AI Weekly Brief
Claude lifts a 37-year-old bound
An unreleased Claude research model raised the proven share of Riemann zeta zeros on the critical line to 67.2%, verified in Lean.
TL;DR
- A frontier release cluster landed inside 48 hours (Google's Gemini 3.7 Flash, xAI's Grok 4.6, DeepSeek's V4-Pro general availability, and Alibaba's first open-weight Qwen Max model), intensifying price and open-weight pressure on the closed frontier.
- OpenAI productized offensive-grade cyber capability for defenders: GPT-5.6-Cyber, delivered through the new Daybreak Red tier, found real zero-days including Chrome V8 CVE-2026-15903 and launched a partner program spanning Accenture, IBM, CrowdStrike, and Palo Alto Networks.
- An unreleased Claude research model raised a 37-year-old number-theory bound, lifting the proven fraction of Riemann zeta zeros on the critical line from 41.6% to 67.2% with a Lean-verified proof.
- Inference speed became a first-class product axis as OpenAI previewed Cerebras-powered Ultrafast (GPT-5.6 Sol at ~750 output tokens/sec) and Gemini 3.7 Flash halved Flash-tier pricing.
- Healthcare AI shifted from clearance toward coverage and platform embedding, as CMS opened the RAPID Medicare pathway and Oracle Health shipped a patient portal powered by OpenAI models.
Editor’s Pick
Asked simply to "take a real stab" at the Riemann hypothesis, an unreleased research version of Claude instead produced a genuine byproduct result: it raised the proven lower bound on the fraction of zeta zeros lying on the critical line from 41.6% to 67.2%, a bound mathematicians had improved by less than a percentage point over the prior 37 years. The methodology is as notable as the number. The model coordinated roughly 60 subagents across two Claude Code sessions and 31M output tokens, generating and discarding 650 failed ideas, running thousands of numerical checks against known zeros, and downloading 54 arXiv papers to confirm novelty. Decisively, it produced a machine-checkable Lean proof and its work was reviewed by external number theorists, so the claim rests on formal verification rather than a lab's self-assessment. The result is not progress on the Riemann hypothesis itself, and Anthropic says the techniques will not lead to a full proof, but it is a concrete, checkable contribution to open mathematics produced with heavy human scaffolding and prior human research. The signal for practitioners is the pairing of large-scale subagent orchestration with an automatic verifier: where a domain admits mechanical proof-checking, long-horizon agent swarms can now do exploratory research whose correctness does not depend on human trust.
Model Releases
Google's Gemini 3.7 Flash. Google's workhorse coding-and-agent model shipped three weeks after 3.6 Flash at half the price ($0.75/$3.75 per 1M tokens introductory), scoring 43.6% on FrontierCode 1.1 (vs 34.4%) and 65.3% on DeepSWE v1.1 (vs 49.0%), with updated CBRN and cyber safeguards.
DeepSeek V4-Pro reaches general availability. After a four-month preview, the flagship shipped with agent-focused re-post-training (Terminal-Bench 2.1 87.9, DeepSWE 62.7) at about $0.435/$0.87 per 1M tokens, roughly 46x cheaper than Claude Fable 5, with peak and off-peak pricing to follow.
xAI's Grok 4.6. A post-training upgrade on Grok 4.5 tuned for long-running agents, with a 500K-token context, a new xhigh reasoning-effort level, and $2/$6 per 1M-token pricing across Cursor, Grok Build, and the API.
Alibaba releases Qwen3.8-2.4T-A95B open weights. Alibaba shipped the weights it promised the week before, its first Max-class open release: a 2.4T-parameter (95B active) MoE with a 256K native context extensible to about 1M, on Hugging Face and ModelScope alongside an FP8 build.
Hugging FacePreviouslyAug 7NextAug 21
AI Tools & Products
OpenAI previews Cerebras-powered Ultrafast mode. A limited-preview API tier runs GPT-5.6 Sol up to 14x faster, at as much as 750 output tokens per second, aimed at latency-critical incident-response, trading, and voice workflows.
ChatGPT ads test expands to five countries. OpenAI extended its ads test to logged-in Free and Go users in the UK, Mexico, Brazil, Japan, and South Korea, keeping ads clearly labeled, independent of answers, and excluded from sensitive topics.
Research Breakthroughs
Claude raises the Riemann zeta critical-line bound. An unreleased research version of Claude, coordinating about 60 subagents across two Claude Code sessions and 31M output tokens, raised the proven fraction of zeta zeros on the critical line from 41.6% to 67.2%, formalized in Lean.
