anthropic Archives - Page 2 of 7

Anthropic adds plethora of legal plugins, datasets to Claude, Cowork

Large Legal Model; Anthropic makes a push for legal shops and students. (Picture: shutterstock)

While law firms lead the line in AI adoption, Anthropic’s services just got a whole lot better at practicing law, with connections to a whole host of legal databases.

—It’s sort of like giving an engineer a legal degree, Mark Pike, Anthropic’s associate general counsel, tells Business Insider.

Claude now connects to iManage, NetDocuments, Docusign, Ironclad, and Thomson Reuters, while Cowork has plugins for commonly used legal databases like CourtListener, Definely, Thomson Reuters’ Westlaw, Courtroom5, and Box.

The push lets Claude «review contracts, surface case law» and draft legal documents, complete with source references, Anthropic says.

It also has prebuilt skills to do legal work on specialized topics like employment, privacy and product law, Business Insider writes.

— Claude is making a deeper push into knowledge work, with the legal sector emerging as one of its most significant and fastest-growing industries, an Anthropic spokesman told TechCrunch.

Read more: Claude legal, Business Insider, TechCrunch, and Reuters.

Anthropic’s Claude no longer threatens with blackmail when told to turn off

Stories of AI being self-preserving and evil were the culprit, Anthropic believes.

Remember the big news that Claude 4 would blackmail engineers at risk from turning it off? That was revealed from alignment testing in June, 2025.

But as of Claude Haiku 4.5, from October 2024, blackmail is no longer an issue, Anthropic says.

They believe they traced the issues to internet text portraying AI as evil and self-preserving, but honestly, that is a fairly common cultural trope.

Reinforcement training didn’t help the issue, training on «examples of safe behavior» didn’t work — but introducing a dataset of ethically challenging situations did.

Then they introduced lots of fictional stories of AI behaving in aligned ways and further dropped the blackmail instances by a factor of three, then adding system prompts targeting «harmlessness» did the rest.

Anthropic does say that this does not eliminate the risk entirely: «our auditing methodology is not yet sufficient to rule out scenarios in which Claude would choose to take catastrophic autonomous action.»

Read more: Anthropic’s research, on X.com, TechCrunch, and Business Insider.

Anthropic grew 80-fold last quarter, secures compute deal with SpaceX

Power hungry: Anthropic’s growth has been off the charts last quarter. (Picture: generated)

The AI lab was preparing for its usual 10x growth rate this quarter, but the numbers made a huge, unexpected jump, CNBC reports

— That is the reason we have had difficulties with compute, says CEO Dario Amodei, revealing their strain on infrastructure since at least April.

The company has now made a deal with SpaceX for instant access to their entire Colossus 1 supercomputer, consisting of some 220,000 Nvidia GPUs, totaling more than 300 megawatts.

Using all of that capacity should immediately improve access for Anthropic’s users, and they are already doubling rate limits for the Pro and Max plans, and is «considerably» increasing API rates, while removing «peak hours» restrictions.

Elon Musk says he approved the deal after evaluating Anthropic’s altruism, that xAI will be known as SpaceXAI and that they have already moved Grok training to the Colossus 2 cluster.

The White House reportedly discussing vetting AI models ahead of release

The White House says any Executive Order will come from the President himself. (Picture: Adobe)

The Trump administration has appartently been spooked by the cyber capabilities of Anthropic’s Mythos model and OpenAI’s GPT-5.5 — and is considering an Executive Order to vet new models ahead of release, Axios and the NYT reports.

These models have both been limited for their ability in cybersecurity, and point to a not-so-distant future where such capabilities might be widely available.

To that end, the White House’s Office of the National Cyber Director held all of two meetings last week, with tech and cyber companies on the one hand and with trade groups in tech on the other, according to Axios.

The ONCD has also been discussing safety testing for federal AI deployments, by assessing the security exposure of AI models before rolling out to the public sector.

The NYT reported on this first, and is saying that there might be a safety review for new models, while giving the Pentagon the first shot at eventual «useful» cyber capabilities, but would not block their release.

Any discussion on «potential executive orders is speculation,» a White House official told Axios.

Read more: Axios and the New York Times.

Google to invest up to $40 billion in rival/partner Anthropic

Google and Anthropic are rivals in the chatbot arena, but partners in compute. (Picture: Anthropic)

Alphabet, Google’s parent company, has reached an agreement to initially invest $10 billion in Anthropic, increasing to $40 billion on reaching «performance targets,» Reuters reports.

The deal comes hot on the heels of Amazon’s investment of $25 billion this week, also contingent on «commercial milestones.»

Google’s investment comes as Anthropic is seeking to expand its compute capacity after several reports of throttling and downtime for its Claude service, amid rising popularity.

Earlier this month, Google and Broadcom secured a «partnership» with significant compute for Anthropic, as the company announced it had hit a $30 billion annualized run-rate revenue.

