• AI KATANA
  • Posts
  • Anthropic ships Claude 4.1 upgrade

Anthropic ships Claude 4.1 upgrade

Also: OpenAI to livestream GPT-5 Thursday 10AM PT

Author

AI KATANA
August 06, 2025

In partnership with

An Entire Month of Videos Before Lunch

Tired of the post-every-day grind?

Syllaby.io automates your entire content workflow. All you need is a topic—our AI does the rest.

✅ Get daily viral content ideas
✅ Auto-generate scripts tailored to your niche
✅ Instantly create faceless videos
✅ Bulk schedule across all your platforms

Syllaby is perfect for coaches, creators, and marketers who want to grow without showing their face or spending hours editing.

Hello!

Anthropic’s surprise Claude Opus 4.1 point-release leads a packed day of AI headlines, trimming latency by nearly a third while stretching context to 384 k and debuting an autonomous validator that pre-screens disallowed content—an aggressive tempo aimed squarely at GPT-5’s impending debut, which OpenAI will preview in a Thursday livestream. Google, meanwhile, is pouring $1 billion into a three-year Gemini training program across more than 100 U.S. universities, coupling free Pro-Plan licences with GPU grants to entrench its stack in higher-ed workflows. On the infrastructure front, Broadcom’s 3-nm Jericho 4 router-on-chip promises 16 Tbit/s links that can knit 100 km-spaced GPU clusters into a single high-bandwidth fabric, while DeepMind’s Genie 3 generates fully explorable 720 p worlds from text that persist for minutes, signaling new horizons for embodied AGI research. Policy watchers are rattled by reports that NIST is withholding a 139-page red-team study exposing fresh jailbreak vectors just as Senate hearings on AI procurement loom, and investors are bracing for SoftBank’s earnings call after an $11 billion spree of AI bets pushes leverage toward pre-2019 levels.

Sliced just for you:

  • 🦾 Anthropic ships Claude 4.1 upgrade

  • 🆕 GPT-5 livestream

  • 🎓 Google puts $1B into campus AI

  • 🌐 Broadcom’s Jericho 4 links 60-mile clusters

  • 🕵️‍♂️ NIST red-team report stays hidden

  • 🎮 DeepMind’s Genie 3 spawns living worlds

  • 💹 SoftBank’s earnings hinge on AI spree

Anthropic rolled out Claude Opus 4.1, a point-release that quietly boosts agentic planning, long-form coding and chain-of-thought reasoning while holding price steady. API logs leaked by early adopters show a 28 % latency cut and a jump to 384 k context, enabling multi-file refactors in a single pass. Anthropic says a new “autonomous validator” now rejects 93 % of disallowed content before it ever reaches the user, answering regulators’ “duty of care” concerns without human in-the-loop. The model lands simultaneously on Anthropic’s console, Amazon Bedrock and Google Vertex, signalling deeper multi-cloud ties even as the company courts a mega-round that could value it at $170 billion. Observers note the cadence—two months after Opus 4—shows Anthropic adopting a rapid, Tesla-style refresh cycle to stay top-of-mind before GPT-5 arrives.

Google pledged a three-year, $1 billion programme to train students and faculty at more than 100 U.S. universities on Gemini-powered tools, complete with free Pro-Plan licences and 5 000 GPU-hour research grants. The announcement dovetails with a new “guided learning” mode inside Gemini that breaks homework into step-wise hints, interactive quizzes and image-based explainers rather than handing out answers. Education deans praised the pledge as the biggest AI-literacy push since Google’s 2012 Cloud credits, while critics question whether locking coursework to one vendor risks future fees. Google counters that all materials will be Creative-Commons and promises annual audits to track equity outcomes, positioning the move as both philanthropy and talent pipeline.

Broadcom unveiled Jericho 4, a 3-nm router-on-chip that can sling 16 Tbit/s across campus fabrics and chain GPUs in data-centres 100 km apart without choking synchronous gradient exchange. Each device packs on-chip high-bandwidth memory, lossless Ethernet buffers and line-rate MACsec, allowing a single petascale installation of 4 500 chips to appear as one giant, secure backplane. Cloud architects say the part lets operators stitch together smaller, cheaper sites instead of betting everything on one hyperscale complex, easing land-use and power constraints. Early sampling is under way with “multiple tier-one” providers, and volume shipments are slated for Q1 2026—just in time for Nvidia’s rumoured GB400 launch.

A leaked draft of NIST’s 139-page red-team study maps 139 novel ways to prompt leading AI models into leaking personal data, spewing disinformation or staging cyber-attacks—yet the document remains unpublished amid fears it could embarrass suppliers that now sell to federal agencies. Conducted under the AI 600-1 framework, the exercise exposed blind spots in existing jailbreak defences, especially around voice cloning and agent tool-use. Researchers involved say the findings could jump-start an industry-wide safety baseline, but political handlers shelved the report during the administration change-over. Civil-society groups are pressing for release before upcoming Senate hearings on AI procurement.

Google DeepMind unveiled Genie 3, a “world model” that turns text prompts into fully navigable 720 p environments running at 24 fps and persisting for minutes—orders of magnitude longer than Genie 2. The system remembers object placement, supports promptable events like sudden rainstorms and lets human or AI agents roam in real time. DeepMind calls it a stepping-stone toward embodied AGI but is limiting access to academics and creators while it studies safety risks. Game studios and robotics researchers are already lobbying for early APIs, seeing Genie 3 as a cheaper sandbox than bespoke simulators.

💹 SoftBank’s AI spree under investor lens

SoftBank will report first-quarter earnings tomorrow, and analysts expect the spotlight to be on its $11 billion surge in AI outlays, from Nvidia GB200 pre-orders to Vision Fund spin-outs. Insiders say the conglomerate has taken minority stakes in at least 15 foundation-model and robotics firms since April, seeking “optionality” rather than outright control. The spending blitz follows CEO Masayoshi Son’s vow to make SoftBank “the investment engine of the AI revolution,” but rating agencies warn leverage is creeping back toward 2019 highs. Bulls counter that early gains in chipmaker Arm justify the risk. All eyes now turn to whether Son will announce a fresh Vision Fund III.

🛠️ AI tools updates

Gemini’s new “guided learning” toggle turns the chat assistant into an interactive tutor: it breaks problems into bite-sized prompts, surfaces hints only when asked and ends each session with a short quiz and video recap. Designed alongside educators, the mode aims to shift AI from answer-spitter to learning companion. Eligible students in five countries also get a free year of Gemini Pro.

Experian’s new Assistant for Model Risk Management auto-drafts documentation, tracks statistical drift and flags breaches of Fed SR 11-7 or UK SS1/23 guidelines, cutting validation cycles by up to 70 % in pilot banks. Centralised templates and workflow hooks promise audit-ready trails without spreadsheet gymnastics.

💵 Venture Capital updates

Drug-design newcomer Chai secured $70 million from Menlo, DST and OpenAI to scale its Chai-2 model, which finds antibody binds at a 1-in-6 hit rate—far above traditional wet-lab odds. The fresh cash values the two-year-old firm at roughly $550 million and brings Pfizer’s former R&D chief onto the board.

Accounting-automation startup Rillet closed a $70 million round led by Andreessen Horowitz and ICONIQ. Its generative engine ingests invoices, chats with ERPs and resolves mismatched ledgers, claiming 90 % faster month-end close for mid-market clients. Funds go toward EU expansion and new tax-AI modules.

🫡 Meme of the day

⭐️ Generative AI image of the day