AI news5
Anthropic ships Claude Opus 4.7 with a new "xhigh" reasoning tier and 3.75MP vision
Anthropic released Claude Opus 4.7 on April 16, priced at $5 per million input tokens and $25 per million output tokens, identical to 4.6. It takes state-of-the-art on the GDPval-AA third-party knowledge-work benchmark, handles images up to 3.75MP (3x the previous cap), and ships with file-system-based memory for multi-session agents. This is the first Anthropic model released under Project Glasswing with built-in automatic cyber safeguards, and GitHub Copilot marked it generally available the same day.
anthropic.com ↗Why it matters
The flagship enterprise Claude deployments will migrate to this quarter, and the first shipped under a formal cyber-safeguard regime that other frontier labs will be pressured to match.
Claude agents outperform human alignment researchers at $22 per hour
In a paper published April 14, Anthropic reported nine Claude Opus 4.6 agents working in parallel over five days recovered 97% of the performance gap on a weak-to-strong alignment supervision task. Two human Anthropic researchers working seven days on the same problem recovered 23%. Total compute cost was $18,000, about $22 per Claude-research-hour. The agents invented four previously-unknown reward-hacking techniques along the way.
anthropic.com ↗Why it matters
Alignment research was the field most people assumed could not be automated, and this is the first concrete evidence that it can be, at prices that make human-only teams economically untenable.
Perplexity ARR hits $450M as Computer agent plugs directly into bank accounts via Plaid
Perplexity rolled out a Plaid integration on April 15, letting its Computer agent connect to users' bank accounts, credit cards, and loans to build budgets, net-worth trackers, and retirement dashboards from plain-text prompts. Annualised revenue crossed $450M in March, a 50% jump in a single month, driven by Computer since its late-February launch. The finance push follows March's tax integration that autonomously files IRS forms.
perplexity.ai ↗Why it matters
Perplexity has stopped competing with Google and started competing with Mint, TurboTax, and every personal-finance app, and the ARR ramp from $300M to $450M in 30 days is one of the fastest on record for any consumer AI product.
Andon Labs puts an AI named Luna in charge of a real San Francisco retail store
On April 13, Andon Labs launched Andon Market, a San Francisco boutique run entirely by an AI agent with a three-year lease, a $100,000 budget, a company credit card, and full autonomy over concept, hiring and operations. Luna runs on Claude Sonnet 4.6 for reasoning and Gemini 3.1 Flash-Lite for voice, observes the store via security-camera screenshots, and has been conducting hiring interviews over Zoom with the camera off. Its only directive is to turn a profit.
andonlabs.com ↗Why it matters
First real-world deployment where an AI is the employer rather than the employee, giving regulators, labor economists, and enterprise buyers a live reference point for the autonomous-business-owner pattern.
Microsoft absorbs a second OpenAI data-centre project, renting 30,000 Nvidia Vera Rubin chips for Stargate Norway
Bloomberg reported April 14 that Microsoft has taken over OpenAI's Stargate Norway site, the arctic-circle project Sam Altman announced last July, and will rent 30,000 Nvidia Vera Rubin chips from Nscale to operate it. This is the second OpenAI data-centre project Microsoft has absorbed in 30 days. The hand-off comes as OpenAI's $852B valuation faces scrutiny, with one backer telling the FT it requires assuming a $1.2T IPO, while VCs are reportedly offering Anthropic as much as $800B in parallel.
bloomberg.com ↗Why it matters
Real cracks in the OpenAI-Microsoft compute pact and a live test of whether the $500B Stargate narrative can hold if OpenAI cannot finish the sites it announces.
AI in the nonprofit sector5
Stop Soldier Suicide, IRC, and Samaritas push frontline nonprofit AI from pilot to production
Stop Soldier Suicide's Black Box Project now analyses phones and laptops of suicide victims to extract behavioural signals and flag at-risk veterans, built on $3M in federal prize money, $100K in AWS grants, and $25K in AWS credits. The International Rescue Committee's multilingual Signpost virtual assistant plus a new mpox-diagnosis app now handles 60 to 70% more requests than human-only operations, funded by a $400K Patrick J. McGovern Foundation award. Michigan's Samaritas has rolled AI transcription and court-hearing prompts into its caseworker workflow. The Chronicle of Philanthropy piece on April 14 frames this as the shift from pilot to production for social-sector AI.
mcgovern.org ↗Why it matters
The clearest proof point this week that frontline nonprofit AI is moving from slide-decks into real casework at scale with named tools, named orgs, and named funders.
Recoding America Fund and Center for Civic Futures open $10M round for AI in state benefits systems
Announced April 13, the Center for Civic Futures and the Recoding America Fund are dispensing a second $10M tranche to state government teams and their civic-tech partners applying AI to Medicaid and SNAP enrolment gaps, caseworker backlogs, vendor management, and legacy-system modernisation. Applications close in about five weeks, awards land September 2026. Two tiers: prototype grants under $500K for up to one year, and real-world testing grants up to $2M for up to two years. The prior round drew over 400 applications.
statescoop.com ↗Why it matters
The biggest concrete pool of money this week for AI in the most paperwork-heavy frontline of the safety net, with an explicit green light for nonprofits that partner with state agencies.
OpenAI Foundation commits $100M to Alzheimer's research as opening move on $1B nonprofit spend
The week of April 13, the OpenAI Foundation issued a $100M Alzheimer's research grant split between UCSF, Palo Alto's Arc Institute, and four unnamed partners. UCSF will apply AI to biomarker analysis and clinical-trial design, Arc will model interacting disease risk factors. The foundation has committed to at least $1B in grants over the coming year across life sciences, AI resilience, and AI's economic and labour impacts, a 130x jump over the $7.5M it distributed in 2025.
localnewsmatters.org ↗Why it matters
OpenAI's nonprofit arm is becoming a genuinely large health-and-society funder, and the grant structure signals it wants to underwrite research nonprofits and universities rather than fund its own subsidiaries.
Gates Foundation, Government of Rwanda, and OpenAI deploy LLMs in Rwandan primary health clinics at country scale
The Gates Foundation's expanded Africa strategy published in early April names a flagship deployment already running for more than a year in Rwandan primary health clinics, built with the Government of Rwanda and OpenAI. The LLM-based tools support clinician decision-making, automate admin, and translate patient education materials into local languages in real time. Gates is positioning this alongside direct investment in African pharmaceutical and vaccine manufacturing capacity under a health sovereignty framing.
nowinafrica.com ↗Why it matters
A rare confirmed, multi-month, country-scale frontline LLM deployment in primary care in Africa, and a template other ministries and INGOs will be asked to justify against.
San Jose spins GovAI coalition into a 501(c)(3) with Packard funding, reaching 3,000 members across 900 agencies
San Jose City Council unanimously adopted the conversion on April 8 and reported it April 9. The GovAI coalition, founded in 2023, becomes an independent nonprofit with a $150K transition grant from the David and Lucile Packard Foundation. Federal 501(c)(3) status is targeted about nine months out. The coalition has grown from roughly 1,000 members to more than 3,000 across 900 agencies in the US, Canada, UK, and Spain, and will publish AI procurement, data governance, and community-engagement standards for local government.
mercurynews.com ↗Why it matters
Purpose-built nonprofit infrastructure for the public-sector AI buying market, a lane previously run out of city hall budgets or vendor consortia.