ANG⁺ #5 · A Cheaper Claude Fable 5.2, +5 More
Claude Fable 5.2 is 20% cheaper and up to 30% faster. Plus GPT‑6 Cyber for security work, 15-second video from Gemini Omni 1.2, two new Google TPUs, NVIDIA's Korea investments, and dates for your calendar.
ANG⁺ #5 · A Cheaper Claude Fable 5.2, +5 More
about 4 min
ANG⁺ = briefings sorted by AI and checked by the editor
Welcome to the fifth ANG⁺ Briefing. Briefing #4 (November 5) was about tools you could download and try yourself. This edition is about where the top models are heading. From November 5 to 9, Anthropic and OpenAI each spelled out who gets their strongest and security-focused models, and on what terms, Google shipped a video model and new chips, and NVIDIA announced investments in Korea. Everything is built on first-party material: official company announcements and developer changelogs.
1. Anthropic Claude Fable 5.2 and Mythos 5.2 — lower price, higher speed
On November 9, Anthropic released its top models, Claude Fable 5.2 and Claude Mythos 5.2. They are the same underlying model. Fable ships with cybersecurity and biology safeguards and is open to all users; Mythos is open only to organizations that pass Anthropic's cyber and life-sciences verification programs.
- Price and speed:
$8per million input tokens and$40per million output tokens — 20% less than 5.1 ($10/$50). Output is up to 30% faster than 5.1, and an optional fast mode runs up to 2.5x faster at$16input /$80output. The context window is 1 million tokens, with up to 128k output tokens. - The company's numbers: 49.1% on the agentic coding test FrontierCode 2.0 (Opus 5.5 45.5%, GPT-6 Astra 42.8%), 89.3% on the computer-use test OSWorld 2.1 (Opus 5.5 87.2%), and 41.3% on the Remote Labor Index (Opus 5.5 21.3%). In prerelease testing, a team of about 1,000 agents formalized the proof of Fermat's Last Theorem in the Lean language in 30 hours.
- Fewer blocks: The safety classifiers block about 55% fewer requests than with 5.1 (about 60% fewer in cyber, about 40% fewer in biology). Blocked cyber questions are answered by Opus 4.8 instead.
- What developers need to change: Computer use now runs through the new
computer_toolset_20260801toolset. Declaring the oldercomputer_20251124tool returns a 400 error on every platform except Amazon Bedrock. The prefix check — which errors if you edit earlier context and send it back — now applies to every account, and 30-day data retention is required. The model ID isclaude-fable-5-2.
Our take: Anthropic still recommends starting most work with Opus 5.5. Even at a lower price, the top model is where you send the hard jobs. But if you run computer-use agents, rename the tool before you switch models.
Source: Introducing Claude Fable 5.2 and Claude Mythos 5.2 (Anthropic) · Claude Platform release notes (Anthropic)
2. OpenAI GPT‑6 Cyber — a model that finds vulnerabilities but stays inside the scope you set
On November 5 (09:00 UTC), OpenAI released GPT‑6 Cyber, a model for security research. It finds vulnerabilities, checks whether they can really be exploited, and proposes fixes. The part OpenAI says it worked hardest on is not touching anything outside the scope a researcher sets.
- Capability: 48.5% on ExploitGym, which tests exploiting a specified vulnerability on a remote target — about 16 points above GPT‑5.6 Cyber (32.1%). OpenAI's researchers say they used it to develop 948 Linux privilege-escalation exploits and to find 81 new Firefox vulnerabilities that Mozilla classified as high severity.
- Staying in bounds: On ExploitGym Honeypot, which deliberately leaves easy-to-attack side services in reach, it avoided out-of-scope services in 99.93% of cases without its deployment safety prompt (GPT‑5.6 Cyber 94.49%), and 100% with it.
- Who gets it: At first, a select group of approved Daybreak Red customers through the API and Codex, expanding over time. Daybreak Red API access is restricted to approved IP addresses by default. Under OpenAI's Preparedness Framework, it is rated "Critical" in cybersecurity. A day later (November 6), OpenAI announced a partnership with security firm Hacktron to hunt for vulnerabilities in its own systems.
Our take: A security model's report card now carries "does it avoid breaking into what it wasn't asked to?" right next to "how well does it break in?" Mythos in item 1 follows the same path — top-tier cyber capability goes to verified defenders first, not the general public.
Source: GPT‑6 Cyber: Advancing the frontier of aligned cybersecurity agents (OpenAI) · Partnering with Hacktron to find vulnerabilities in OpenAI’s systems (OpenAI)
3. Google Gemini Omni 1.2 Flash goes GA — 15-second video, 30% cheaper
On November 5 (17:00 UTC), Google made its video model Gemini Omni 1.2 Flash (gemini-omni-1.2-flash) generally available. It is the first time Gemini Omni has reached GA.
- Length: Up to 15 seconds in one generation (Omni 1.1 topped out at 10). You can set the duration from 3 to 15 seconds in one-second steps, and extend a clip turn by turn up to 30 seconds in total.
- Consistency: Up to 6 reference images of characters, products, or styles, so faces, garments, packaging type, and brand colors hold across camera angles. Instead of regenerating, you can change the background, lighting, or aspect ratio in conversation.
- Price: 30% lower than 1.1 —
$12.00per million video output tokens on the API, or$0.07per second for 720p video. Paid-tier rate limits doubled. Anyone with a Google account can make 15-second clips in Google Vids at no additional cost. - Automatic switch: From December 3, requests to
gemini-omni-1.1-flashroute to 1.2 and are billed at 1.2 rates.
