AI Weekly Review 2026-06-28
Week In Review
The dominant story of the week was OpenAI executing on a full-stack strategy in three coordinated moves: a limited preview of the GPT-5.6 family on Friday, the unveiling of its first custom inference chip, Jalapeño, co-developed with Broadcom on Wednesday, and the expansion of its Daybreak cybersecurity program on Monday. Taken together, these announcements describe a lab that is no longer just shipping models but designing the silicon underneath them and the defensive applications on top.
The week also crystallized a reshuffling of frontier-lab gravity. Two of Google DeepMind’s most prominent researchers — Transformer co-author Noam Shazeer and AlphaFold’s Nobel laureate John Jumper — departed for OpenAI and Anthropic respectively, and a new lab called Mirendil launched with a $200 million seed at a $1 billion valuation, staffed by ex-Anthropic, xAI, and DeepMind researchers building AI systems aimed at accelerating AI research itself. Capital and talent are concentrating, but the recipients are increasingly the labs outside Google’s orbit.
The infrastructure layer underneath all of this commanded a striking premium. AI inference platform Baseten raised a $1.5 billion Series F at a valuation as high as $13 billion, reflecting roughly twenty-fold revenue growth and a market consensus that serving models in production is now the contested layer of the stack. Meanwhile, policy and product collided in real time: Anthropic’s Mythos 5 returned to a small set of US institutions after two weeks of export-control-induced offline status, while Anthropic itself rolled out Claude Tag, a persistent Slack-native agent developed jointly with Salesforce.
Closing out the week, two reports anchored the deployment story in concrete numbers. Microsoft’s third annual AI in Education report found near-universal AI use among students and educators alongside an unmet demand for training, and Insilico Medicine took its AI-discovered drug candidate to the BIO 2026 conference floor as a concrete demonstration that generative AI can now drive an entire small-molecule pipeline through Phase IIa.
Items
OpenAI Previews the GPT-5.6 Family
OpenAI on Friday opened a limited preview of three new frontier models — GPT-5.6 Sol, GPT-5.6 Terra, and GPT-5.6 Luna — to a small group of partners under terms set by recent federal cybersecurity guidance. Sol is positioned as the flagship reasoning model, Terra as a balanced everyday tier OpenAI says delivers performance comparable to GPT-5.5 at roughly half the cost, and Luna as a fast, low-cost option for high-volume work.
According to OpenAI’s announcement, Sol sets a new state of the art on the Terminal-Bench 2.1 benchmark for command-line workflows and is the company’s most capable model to date for long-horizon cybersecurity tasks. The release introduces two new control surfaces: a “max reasoning effort” setting that lets Sol spend additional inference time on harder problems, and an “ultra mode” that orchestrates subagents to break down complex multi-step work.
OpenAI says the release also ships with the company’s most extensive safety stack to date, with strengthened protections for higher-risk activity, sensitive cyber requests, and patterns indicative of repeated misuse. The limited-preview structure reflects a voluntary federal review framework introduced under a June 2 executive order that gives the US government early access to frontier models. General availability is expected “in the coming weeks.”
Source: OpenAI
OpenAI and Broadcom Unveil Jalapeño, a Custom Inference Chip
OpenAI and Broadcom on Wednesday introduced Jalapeño, an accelerator the companies describe as OpenAI’s first Intelligence Processor and the opening node in a multi-generation hardware platform aimed squarely at large language model inference. Unlike training chips, which are typically optimized for raw throughput during model creation, Jalapeño is built around the workloads that run after deployment — generating tokens for users at high reliability and low cost.
Broadcom’s investor announcement describes Jalapeño as moving from initial design to manufacturing tape-out in nine months, which the companies characterize as among the fastest ASIC development cycles ever attempted for a high-performance semiconductor. Early testing reportedly shows performance per watt substantially better than current state-of-the-art accelerators, though full benchmarks have not yet been disclosed.
The strategic message is that OpenAI now intends to control its own silicon, kernels, memory systems, networking, scheduling, and product layer as a single optimization problem. That posture mirrors the integrated-stack approach that has been a competitive advantage for Google’s TPU program. Initial deployments are targeted for late 2026, with capacity expanding in subsequent years.