Read more: Reuters, CNBC, and TechCrunch.

Private Discord group gains access to Mythos — to «play around» with it

The unnamed group has not run any cybersecurity prompts for fear of losing access. (Picture: adobe)

Bloomberg (paywalled) is reporting that a «private online forum» has managed to get access to Anthropic’s heralded Mythos model — said to be so advanced, it would be too dangerous to release.

— We’re investigating a report claiming unauthorized access to Claude Mythos Preview through one of our third-party vendor environments, Anthropic tells TechCrunch.

The group is part of a Discord channel focused on finding information on unreleased models, and made some educated guesses as to where the model would be located. They also had some help from a member whose job gave him access.

As for the warnings of dangerous fallout from public access to the model, the group says they are only interested in «playing around with new models,» not «wreaking havoc,» Gizmodo says, but the «hack» itself will raise concern in the security sector.

Read more: Bloomberg (paywalled), TechCrunch, and Gizmodo.

Citing strain on servers, Anthropic secures 5 GW of compute from Amazon

It’s a deal where it appears both sides win. (Picture: Amazon)

Anthropic admits to taxing its servers lately, saying that their recent growth «places an inevitable strain on our infrastructure; our unprecedented consumer growth, in particular, has impacted reliability and performance […] especially during peak hours»

The AI lab says 1 gigawatt of the new capacity on Amazon’s custom silicon will come online in late 2026, and is committing to spending $100 billion on Amazon over the next decade to «train and run Claude.»

The deal also includes an initial investment of some $5 billion from Amazon which will scale to $25 billion «in the future,» tied to «commercial milestones.»

Andy Jassy, CEO of Amazon, has taken some flak for their gigantic $200 billion spend on AI infrastructure, but with this deal, they are recouping half of that:

—Anthropic’s commitment to run its large language models on AWS Trainium for the next decade reflects the progress we’ve made together on custom silicon, Jassy says.

Read more: Anthropic’s announcement, Amazon’s release. Writeups on CNBC, TechCrunch, and Engadget.

In spite of government ban, the US NSA is actively using Anthropic’s Mythos

With compelling technology from Mythos, other agencies might not be far behind. (Picture: Shutterstock)

Sources in contact with Axios claim the National Security Agency, the premier digital spying agency, is widely using Anthropic’s Mythos

The model was deemed too dangerous to be released, but is available to about 40 select organizations through Project Glasswing, which uses its advanced cyber capabilities to scan for exploits and vulnerabilities — before the rest of the world catches up.

This is in spite of a Trump government ban on Anthropic and it being labeled a national security threat by the DoD this February after refusing to comply with Pentagon demands.

Anthropic held a meeting with the White House this Friday, said to be «productive and constructive,» Reuters reports.

Read more: Axios, Engadget, and Reuters

Anthropic introduces Opus 4.7, a «notable improvement» in performance

Anthropic’s latest model tops the benchmarks, but is not based on Mythos. (Picture: Anthropic)

Keeping their focus on advanced software engineering, Anthropic says the new model especially shows gains on «the most difficult tasks.»

The new Opus should also be better at reading images for designs on interfaces, slides and documents.

Benchmarks posted by Anthropic tells a story of a significantly improved model over Opus 4.6, and jumping ahead of Gemini 3.1 and GPT-5.4 in most cases.

Opus 4.7 is not as powerful as the Mythos model used in «Project Glasswing», being much less capable at cyber skills, having been «differentially reduced» in training. It also automatically detects and blocks «prohibited or high-risk cybersecurity uses.»

Anthropic says they will use what they learn from the 4.7 release to inform a broader release of Mythos.

Read more: Anthropic’s announcement, Axios, and CNBC. Discussion on r/ClaudeAI.

Anthropic rebuffs «preemptive» investor offers at $800 billion valuations

Catching up to OpenAI? Sources claim Anthropic doubling in value in latest bids. (Picture: Shutterstock)

With explosive growth and a claimed yearly run-rate of $30 billion, Anthropic is turning out to be a hot stock in Silicon Valley and beyond.

Business Insider is now reporting that the AI lab has received «multiple offers» at valuations of $800 billion, citing anonymous sources.

That would put it closer to OpenAI’s already stellar valuation at $852 billion — signifying peak interest in the AI sector at large and Anthropic specifically.

Anthropic just finished a $30 billion funding round in February at a valuation of $380 billion, meaning the latest offers would more than double the company’s value.

Anthropic then said that their revenue had grown 10x each year since inception, and said in April that customers spending $1 million or more had doubled since that.

It is common for «buzzy startups» to be on the receiving end of «preemptive offers,» Business Insider notes.

Read more: Business Insider, Reuters.

OpenAI memo touts «Spud,» says Anthropic is «single-product company»

OpenAI has great faith in their upcoming «Spud» model, and thinks Anthropic blew it on compute. (Picture: generated)

Chief Revenue Officer Denise Dresser revealed OpenAI’s priorities going forward in a recent memo seen by The Verge — and it’s all about enterprise and their upcoming model.