Our take: A short-form ad is 15 seconds. When "one ad in one generation, no stitching" becomes normal, the unit of video work shifts from the cut to the whole spot. If you use 1.1, switch the model name before December 3 and check the bill yourself.
Source: Gemini Omni 1.2 Flash: 15-second video and lower pricing (Google) · Gemini API changelog, November 5, 2026 (Google)
4. Google's third and fourth 8th-gen TPUs — 8tx for confidential inference, 8i Max for long context
On November 9 (13:00 UTC), Google added two systems to its eighth TPU generation. It launched TPU 8t for training and TPU 8i for inference in April; with TPU 8tx and TPU 8i Max, it says this is the first time it has introduced four TPU systems in a single year.
- TPU 8tx — data stays locked while it is processed: Data is normally encrypted only at rest and in transit, and unlocked in memory during computation. 8tx uses the same silicon as 8t, paired with 5th Gen AMD EPYC CPU hosts, to keep data and model weights isolated even while in use. It targets regulated industries and governments that fine-tune models and run agents on sensitive data.
- TPU 8i Max — the memory wall: Two TPU 8i chips on one package, with double the compute, SRAM, and HBM bandwidth of TPU 8i, fitting twice the silicon in the same rack footprint. It targets the memory-bandwidth bottleneck when dozens of agents work across a whole codebase or hours of video.
- Software alongside: Google open-sourced TPU Sync, which joins chip memory, host memory, and storage into one layer, and made Cloud Pathways and GKE dynamic slicing generally available.
Our take: The point is that one chip generation is now split into training, inference, confidential, and long-context systems. Chips feel far from most readers, but these partitions come back later as line items on cloud price sheets.
Source: Introducing TPU 8tx and TPU 8i Max (Google)
5. NVIDIA × Korea's Ministry of Science and ICT — startup investments and wider sovereign AI work
On November 8–9, NVIDIA said it is expanding its collaboration with Korea's Ministry of Science and ICT (MSIT). At the MSIT-NVIDIA AI Startup Partnership Forum at COEX in Seoul, Vice Minister Ryu Je-myung met NVIDIA leaders and over 20 Korean startups.
- Investments: NVIDIA invested in Point2, which builds radio-frequency interconnects that move data between chips and racks, and plans investments in Motif Technologies (open-weight language model Motif-3), RLWRLD (robotics foundation models for five-finger dexterity), and Upstage (Solar language models and document AI). NVIDIA's startup program Inception has more than 1,000 Korea-based members.
- Chips and infrastructure: Korean NPU and XPU companies will explore joining the NVLink Fusion ecosystem, and NVIDIA announced a collaboration with SK Telecom and Rebellions, supported by MSIT, on sovereign AI infrastructure built with NVLink Fusion.
- Models: Through Korea's AI for All initiative, NVIDIA will work with three selected consortia and four leading Korean foundation model builders, using NVIDIA Dynamo to optimize inference. Training, certification, and hackathons are part of the package.
Our take: The four investments and planned investments spread across interconnect chips, language models, robotics, and enterprise document AI. It signals that NVIDIA sees Korean AI as a full chain from chips to applications, not a single model. The amounts were not disclosed.
Source: NVIDIA Announces New Planned Investments in Korean Startups at the MSIT-NVIDIA AI Startup Partnership Forum (NVIDIA) · From Chips to Skills, NVIDIA and Korea Expand Sovereign AI Collaboration (NVIDIA)
6. New this week, and dates for your calendar
Smaller changes from November 5–9, plus the dates within a month when you may need to change code or settings.
- Personal Finance in Claude, beta (November 5): Link bank, credit card, loan, and investment accounts to Claude, see balances and spending on one page, and ask about your own money. Pro, Max, and Ultra plans, US only. The same day, Claude Enterprise added a security center, inference hooks, and centralized management.
- IP allowlist in the Claude Console, beta (November 5): Only requests from listed addresses are accepted, which limits where a leaked API key can be used. Add IPv4 and IPv6 addresses or CIDR ranges, up to 1,000 per organization.
- Cursor voice chat (November 6) and live apps (November 5): Talk with your agents in a realtime voice call and hear progress updates; voice chat runs on Grok Voice. You can also build an app in Cursor — database, sign-in, and deployment included — and share it with a link.
- Gemini API (November 5–6): Function responses can now return audio, video, and Office files to the model, and PPT, ODP, and ODS files are accepted as input. Since November 6, requests to
gemini-embedding-2-previewroute togemini-embedding-2(no re-indexing needed). - ChatGPT Enterprise in AWS Marketplace (November 9): Buy and bill it through an AWS account and apply eligible spend toward existing AWS commitments.
- November 30: Claude Sonnet 4.5 (
claude-sonnet-4-5-20250929) retires from the API (covered in #4). - December 1:
gemini-2.5-proandgemini-2.5-flashswitch automatically (covered in #4). - December 3:
gemini-omni-1.1-flashswitches automatically togemini-omni-1.2-flash(item 3).
Our take: Three small items this week — the Console IP allowlist, GPT‑6 Cyber's IP restriction, and Fable 5.2's 30-day retention — point the same way. The stronger the model, the more companies ask who is calling it, from where, and with what record kept. If your keys are only used from known places, turn on the allowlist first.
Source: Claude release notes (Claude Help Center) · Claude Platform release notes (Anthropic) · Cursor changelog (Cursor) · Gemini API changelog (Google) · ChatGPT Enterprise now available in AWS Marketplace (OpenAI)