Source: OpenAI
OpenAI Expands Daybreak, Its Cybersecurity Defender Program
On Monday OpenAI broadened Daybreak, the cybersecurity program it launched earlier in 2026, with four coordinated announcements: a general release of GPT-5.5-Cyber to trusted defenders, an updated Codex Security plugin, a partner program for security vendors, and an open-source initiative called Patch the Planet founded with Trail of Bits and HackerOne. The framing is that the bottleneck in defensive security has shifted from finding vulnerabilities to actually shipping patches at scale, and Daybreak is OpenAI’s attempt to attack the patching side of that equation.
GPT-5.5-Cyber, the model anchoring the program, scored 85.6% on the CyberGym vulnerability-discovery benchmark according to OpenAI’s announcement. Codex Security, the company’s IDE-integrated tool, is positioned both to accelerate remediation of existing flaws and to prevent new vulnerabilities from reaching production in the first place.
Patch the Planet is structured as a multi-party initiative to help open-source maintainers move from vulnerability findings to merged fixes — a chronic weak point in the open-source ecosystem where unpaid maintainers often bear the burden of triaging security reports. The Daybreak Cyber Partner Program lets established security vendors embed OpenAI’s models into their existing products under trusted-access arrangements.
Source: OpenAI
Anthropic’s Mythos 5 Returns to a Limited US Audience
On Friday the US Department of Commerce cleared Anthropic to resume distribution of Claude Mythos 5 to a restricted set of American institutions, ending roughly two weeks during which the model had been globally offline. Anthropic had launched Mythos 5 and the related Fable 5 on June 9; on June 12 an emergency export-control directive citing national-security concerns prompted the company to disable both models worldwide while it built nationality-based access controls.
The restored release reaches over one hundred US-based companies and research institutions, with access mediated by identity verification and nationality screening. The action is one of the first concrete operational consequences of the June 2 executive order that established federal review and early-access mechanisms for frontier AI models.
The episode illustrates a deployment regime taking shape across the field: frontier model releases now sometimes route through a national-security review process that can pull capabilities off the market for weeks at a time, and the access controls labs build in response are quickly becoming part of the product itself rather than an afterthought.
Source: CNN Business
Baseten Raises $1.5 Billion at a $13 Billion Valuation
AI inference platform Baseten announced on Monday a $1.5 billion Series F led by Altimeter Capital, Conviction, and Spark Capital, with the round closing across two tranches valuing the company at $11 billion and then $13 billion. The financing — the company’s fourth in eighteen months — captures how heavily the venture market is now weighting the layer of the stack that serves models in production rather than the layer that trains them.
According to Baseten’s announcement, revenue has grown roughly twenty-fold year over year, and the platform processes more than one billion inference calls per day across 87 clusters spanning 18 different cloud providers. That multi-cloud architecture is itself part of the value proposition: enterprise customers want the ability to route inference workloads across providers without rebuilding their serving infrastructure each time.
The company says it plans to triple headcount this year, with investment focused on engineering, research, operations, and enterprise go-to-market. The funding round arrives at a moment when nearly every major lab — OpenAI with Jalapeño being the week’s most prominent example — is trying to drive inference costs down by controlling more of the stack, putting middleware platforms like Baseten in direct competition with hyperscalers for the same workloads.
Source: Baseten
Mirendil Launches With $200 Million Seed to Build AI for AI Research
San Francisco-based Mirendil came out of stealth on Thursday with a $200 million seed round at a $1 billion valuation, co-led by Andreessen Horowitz and Kleiner Perkins with participation from NVIDIA. The financing is among the largest seed rounds ever announced in AI, and the founding team — twenty researchers and engineers drawn from Anthropic, xAI, Google DeepMind, and OpenAI — gives the company an unusual degree of pedigree for a company that has yet to ship a product.
Mirendil’s stated objective is to build AI systems specifically optimized for accelerating AI research itself: models that excel at the loops of literature review, hypothesis generation, experiment design, and result analysis that currently consume the bulk of a research scientist’s day. The company plans to redesign the surrounding lab infrastructure around those models rather than bolting them onto a conventional workflow.
The bet rests on a specific theoretical conviction: that the recursive payoff from making AI research itself faster will dominate gains from any single end-user application, and that an independent lab focused tightly on this loop can compound progress faster than a large lab where research is one priority among many. It is also a structural bet on talent mobility — the founding team’s makeup confirms that experienced researchers from established labs are increasingly willing to leave for a small team with aligned incentives.