Shanghai AI Lab's Intern-S2-Preview. A 397B scientific agentic foundation-model series combining multimodal pretraining, multi-task and agentic RL, and on-policy distillation, with native time-series forecasting and a Memory Decoder that specializes the frozen backbone without modifying it.
Microsoft Research's full-bandwidth transformer. Fusing the previous top-layer hidden state with the sampled-token embedding through a gated linear unit widens the feedback channel between decoding steps, matching standard transformers trained on about 1.5x more tokens with negligible overhead and an intact KV cache.
Alibaba's DreamX-Phi 1.0 manipulation world model. An action-conditioned video world model that injects per-arm SE(3) transforms through geometric attention, plus a depth branch and object masks, to predict faithful manipulation rollouts, taking first place on Track 1 of the WorldArena 2.0 Challenge.
Workflows & Techniques
Meituan's AutoDesign rewrites its own harness. A meta-harness optimizer has a code agent recursively rewrite its harness from rollout feedback, lifting average PosterBench scores from 54.99 to 67.39 and beating the closed Claude Design system by 7.45 points with a frozen base model.
LycheeMemory V2 batches agent memory by segment. Replacing per-turn memory consolidation with semantic segment-level batching reaches 89.22% on LoCoMo and 92.20% on LongMemEval-S while cutting memory-construction tokens by up to 86% versus A-Mem, without raising query-time cost.
OpenDataLab's SKILLER writes skills for small models. A natural-language RL loop uses a strong model as actor and critic to generate executor-specific skills for small open models such as Qwen3.5-9B and 4B, adding up to 20.4 points and matching closed models on single-skill tasks.
Industry & Ecosystem
OpenAI's GPT-5.6-Cyber and Daybreak partner program. Delivered through a new Daybreak Red tier, GPT-5.6-Cyber completes 95.0% of advanced exploit-development prompts (versus 1.5% for GPT-5.6 Sol) and has disclosed real zero-days including Chrome V8 CVE-2026-15903 and 400+ kernel privilege-escalation bugs; partners include Accenture, IBM, CrowdStrike, and Palo Alto Networks.
Thrive Holdings raises $2B at $12B. The OpenAI-backed AI rollup raised from SoftBank, D1 Capital, and Altimeter to buy traditional accounting, IT, and physical-asset businesses and embed AI agents into their operations.
Progress to acquire Domo's AI platform for $400M. Progress Software agreed to buy substantially all assets of Domo's agentic AI and data-products platform for $400M in cash, broadening its enterprise data-readiness stack.
Anthropic to watermark AI-generated text. To comply with the EU AI Act Transparency Code, Anthropic will watermark text and files from all Claude models released after the code took effect, using the C2PA standard so the mark persists through copy-paste.
AI Startups
CodeRabbit raises $143M at $1.5B. The AI code-review company, running more than 2M reviews a week for 17,000+ customers, raised from Atomico and Smash Capital and launched Agentic Change Management, a governance layer for human- and agent-written code.
Skan AI raises $63M Series C. Skan builds process-intelligence maps by observing how office staff work, then uses them to design and deploy enterprise agents.
Corma Labs raises $60M for defensive security AI. Backed by Sequoia, Khosla Ventures, and Coatue, Corma is building a defense-specific security foundation model whose agents onboard like new hires and reportedly cut threat-response times by more than 94% at Fortune 100 and 500 customers.
Noosphere Labs raises $10.25M for physical AI. A former Meta AI research director raised from Trilogy Equity Partners and Madrona for a stealth startup pursuing human-centered physical intelligence beyond text and images.
Healthcare AI
CMS opens the RAPID Medicare coverage pathway. The pathway issues a proposed Medicare National Coverage Determination the same day the FDA authorizes an eligible Breakthrough device, compressing coverage from roughly a year to about two months.
Oracle Health launches an AI patient portal. The OpenAI-model-powered portal, now generally available in the US, grounds plain-language explanations of records, labs, and scheduling in the EHR, with non-diagnostic guardrails and source citations.
Atlantic Health and K Health launch PatientGPT. Deployed inside Epic MyChart as an EHR-integrated digital front door, PatientGPT grounds symptom triage in a patient's own record and escalates to same-day virtual primary care.
FDA clears Roen Surgical's Zamenix kidney-stone robot. Described as the first AI-based flexible ureteroscopic kidney-stone surgical robot, Zamenix checks in real time whether a stone is small enough to pass and fragments larger stones by laser.
Epic's Chart with Art extends to nursing. Northeast Georgia Health System became the fifth US health system to pilot Epic's ambient documentation for nurses, extending the clinician scribe into inpatient nursing workflows.
