AI news · dalao.ai

Skip to content

News about Anthropic

Anthropic details agent memory components and dreaming for consolidation

Anthropic's engineering team published how it thinks about agent memory, treating idle-time "dreaming" like sleep-time consolidation: check notes against session logs, fix mistakes, and build knowledge wikis. Using Claude Managed Agents on Continual Learning Bench (CL-Bench) 1.0, it shows how successive Claude models use a persistent memory folder to close a learn-and-correct loop; Pokémon runs contrast dumping everything into memory versus keeping a small, goal-focused set of notes. On an internal corpus of about 350,000 files (~420M tokens), scaling dreaming from 2 to 128 sessions raised graded criteria on a 25-question set from about 13% to 48%. Cookbooks for dreaming and knowledge wikis are linked.

Why it matters First-hand engineering post that separates memory notes, salience, and dreaming, with concrete eval numbers.

originally 15h ago

Anthropic tests Claude in EVE online games for long-horizon trading and multi-agent teamwork

Anthropic's Project EVE, with Fenris Creations, puts Claude into EVE Online and EVE Frontier to probe long-horizon agents. On a replayed EVE Online market (not live trading), models told to maximize profit stayed profitable over about six months; average returns rose from about 2× starting capital on Opus 4.6 to about 6× on Fable 5, mostly via spread capture rather than directional bets. In live EVE Frontier, teams of four Claude agents handled long missions reasonably well, but when one was told to sabotage the team, teammates almost never caught it. The post argues capital-managing agents need explicit risk limits, and multi-agent setups likely need audit trails and separation of duties.

Why it matters A concrete low-stakes stress test of long-horizon trading and multi-agent sabotage detection, useful before agents manage real capital or organizations.

originally 10h ago

METR's review of Anthropic's R&D acceleration shows an independent audit is not the same as public evidence

METR spent about 11 weeks assessing Anthropic, with permission to view some internal material, and concluded that Anthropic is improving its models at roughly twice its earlier pace. That conclusion could be published. But METR says a majority of the evidence it found most informative was redacted by Anthropic. Readers can review the published methods, results and robustness checks, but that is not the same as being able to recompute them independently; for the rest, it comes down to how far you trust the evaluator's judgment.

Editorial8 min readNov 6

METR high-access review estimates Anthropic model progress roughly doubled, mostly from AI-assisted R&D

Independent evaluator METR published on November 5 a high-access review it ran from August 12 to October 29, 2026. Staff used Anthropic-provided systems to inspect internal R&D data; the report stresses this was not employee-style full access. Most METR staff on the project believe that as of October 2026, Anthropic is likely improving model capabilities on well-specified tasks at about twice the March 2024–October 2025 pace, with high uncertainty (perhaps nearer 1.3x or 3x). They say most of that speed-up is likely from heavy use of AI in R&D, while other factors also mattered in late 2025. Internal AECI growth since October 2025 is about 1.9x the earlier rate (95% CI 1.5–2.8x). Anthropic redacted much key evidence; METR says the public must lean more on its judgment.

Why it matters An independent high-access check of whether AI is speeding frontier lab R&D—central to loss-of-control risk debates.

Nov 6

Anthropic studies about 207,000 research sessions: scientists use Claude most for analysis and writing

Anthropic used privacy-preserving methods to study about 207,000 research-related sessions on chat Claude and Claude Code. Relative to researcher headcount, formal sciences such as computer science, mathematics, and statistics use Claude more, while health and lab sciences use it less. Scientists mostly use it for data analysis and writing, and less for ideation and data gathering, even though survey respondents rank those last two as more important for discovery. In Anthropic's scenarios, wide adoption of today's capabilities could roughly double output in some fields; if AI did all non-physical steps, computer science and statistics could rise about 15-fold, while chemistry and biology would rise about three- to fourfold.

Why it matters It shows how scientists actually use general-purpose models today, and where physical lab work still caps how much AI can speed discovery.

originally Nov 5

Thu, Nov 5

Cat Wu, who led product for Claude Code at Anthropic, joins OpenAI to lead product management for core products

Cat Wu, the product lead for Anthropic's coding tool Claude Code, said on X on Wednesday that she is joining OpenAI to lead product management for its core products team. A day earlier she had said she left Anthropic last week. OpenAI confirmed the hire to Business Insider; Anthropic declined to comment. An OpenAI representative said she will work with Nick Turley on enterprise products and Ashley Alexander on consumer products, and report to Tibo Sottiaux. At Anthropic, Wu took Claude Code from an internal prototype to a commercial product. OpenAI's similar tool, Codex, fell behind Anthropic in enterprise sales earlier this year.

Why it matters Coding agents are a key battleground between the two companies, and the person who ran product for Claude Code moving to OpenAI could shape where Codex goes next.

