropic’s November 2 ledger records 5,087 disclosed findings and 629 known patches, while a separate n8n advisory gives workflow operators a concrete isolation vulnerability to fix.
A proportional ring distinguishes 629 known patched from 4,458 with unconfirmed patch status. It does not measure installed upgrades.
disclosed findings
Anthropic disclosure ledger, November 2
Whole: 5,087 findings
Cumulative disclosures; unconfirmed status is not proof of no patch. An upstream patch is not installed fleet coverage.
The UK’s October 28 Budget commits £40 million to a National Agentic Security Unit and an additional £120 million to the AI Security Institute, while NCSC’s new cyber scenarios stress uncertainty about future attacks.
Two bars share a pounds-million scale: 40 for the agentic unit and 120 additional for AISI. These are distinct commitments with different stated time scopes.
£ million
October 28 Budget
2026–27 through 2029–30
Commitments, not spending or measured effectiveness. AISI funding spans 2026–27 to 2029–30; no identical period is asserted for the unit.
DOJ announced its Super Intelligence Task Force on November 5 under an October 13 memorandum, while an October 30 FOIA response withheld 60 pages concerning federal AI coordination and review.
OpenAI and Anthropic support a common robots.txt signal for AI-training preferences, but their implementation timelines differ and the signal relies on crawler compliance.
Menlo estimates a $176.8 billion annualized enterprise-AI deployment market, while Deloitte finds only 44% of surveyed organizations track per-unit inference costs; neither result proves investment returns.
Deloitte surveyed 3,500 AI-involved leaders across 25 countries in May–June 2026. 71% report tracking total inference/API spend, 44% cost per inference unit. Bars share a 0–100% scale. Separate questions do not establish a nested subgroup or causal economic return.
Deloitte respondents report less visibility into per-unit inference costs.
3,500 leaders in 25 countries; May–June 2026 survey. Separate questions, self-reported practices; not a causal ROI result.
AMD reported 97% growth in data-center revenue, while Arm’s license and other revenue grew 50% and royalties 8%, showing distinct routes from AI demand to recognized sales.
AMD data-center revenue grew 97%, from $4.341B to $8.546B. Arm license/other revenue grew 50%, from $515M to $773M; royalties grew 8%, from $620M to $670M. Rates compare each business with its own year-earlier quarter. AMD includes CPUs and accelerators; Arm royalties use estimates of customer shipments. Not a GPU market-share chart.
AMD hardware revenue nearly doubled; Arm licensing grew faster than royalties.
Mixed hardware: CPUs, GPUs and other products
IP and other contractual revenue
Linked to customer product shipments; estimated
Within-company year-on-year growth. AMD quarter ended September 26; Arm September 30. Different revenue models, not market shares.
Cohere announced an initial Series E close targeting up to C$7 billion, with Canada committing up to C$1.6 billion in staged equity; neither figure establishes cash already received.
Cohere’s Series E targets up to C$7B. Canada announced up to C$1.6B of staged equity within that financing. Both bars use the same C$0–7B scale; the total target is hollow and government commitment hatched. Neither mark represents disclosed cash already received; initial close amount is unknown. C$35.7B valuation is not plotted as funding.
Cohere targets up to C$7B; Canada plans up to C$1.6B in stages within the round.
Target for the round, not disclosed cash received
Staged investment within the round
Initial close amount is undisclosed. These are a financing target and a maximum staged commitment; C$35.7B valuation is a separate quantity.
Shield announced an Echodyne acquisition worth up to $5.3 billion and $2.5 billion of supporting financing, with closing expected in early 2027 and part of the purchase price contingent.
Acquisition up to $5.3B comprises $4.8B at closing and up to $0.5B contingent. Announced $2.5B financing comprises $1B Series H equity, $1B structured equity, $0.5B debt. Each strip is normalized to its own total, not a common dollar length. Early 2027 closing is conditional; the separate $28.8B company valuation is not part of either total.
A conditional acquisition price and a mixed financing package describe different totals.
Whole: 5.3 US$ billion
Whole: 2.5 US$ billion
Each strip shows its own total. Announced agreement; early 2027 closing remains conditional. Valuation is not included in either strip.
Ballotpedia’s post-election tally finds the restrictive side won 40 of 41 November 3 local data-center contests, including advisory votes and rejected project approvals rather than forty outright bans.
NVIDIA, Rebellions and SK Telecom plan a mixed-processor inference system using GPUs for prompt processing and Korean NPUs for token generation, with R200 systems expected to be deployable in 2028.
At the edge of what we know
Anthropic researchers report a 96-subunit vault-like shell assembled from a phage protein in bacteria, an experimentally observed structure whose natural function and therapeutic utility remain unknown.
Rosalind, Antigravity and Hydrogen-1.0 offer approved research access, chemistry-tool orchestration and downloadable specialist weights respectively, with different operational requirements and no shared proof of scientific reliability.
A new AutoDNA study reports an instrument-connected workflow that executed a DNA-storage experiment, while a separate laboratory-agent preprint illustrates the simulation and human-approval work still needed around physical automation.
Four recent benchmark releases or updates report substantial failures on regulatory, antibody-development, rare-disease and surveillance tasks, but their different task sets cannot be read as one clinical performance ranking.
Four separate strips: LifeSci Diamond, 16 of 55 tasks; TxBench, 39.4% of attempts on 60 tasks; RareDiseaseBench, 43.8% of attempts on 35 cases; BioSecBench, 25.5% of attempts on 110 tasks. Each Latch study uses three attempts per task; these are not pass-at-least-once rates.
task or attempt pass rate
Best reported configuration within each study
LifeSci counts tasks; Latch counts attempts, three per task. Not the share of tasks solved within three tries or a clinical ranking.
Chai Discovery announced a collaboration applying its molecular design platform across Roche discovery programs, without disclosing financial terms, targets or a clinical result.
A November 5 US policy sets phased screening requirements for nucleic-acid procurement by recipients and subrecipients of federal funding, with NIH confirming that existing requirements remain in force during the transition.
Four milestones share a zero-to-340-day scale: NIST standards within 90 days, initial supplier list within 160, nucleic-acid procurement within 250, and benchtop procurement within 340. Covered recipients’ relevant procurement includes orders paid from other funds.
days after memorandum
Intervals from the November 5 memorandum
All relevant procurement by federal-funding recipients/subrecipients is covered, including purchases using other funds. Milestones are prospective; existing requirements remain in force.
Take the reporting, visual explanations and original sources into a conversation. Copy the brief, paste it into your preferred assistant, then start its voice mode.
Paste the brief into your voice app to discuss this reporting.