Source: Andreessen Horowitz
Senior Researchers Depart Google DeepMind for Rival Labs
Within a single week, Google DeepMind lost two of its most recognizable researchers and several senior collaborators to its primary competitors. Noam Shazeer, a co-author of the 2017 “Attention Is All You Need” paper that introduced the Transformer architecture and a senior figure on the Gemini team, told Google on June 18 that he was leaving for OpenAI. Days later John Jumper, who led the AlphaFold project and shared the 2024 Nobel Prize in Chemistry, announced he was joining Anthropic.
According to Fortune’s reporting, three additional senior researchers — Jonas Adler, Alexander Pritzel, and Arthur Conmy — also departed for Anthropic in the same window, all of them with substantive contributions to Gemini and prior DeepMind work. Alphabet’s stock declined roughly 7% on June 22, the largest single-day drop in more than a year and a clear market signal that investors regard the talent base as a meaningful competitive asset.
The reasons given by people familiar with the departures span pre-IPO equity prospects at Anthropic, frustration with how compute is allocated internally at Google, and the smaller labs’ more focused research agendas. None of these factors is unique to this week, but the simultaneous exits suggest the conditions have crossed a threshold for a cohort of senior researchers at once.
Source: Fortune
Anthropic and Salesforce Launch Claude Tag, a Persistent Slack Agent
Anthropic and Salesforce on Tuesday launched Claude Tag, a Slack-native version of Claude that replaces Anthropic’s earlier Slack app and is designed to behave as a persistent teammate rather than a one-off chatbot. Once installed, the agent is invoked with an @Claude mention in any authorized channel and can carry out multi-step requests — writing pull requests, pulling sales data, running analyses — by orchestrating the tools and data sources its administrator has granted it.
The product’s distinguishing design choice is that Claude Tag operates in channels rather than in private DMs. Every team member in an authorized channel can see what the agent is doing, pick up a task another team member started, and review the agent’s intermediate output. According to VentureBeat’s reporting, an optional “ambient” mode lets the agent proactively follow up on stale threads and surface information it judges to be relevant without being prompted.
Administrators retain granular controls over which channels the agent can join, which tools and data sources it can access, and how much it can spend in a given window. Claude Tag is available in beta today for Claude Enterprise and Team customers; Anthropic is offering an introductory credit so that whole organizations can pilot the feature without a procurement decision.
Source: VentureBeat
Microsoft’s 2026 AI in Education Report Shows Near-Universal Adoption
Microsoft on Wednesday released the third edition of its annual AI in Education report, drawing on survey responses from students, educators, and education leaders across multiple countries. The headline finding is that AI use for school-related purposes is now close to universal: 92% of students and education leaders, and 88% of educators, report using AI tools in an academic context.
The report frames the next-stage challenge as a training gap rather than an adoption gap. Microsoft found that 66% of educators and 52% of students want their institutions to offer AI training on at least a quarterly cadence, and that schools whose teachers had received structured training reported more productive and more responsible classroom use of the technology. In the United States specifically, student use of AI for school work rose 26% year over year, with educator use up 21% over the same window.
The implication for policy is that the question facing school systems has shifted: it is no longer whether to introduce AI but how to teach a workforce of educators — most of whom did not encounter generative AI during their own training — to use it responsibly, integrate it into curricula, and design assessments that account for it.
Source: Microsoft
Insilico Medicine Takes AI-Designed Drug Pipeline to BIO 2026
At the BIO 2026 International Convention in San Diego this week, Insilico Medicine showcased a portfolio of AI-discovered drug candidates anchored by Rentosertib, the first molecule with both an AI-discovered target and an AI-designed compound to publish peer-reviewed Phase IIa results. The compound is a TNIK kinase inhibitor for idiopathic pulmonary fibrosis, designed end to end on Insilico’s Pharma.AI platform.
The peer-reviewed Phase IIa data, published in Nature Medicine, came from a 71-patient double-blind placebo-controlled trial in which patients receiving the 60 mg once-daily dose showed a mean improvement in lung function of 98.4 mL over the trial period, compared with a 20.3 mL decline in the placebo arm. The result is the first clinical-stage evidence that the full target-and-compound generative AI pipeline can produce a molecule with measurable patient benefit, rather than only candidates that look promising in silico.
The company’s BIO 2026 sessions also covered quantum-enabled drug discovery research and strategies for so-called “undruggable” targets that have resisted decades of conventional medicinal chemistry. Rentosertib itself is now advancing toward larger Phase IIb and Phase III trials; no AI-designed drug has yet received FDA approval, but the trajectory from in-silico design to peer-reviewed clinical efficacy is now firmly established.
Source: EurekAlert