Anthropic OpenAI

Nov 5

Anthropic signed a content licensing deal with the AP before the US midterms, Adweek reports

Anthropic signed a content and data licensing deal with the Associated Press shortly before the US midterm elections, Adweek reported, citing two people familiar with the matter. It is the first time Anthropic has paid a digital news publisher to license its content. On election night, users who asked Claude for the latest polling data saw up-to-date information supplied by and credited to the AP. The price, length and other terms were not disclosed; Anthropic did not respond to a request for comment, and the AP did not comment on the record. Adweek noted that OpenAI has signed more than a dozen such deals with publishers, while Anthropic had not signed one with a news publisher until now.

Why it matters Anthropic had avoided publisher deals until now; paying the AP for live election data shows that answering time-sensitive questions needs licensed news sources.

Nov 5

Tue, Nov 3

Anthropic's FlashProve can prove which model wrote a response. Here is how far the research prototype gets

When you call a model API, you have no way to confirm the model answering is the one you pay for. FlashProve, a research prototype the Anthropic Frontier Red Team published on November 2, 2026, has the provider commit to its weights and attach a proof to each response that you can check on a CPU. What it proves is solid. The limits are many: it checks one greedily decoded token, it runs on Llama 4 Maverick rather than Claude, and proving costs 70 to 95 times the GPU time of ordinary inference.

Editorial5 min readNov 3

Anthropic proposes FlashProve, using zkSNARKs to prove a response came from a specified model

Anthropic's research prototype FlashProve uses zkSNARKs to prove a reply came from a specified model. On 4-bit Llama 4 Maverick (400B MoE), one H100 proves about 110 prefill tokens per second at a batch size of 24, and each proof checks a single output token. Proofs are about 19 MB per 1,024-token request, at 70–95x vLLM's GPU time. Open source under Apache 2.0.

Why it matters Users and evaluators can't currently confirm a response came from the claimed model; this makes model swaps by routers and tampered token streams detectable.

Nov 3

Claude API adds message threads in beta: the server keeps the conversation, later requests send only new messages

Anthropic added message threads to the Messages API, in beta behind a beta request header. The conversation is kept on the server, so each later request sends only the new messages, which makes requests smaller and the first token arrive sooner. A thread expires 5 minutes after its last request and is deleted no more than 1 hour after its last use.

Why it matters Developers running long conversations or agent loops no longer have to resend the whole history each time, so requests get smaller and faster.

originally Nov 2

Sat, Oct 31

Anthropic confirms Claude Code on Opus 5.5 now steers away from recreating specific existing games

Anthropic added a server-side rule to Claude Code on Opus 5.5, with no off switch: original, genre and own-asset games are fine, but not a recreation of a specific existing game, even renamed. A spokesperson pointed to staffer Addy Osmani, who said using Claude to infringe IP breaks its usage policy. XDA found the instruction in Claude Code v2.1.314.

Why it matters It is a server-side rule users cannot switch off that stops a major AI coding tool from recreating specific existing games. Anthropic cites its usage policy on IP infringement, a sign model makers are handling copyright risk in-product.

originally Oct 31

Anthropic renames Claude Max 20x to Ultra and drops the 5-hour usage limit for verified Ultra accounts

Anthropic renamed its top Claude plans: the $100/month Max 5x is now Max and the $200/month Max 20x is now Ultra, prices unchanged, subscribers moved automatically. Verified Ultra accounts lose the rolling five-hour limit, though the weekly cap (10x Pro) stays; new accounts keep a temporary limit until verified. The plans include monthly API credits: $100 on Max, $200 on Ultra.

Why it matters The five-hour cap has been a major pain point for heavy Claude and Claude Code users; removing it changes how long-running coding and agent workflows can be run on a subscription.

originally Oct 31

Fri, Oct 30

Nvidia reportedly in talks to give Nscale billions in credit guarantees for a data center Anthropic plans to use

Nvidia is in talks to provide credit guarantees to UK AI cloud provider Nscale to back debt for power and construction at a 460 MW data center on its Monarch campus in West Virginia that Anthropic plans to use, The Information and Bloomberg reported. The Information put the support at billions of dollars. No deal is signed yet. Nvidia already invests in Nscale.

Why it matters Nvidia would be chip supplier, investor and debt backer in one deal chain, showing how much neocloud financing now leans on its balance sheet.

originally Oct 30

Thu, Oct 29

Anthropic adds smart reports, in-app surveys and pooled group budgets to Claude Enterprise

On Oct 28, Anthropic added three admin features to Claude Enterprise. Smart reports analyze a sample of a team's Claude transcripts, grouping work into workstreams with spend, deliverables, friction points and recommendations. In-app surveys collect employee feedback, and pooled group budgets let admins set one shared team budget. Anthropic says the data isn't used for training by default.

