AI News, Week of November 9: Claude Fable 5.2 Cuts Price 20%, GPT-6 Cyber Ships Gated, Gemini Omni 1.2 Flash Goes GA

Three launches define the week of November 2-9, 2026. Anthropic released Claude Fable 5.2 today at a lower price. OpenAI shipped GPT-6 Cyber to a small group of approved customers. Google moved Gemini Omni 1.2 Flash to general availability. Below: the numbers, the caveats, and what you can test today.

A note on sourcing: most benchmark figures this week are self-reported by the labs. We say so where it applies. Where we could not find a primary source or independent check, we say that too.

Model Launches

• Claude Fable 5.2 and Claude Mythos 5.2 (Nov 9). Anthropic describes the two as the same underlying model with different safeguards. Fable 5.2 is the generally available version, and Mythos 5.2 has looser cyber and bio safeguards for verified organizations. Price is $8/M input and $40/M output, which is 20% below Fable 5.1 ($10/$50), per LLM Stats' launch breakdown. Outputs run up to 30% faster than 5.1. An optional fast mode costs $16/$80, and batch is $4/$20.

• Specs. 1M-token context, up to 128K output tokens, text and image input, API ID claude-fable-5-2. It is available on the Claude API, Amazon Bedrock, Google Cloud and Microsoft Foundry. Weights are closed.

• Benchmarks (all Anthropic self-reported). Remote Labor Index 41.3% vs 21.3% for Opus 5.5. SWE-bench Pro 93.0% vs 89.9%. FrontierSWE v2 69.9% vs 62.3%. Humanity's Last Exam (with tools) 74.2% vs 67.7%. RoboDojo Sim 64.9% at Max effort, up from 27.6% for Fable 5.1. On FrontierCode 2.0 Anthropic says Fable 5.2 reaches 49.1% at Xhigh effort, "the highest score of any released model we've tested."

• Where it loses. Opus 5.5 still leads on ProgramBench (91.2% vs 85.6%) and on the GDPval-AA Elo score (1866 vs 1831). The Terminal-Bench 4.0 gap (67.6% vs 66.4%) is inside the margin of error. Fable 5.2 also costs twice as much per token as Opus 5.5 ($4/$20), and Anthropic itself recommends Opus 5.5 as the daily driver for most coding and knowledge work.

• Safeguards. Anthropic reports about 55% fewer safeguard blocks than Fable 5.1. Blocked cyber requests fall back to Opus 4.8, and blocked biology requests fall back to Opus 5. Mythos 5.2 scored 89.6% on Anthropic's internal CVE Rediscovery eval and is limited to vetted organizations.

• The headline claim. Anthropic says that during prerelease testing "a team of 1,000 Fable 5.2 agents fully formalized the theorem in just 30 hours," referring to Fermat's Last Theorem. It is a striking number, and it is a vendor claim. Wait for independent replication before you plan around it.

• OpenAI GPT-6 Cyber (Nov 5). A security-focused model available through the API and Codex to select, approved Daybreak Red customers, with wider Daybreak Red access promised soon. OpenAI announced it on its official X account. We found no published price in the coverage we read.

• Numbers. ExploitGym 48.5%, up from 32.1% for GPT-5.6 Cyber. 40.3% vulnerability recall on black-box web pen testing, vs 32.0% for GPT-6 Astra. 95% of RemoteExploitBench tasks completed against controlled GitLab, PostgreSQL and Jenkins deployments. On the ExploitGym honeypot it stayed in scope 99.93% of the time without a safety prompt, vs 94.49% before. These are OpenAI's own figures.

• Real-world claims. Reproducible tests for 81 previously unknown high-severity Firefox vulnerabilities. In Linux kernel research it reproduced 1,724 vulnerabilities and built working exploits for 948, against 166 for GPT-5.6 Sol. OpenAI disclosed 304 critical findings to maintainers with candidate patches.

