이미지: X — 뉴스 앰프 영상 갈무리
Summary
- OpenAI unveiled an 'Ultrafast' mode, in limited API preview, that runs GPT-5.6 Sol up to 14 times faster on Cerebras hardware
- Anthropic said an unreleased research version of Claude raised the lower bound on the Riemann Hypothesis from 41.6% to 67.2%, and it upgraded the misalignment risk rating of an internal model called 'Model 2' to "low"
- The two companies also saw overlapping personnel moves the same week, including former OpenAI COO Brad Lightcap's departure and the arrival of a new CRO
- Ultrafast 모드
- GPT-5.6 Sol을 Cerebras 하드웨어에서 최대 14배 빠르게 구동, 제한적 API 프리뷰
- Daybreak 확장
- 보안 이니셔티브 Daybreak를 AWS 환경으로 확대
- 리만 가설 하한
- 미공개 연구용 클로드 모델이 41.6%에서 67.2%로 상향
- 내부 모델 Model 2
- Mythos 5보다 다소 향상, 정렬 실패 위험 등급 '매우 낮음'→'낮음'
- 인사 이동
- 달리 라지치 CRO 합류, 데니즈 드레서·전 COO 브래드 라이트캡 퇴사
- Sonnet 5 요금
- 출시 초기 요금제를 영구 적용으로 전환
- 텍스트 워터마크
- 앤스로픽이 도입 예정인 클로드 텍스트 워터마크 방식을 설명
In a week that opened with an experiment letting ChatGPT users get answers up to 14 times faster, Anthropic reported around the same time that an unreleased research version of Claude had pushed forward the lower bound on the Riemann Hypothesis, a math problem that has resisted proof for more than a century. Taken together, the announcements the two companies made within the same week reveal where each is currently putting its energy.
14x faster inference, put to the test on Cerebras
OpenAI released, in limited API preview, an 'Ultrafast' mode that runs GPT-5.6 Sol up to 14 times faster. It reportedly runs on hardware from chipmaker Cerebras, with the stated goal of cutting the wait time for responses. The feature is currently in an early stage open only to select developers, so it's too soon for regular ChatGPT users to notice the difference directly.
Dedicated security stack expands to AWS
OpenAI had previously announced on August 10 the release of GPT-5.6-Cyber, a model dedicated to cybersecurity, saying it came with a Daybreak Blue track for defensive work and a Daybreak Red track for authorized vulnerability research. This week, the company said it had extended this Daybreak system to the Amazon Web Services (AWS) environment as well. The principle that only approved defense teams can access GPT-5.6-Cyber reportedly remains unchanged.
Claude's touch on the Riemann Hypothesis: lower bound to 67.2%
Anthropic said an as-yet-unreleased research version of Claude raised the lower bound figure related to the Riemann Hypothesis from 41.6% to 67.2%. The Riemann Hypothesis is a long-standing problem in mathematics that describes the distribution of prime numbers. The AI did not complete the proof itself, but the fact that it pushed the related figure up so significantly in a short period is a result likely to draw attention from the math community.
Internal model 'Model 2': risk level went up, not down
In its second Risk Report, Anthropic disclosed the existence of an internal model called 'Model 2' that has not been released externally. The company said this model performs somewhat better than the earlier 'Mythos 5,' but it also raised the model's misalignment risk rating one notch, from "very low" to "low." In effect, the company itself pointed out that as models grow stronger, the risk of unintended behavior also becomes something that needs to be managed.
Watermarks and people on the move
Anthropic explained a text watermarking method it plans to introduce soon, changed the Claude feature in Chrome's side panel into 'Cowork' sessions, and decided to permanently apply the early-access pricing for Sonnet 5. In the same week at OpenAI, Dali Rajic joined as Chief Revenue Officer (CRO), Denise Dresser announced her departure, and former Chief Operating Officer (COO) Brad Lightcap also announced he is leaving the company. Set alongside the product announcements, this suggests that organizational changes are underway at both companies at the same time.
Editor's take
Laying out a week's worth of announcements side by side makes the strategic differences between the two companies stand out clearly. OpenAI is pushing on two fronts at once—speed (Ultrafast) and security (Daybreak/Cyber)—continuing to expand tools that enterprise customers can actually put to use. Anthropic, on the other hand, released a research achievement (raising the Riemann Hypothesis lower bound) in the same week as a risk report that voluntarily raised its own model's risk rating, repeatedly sending the message that "we manage capability and risk together." This isn't a coincidence—it's a signal that the two companies are speaking to different audiences. OpenAI is talking to developers and enterprise IT staff, while Anthropic is talking to regulators and the research community.
From a practical standpoint, there's a clear line between what to act on and what to set aside. Ultrafast mode is still in a limited API preview stage, so it can't be put into production right now, but teams that have been struggling with response latency should consider getting on the waitlist. News like the Riemann Hypothesis result or Model 2, on the other hand, is best read as research- and PR-oriented and not something with immediate operational impact. The rise in the misalignment risk rating shouldn't be taken as alarming either—it's better understood as a normal curve in which model capability and risk management grow together.
In the past, a risk-rating upgrade like this might have quietly passed as a footnote, but what's changed recently is that both companies have settled into a practice of regularly publishing risk reports. Points worth watching next week include whether Ultrafast mode's public rollout expands, and whether Anthropic releases the research model behind the Riemann Hypothesis work to the outside world in some form.