Northeast Georgia Health SystemPreviouslyAug 7
Trends & Predictions
Chinese labs take the lead in open-weight models
Open-weight modelsStrengthenedContinues fromAug 7
Alibaba shipped the promised Qwen3.8-2.4T-A95B, its first Max-class open weights, and DeepSeek took V4-Pro to general availability at about $0.435/$0.87 per 1M tokens, roughly 46x cheaper than Claude Fable 5.
- Prediction (Feb 2027): Closed labs keep answering on price and speed rather than raw capability, and more Western teams pilot self-hosted Chinese weights for cost-sensitive agentic work.
- Resolves: Yes if by February 14, 2027 OpenAI, Anthropic, and Google have each launched or repriced a top-tier API model at least 30% below the per-token price of their top tier on August 14, 2026, per their official pricing pages.
Based on Alibaba releases Qwen3.8-2.4T-A95B open weightsDeepSeek V4-Pro reaches general availability
Frontier cyber capability ships to defenders behind gated access
AI cyber offense and defenseStrengthenedRevivesJun 26
OpenAI's GPT-5.6-Cyber, gated behind a Daybreak Red tier, completes 95.0% of advanced exploit-development prompts versus 1.5% for GPT-5.6 Sol and has disclosed real zero-days, while Corma Labs raised $60M for a defense-specific security foundation model.
- Prediction (Feb 2027): Governed access tiers, hardware-key requirements, and auto-review sandboxing become standard around cyber-capable models, and managed AI-versus-AI defense programs proliferate through systems integrators.
- Resolves: Yes if by February 14, 2027 at least three of OpenAI's Daybreak partners Accenture, IBM, CrowdStrike, and Palo Alto Networks launch a managed security service built on GPT-5.6-Cyber or a successor, per their press releases.
Based on OpenAI's GPT-5.6-Cyber and Daybreak partner programCorma Labs raises $60M for defensive security AI
OpenAI previewed a Cerebras-powered Ultrafast tier running GPT-5.6 Sol up to 14x faster at 750 tokens per second, DeepSeek priced V4-Pro about 46x below Claude Fable 5 with peak and off-peak rates to follow, and Gemini 3.7 Flash launched at half 3.6 Flash's price.
- Prediction (Feb 2027): Speed-class tiers and specialized inference hardware move into frontier serving, and cost and latency per completed task remain the primary axis of model competition.
- Resolves: Yes if by February 14, 2027 OpenAI makes Ultrafast generally available and Anthropic or Google offers a frontier model on Cerebras, Groq, SambaNova, or Etched hardware as a paid API tier, per official pricing pages.
Based on OpenAI previews Cerebras-powered Ultrafast modeDeepSeek V4-Pro reaches general availabilityGoogle's Gemini 3.7 Flash
Agents improve themselves through skills, harnesses, and memory
Agent training and memoryStrengthenedContinues fromJul 31
Meituan's AutoDesign has a code agent recursively rewrite its own harness, lifting PosterBench from 54.99 to 67.39 and beating Claude Design by 7.45 points with a frozen model, and SKILLER's generated skills add up to 20.4 points to small open models.
- Prediction (Feb 2027): Harness and skill search become a standard way to convert evaluation compute into durable capability from frozen models, reducing reliance on retraining for incremental gains.
- Resolves: Yes if by February 14, 2027 at least two of OpenAI, Anthropic, Google DeepMind, and Meta publish research or engineering posts describing automated harness or skill search used to improve a shipped agent.
Based on Meituan's AutoDesign rewrites its own harnessOpenDataLab's SKILLER writes skills for small models
Clinical AI moves into EHR platforms and reimbursement
Clinical AIStrengthenedContinues fromAug 7
CMS's RAPID pathway proposes Medicare coverage the same day the FDA authorizes an eligible Breakthrough device, cutting about a year to two months, while Oracle Health's portal, PatientGPT in Epic MyChart, and a nursing pilot of Chart with Art embed AI in the EHR.
- Prediction (Feb 2027): Reimbursement pathways such as RAPID and the proposed Software-as-a-Medical-Service payment framework, together with EHR-native patient- and clinician-facing AI, become the deployment default, with coverage rather than clearance increasingly the gating step.
- Resolves: Yes if by February 14, 2027 CMS issues a proposed National Coverage Determination through the RAPID pathway for at least one AI-enabled Breakthrough device, per the CMS Medicare Coverage Database.
Based on CMS opens the RAPID Medicare coverage pathwayOracle Health launches an AI patient portalAtlantic Health and K Health launch PatientGPTEpic's Chart with Art extends to nursing