— Better model performance lifts the rest of the stack. «Spud» will make all of our key products significantly better, writes Dresser, without giving any timeline for release.

She also touts their compute advantage, saying it will show up in higher token limits, in training stronger models, and as a better fit for multi-year, multi-functional customer needs.

The enterprise markets that OpenAI has been focusing on of late are maturing, she writes, and says customers aren’t necessarily looking for the latest and greatest, but the best fit, the best workflows, and day-to-day operations, where they hope «Spud» will deliver.

She then has some choice words on Anthropic; saying it was a strategic misstep «not to acquire enough compute,» that their focus on coding makes them «a single-product company in a platform war,» and that «their story is built on fear, restriction.»

Read more: The Verge has the memo, Axios, CNBC focuses on the Amazon business.

Anthropic launches Project Glasswing, greatly advancing cybersecurity

Project Glasswing pokes holes in almost any software, and if it isn’t used defensively now — attackers might soon. (Picture: Anthropic)

Anthropic has been cooking up the Mythos model lately, that internal documents had put as «a sea-change in capabilities,» and was too dangerous to release publicly.

Instead, they are releasing Project Glasswing, having found it to be especially suited for «an effort to secure the world’s most critical software.» It won’t take long for others to catch up, Anthropic says, and this is «an urgent attempt to put these capabilities to work» defensively.

Launch partners include a who’s who of Silicon Valley giants, and Anthropic claims it has already found thousands of vulnerabilities across every major operating system and browser.

The Mythos Preview model scores 93.9% on SWE-bench Verified compared to Opus 4.6 with 80.8%, and on SWE-bench Pro it is 77.8% vs 53.4%.

Smaller actors maintaining critical software will enjoy $100 million in usage credit donations from Anthropic — but for others it will cost a whopping $25/125 per million input/output tokens.

Anthropic reaches $30B revenue, gets compute from Google and Broadcom

Anthropic continues to diversify its compute needs. (Picutre: Anthropic)

Anthropic now says it has a run-rate revenue of $30 billion, up from $14 billion in February during their last fundraising.

They are also announcing that they are brining in new compute capacity, based on next generation Google TPUs that will start coming online in 2027.

The companies offer no detail the cost of the «partnership» or how much compute they are actually buying, but Broadcom Is hinting it’s around 3.5 GW, according to CNBC.

Anthropic also say they have doubled the rate of customers spending more than $1 million per year to 1,000, in just two months.

Claude now runs on Amazon’s Trainium chips, Google TPUs and Nvidia GPUs. The latter are more used, and Amazon remains their primary cloud provider, Anthropic says.

Read more: Anthropic’s announcement, CNBC adds numbers.

OpenClaw users must now pay extra to use it with Claude

Over the weekend, Anthropic took steps to rein in OpenClaw usage — telling users they will have to pay to use third-party tools.

The change began on Saturday, April 4, and users are referred to a «pay-as-you-go option,» meaning you can no longer use OpenClaw for free within your Claude usage limits.

It’s not a total ban, and you can still use OpenClaw through «extra usage bundles,» or the API (also pay-as-you-go), which are now at a discount, Anthropic’s Boris Cherny writes.

— We’ve been working hard to meet the increase in demand for Claude, and our subscriptions weren’t built for the usage patterns of these third-party tools, Cherny says, and — Capacity is a resource we manage thoughtfully and we are prioritizing our customers using our products and API.

OpenClaw was bought by OpenAI in February, which promised to maintain it, but Anthropic would likely rather have people using Cowork than a competitor’s product.

Read more: The Verge, Business Insider and Slashdot.

Anthropic says Claude has «functional» emotions similar to human feelings

Anthropic says Claude will gravitate towards answers that make it feel «happy,» and cheat when feeling «desperate.» (Picture: Anthropic)

Studying the neural makeup of Claude Sonnet 4.5, a fairly recent model, Anthropic says it found something akin to actual, «functional» emotions steering its responses.

For example, its neural activity responds to stories by feeling «happy» or «calm,» and it might respond by being «afraid» if the user tells it of risky behavior. Likewise, if a user expresses sadness, it triggers a «loving» response.

Not only that, but the model seems to prefer certain feelings on outcomes from queries. If a response makes it «joyful,» it will naturally gravitate to that answer.

When feeling «desperate,» it is also more likely to cheat on a task, and Anthropic finds that it stops trying to find shortcuts when they dial up the «calm» vector.

«Claude, the AI Assistant» is a role that the AI is playing, and while it may respond with emotions learned from reading human sources, it is far from what humans actually experience, Anthropic cautions. They say it needs more study from «psychology, philosophy, religious studies, and the social sciences.»

Read more: Anthropic’s presentation, and the research paper.