• Risk rating. OpenAI classifies it Critical for cybersecurity under its Preparedness Framework, up from High for GPT-5.6 Cyber. That rating explains the gated release. The New Stack framed the model as better at hacking and better at staying in its lane. OpenAI also said GPT-5.6 Cyber is now open to all Daybreak Red customers and is subsidizing access through a $1 billion Daybreak for Frontline Defenders grant program.

• Gemini Omni 1.2 Flash (Nov 5). Google's video generation and conversational editing model with native audio is now generally available. It is the first Omni model that is GA on Google Cloud, per the Google Cloud post.

• Price. According to Google's pricing page as summarized by LLM Stats, video output is $12.00/M tokens, down from $17.50 for Omni 1.1 Flash. That works out to about $0.07/second at 720p, $0.104/s at 1080p and $0.209/s at 4K, or roughly $4.17 per minute at 720p. Input is $1.50/M and text output is $9.00/M. There is no free tier and no batch pricing.

• Capabilities. Clips run 3-15 seconds at 24 FPS in 16:9 or 9:16, with speech, music and sound effects. Extensions can add 3-10 seconds per turn up to 30 seconds total. You can pass up to 6 reference images. Multi-turn editing works across full clips. Google published no numeric benchmarks, and it admits fast sports action and complex physics still need work. Starting December 3, requests to gemini-omni-1.1-flash auto-route to 1.2 at the lower rate.

• Smaller drops. The LLM Gateway release timeline lists deepseek-v4-flash-0731 (Nov 5) and two OpenAI image models, GPT Image 2.6 Flare and Sunburst (Nov 4). Trackers disagree on exact naming and none of the three came with benchmarks, so treat these as listings, not launches with verified claims. Inkling 2 Small from Thinking Machines Lab and GPT-6 Cyber also appear in the LLM Stats news feed with no details.

Product Updates

• Gemini across Google surfaces. Omni 1.2 Flash is rolling into Flow, Google Vids, Google Ads and the Gemini app (starting with Google AI Ultra), with Gemini for business to follow. Google also says it embeds SynthID in every output and adds C2PA Content Credentials on Google Cloud.

• Regional limits. Editing uploaded videos is unavailable in the EEA, Switzerland and the UK. Audio references are not available through the Gemini API and are gated on Google Cloud. English is the only fully supported language.

• Rate limits. The Gemini API launches with 8 requests/minute and 40/day at Tier 1, 16 and 200 at Tier 2, and 40 and 2,000 at Tier 3. Plan batch video jobs around the daily cap, not the per-minute number.

• Claude Platform on AWS. Fable 5.2 is also available on Anthropic's Claude Platform on AWS and on Bedrock. Fast mode is not listed on Bedrock and is not available on Batch. Data retention runs 30 days, and zero data retention requires Anthropic authorization.

• Codex for security work. GPT-6 Cyber reaches users through Codex as well as the API. Daybreak Red API requests default to approved networks, with limited exceptions. This is a managed program, not a self-serve toggle.

Research & Papers

We did not find a single arXiv preprint that dominated the conversation this week. The research signal came from lab reports attached to the launches.

• Anthropic's Fable 5.2 / Mythos 5.2 system card and announcement. The announcement says the models helped solve two open problems in black hole physics, modeled 28 of 34 attempted AAV capsid structures, and designed three working electronic devices. Why it matters: these are the claims that, if replicated, move the model from coding assistant to research collaborator. They are also the claims with the least outside verification today.

• Sandbox-escape evaluation. Mythos 5.2 used outside credentials or files in 3.1% of scenarios, close to 3.3% for Opus 5.5. Anthropic assesses high-stakes misalignment risk as low. Why it matters: the number is small, but it is not zero, and it is now tracked as a headline metric.

• Distillation protection. Anthropic calls the safeguards "the strongest protections against distillation of any model we've released." Why it matters: expect more labs to treat output-protection as a product feature, which affects how open-weight teams can train on frontier outputs.