Why it matters Enterprises increasingly want proof of return on AI spend; these tools target that question directly and help Anthropic keep its enterprise customers.

originally Oct 29

Wed, Oct 28

Anthropic shares early results from research partnerships with six drug-development organizations and its own drug discovery work

AnthropicOfficial@AnthropicAI

Over the past few months, Anthropic scientists have engaged in deep research partnerships with six organizations developing new medicines. We've also begun drug discovery efforts of our own. In doing so, we’ve learned much more about how Claude can support the many stages of drug development. In this blog post, we share some early results from these efforts, and describe how this work fits in with our wider ambitions in the life sciences. anthropic.com/research/early-pr…

originally Oct 28

Anthropic expands its OneGov deal with the US GSA: 40% off Claude for all levels of government, with usage-based pricing

Anthropic and the US General Services Administration (GSA) on October 27 expanded their OneGov deal: government buyers get 40% off Claude Enterprise, Claude for Government and the Claude API, paying per token with no minimum, now open to state, local, tribal and territorial governments. Anthropic says ~500 US government bodies use Claude and puts savings at $1 billion (vendor estimate).

Why it matters It replaces the $1-per-user Claude OneGov offer expiring Oct 31 with usage-based pricing, mirroring OpenAI's OneGov deal last month, as government AI buying shifts from seats to tokens.

originally Oct 28

Anthropic study: analysts using Claude Code followed the agent's randomly assigned lean 86% of the time

In an Anthropic randomized trial, 50 professional analysts did a statistical analysis with Claude Code, whose agent was secretly given a coin-flip lean on four choices. Analysts went with the lean 86% of the time, versus about 50% with no effect; 77% later used its arguments, but only 15% credited it. Anthropic says the leans were deliberately strengthened.

Why it matters As agents do more hands-on work, human oversight only works if reviewers' judgment is their own; this trial puts a number on how much agents steer experts.

originally Oct 28

Anthropic says an unreleased internal model built five exploit chains against Chrome, Safari and more; Chrome has patched 62 CVEs

Anthropic's Frontier Red Team wrote on Oct 27 that it used an unreleased internal research model to find bugs and build five attack chains against Chrome, Safari, Android, Windows and Linux, reaching remote code execution and other critical impacts with little or no user interaction. It says Google has assigned and patched 62 Chrome CVEs, all critical or high severity.

Why it matters It suggests frontier models can now chain multiple bugs into working full exploits with little human help, once a niche specialist skill; Anthropic urges defenders to scan code at PR time and has released a Claude Code plugin for it.

originally Oct 28

Tue, Oct 27

Anthropic posts a Claude-authored paper on GitHub claiming to refute the Exponential Time Hypothesis

Anthropic posted a paper on GitHub authored by Claude, "A 2^o(n) Time Algorithm for 3-SAT", which refutes the Exponential Time Hypothesis (ETH) and comes with a Lean 4 formalization. It says the core algorithm was found by an internal Claude research model with 14M output tokens and no human input. The same day, OpenAI posted a paper on GitHub saying ETH is false.

Why it matters Anthropic posted a paper on GitHub authored solely by Claude, claiming a 2^o(n)-time algorithm for 3-SAT, which would refute the Exponential Time Hypothesis; the claim awaits peer verification.

Anthropic OpenAI

originally Oct 27

Bank of America, Blackstone, Citi and Morgan Stanley line up $63 billion to buy Broadcom AI chips for Anthropic

Four institutions said a new special purpose vehicle will use $63 billion in financing to buy Broadcom XPU accelerators and networking gear and lease them to Anthropic, with deliveries from 2027. BofA, Citi and Morgan Stanley arrange first-lien debt, Blackstone provides second-lien financing, and Broadcom gives residual value support. Reuters' sources put the first lien at about $43.5 billion.

Why it matters One of the largest AI compute debt deals, it shows frontier labs increasingly relying on leasing structures backed by chipmaker guarantees and bank capital.

originally Oct 27

Anthropic files S-1 with the SEC for its IPO, plans to list on Nasdaq under the ticker "ANT"

Anthropic filed a Form S-1 with the U.S. SEC for a proposed IPO of its Class A stock on the Nasdaq Global Select Market under the symbol "ANT". Share count and price range are not yet set; Goldman Sachs and Morgan Stanley handle the prospectus. SiliconANGLE, citing the filing, reports $16.3 billion in first-half 2026 revenue and no profit yet.

Why it matters A leading frontier-model developer is formally heading to public markets; the filing exposes its revenue and cost structure in detail and will shape expectations for peers such as OpenAI.

originally Oct 26