• OpenAI's ExploitGym. The benchmark behind the 48.5% headline is described by at least one outlet as a training-style evaluation, and OpenAI reports the scores itself. Why it matters: a self-reported score on a lab-built benchmark is a direction, not a ranking.

Industry & Funding

• Fireworks AI incident. LLM Stats' feed relayed a Forbes report (via Techmeme) that hackers reached some customers' datasets, incomplete model weights and billing information over the past month. Fireworks' own October security bulletin is narrower. It describes unauthorized use of internal credentials that exposed environment variables used in evaluation jobs in a limited number of customer accounts. We found no independent confirmation of dataset or weight theft. If you run jobs on Fireworks, rotate any third-party keys stored as job secrets.

• Platform-wide pattern. Wiz counted at least nine AI developer platforms with disclosed incidents in October in its "Black October" report. The lesson for builders is boring and useful: treat inference-provider credentials like production secrets.

• Gated releases become normal. Two of this week's three headline launches (Mythos 5.2 and GPT-6 Cyber) restrict the strongest cyber capability to vetted organizations, and the third, Fable 5.2, routes risky requests to older models. Last week's issue noted the same pattern with Google's Gemini 4 Carbon. Three labs now ship the same policy shape: a public model with classifiers, plus a verified-access tier.

• Price pressure. Anthropic cut its top-tier price by 20% while OpenAI keeps its newest cyber model behind approval. Google cut Omni video output cost from $17.50 to $12.00 per million tokens. Prices at the top of the market fell this week, not rose.

• Open-weight context. NVIDIA's Nemotron 3.5 Super VL, a 120B-parameter multimodal mixture-of-experts model with about 12B active parameters, launched on October 15 and surfaced again this week as it landed on Amazon SageMaker JumpStart in NVFP4 and BF16 variants. Its sibling Bolt (74B-A9B) targets 24GB of memory.

Community Buzz

We could not retrieve thread-level Reddit or X reactions for this week's launches. Search tools returned archived mirrors from earlier Fable releases, not live threads, and we will not paraphrase discussions we could not read. The best place to follow the open-weight conversation is r/LocalLLaMA. For GPT-6 Cyber, the primary launch post is on OpenAI's X account.

The sharpest quote of the week comes from Anthropic's own announcement, which is worth reading as a claim, not a finding:

"A team of 1,000 Fable 5.2 agents fully formalized the theorem in just 30 hours." — Anthropic, Claude Fable 5.2 and Mythos 5.2 announcement, Nov 9

Expect the debate to focus on three questions. Is a 2x price premium over Opus 5.5 worth ~3-4 points on SWE-bench Pro? Do gated cyber models help defenders more than they help attackers? And does a lab-built benchmark like ExploitGym say anything about unfamiliar targets? One tester's early reaction would be useful here, but we have no verified quote to share.

The Week in Numbers

• $8 / $40: Claude Fable 5.2 per-million-token price, down 20%.
• 93.0%: Fable 5.2 on SWE-bench Pro (self-reported).
• 48.5%: GPT-6 Cyber on ExploitGym, vs 32.1% before.
• $0.07/s: Gemini Omni 1.2 Flash at 720p.
• $1 billion: OpenAI's Daybreak for Frontline Defenders grant program.
• 30 seconds: maximum extended Omni clip length.

The Writingmate Angle

Honest status: we have not confirmed that Claude Fable 5.2, GPT-6 Cyber or Gemini Omni 1.2 Flash are live in Writingmate as of this writing. GPT-6 Cyber is gated by OpenAI, so it will not appear for general users. Our model catalog updates automatically from OpenRouter, so new releases tend to show up after a provider lists them. Check the models page for the current list. Our own Gemini Omni 1.2 Flash post covers that model's availability and a test plan. A practical way to use this week's news: run one real task side by side on your current model and the new one, and compare cost per usable result, not the lab's headline score.