LONG AI
Home › Blog › Claude Fable 5.2 Tops: 49.1% on Coding

Claude Fable 5.2 Tops: 49.1% on Coding

Category:行业资讯 · 21 min read · 2026-11-10

This column aggregates the latest large-model releases and open-source drops from the past one to two days, cross-checked against official announcements and multiple media outlets. Performance figures distinguish between vendor claims and independently verified third-party results; anything not yet replicated is clearly marked. Illustrations are original.

Stacked glass panels with a small robot on top, connected by light arcs illustration
One base, two packagings: same capability, how much you get depends on the gate (illustration)

Headline: Anthropic splits its strongest model into two, coding jumps 16.7 points, price drops 20%

On Nov 9, Anthropic released Claude Fable 5.2 and Claude Mythos 5.2. The most easily misread point first: these are not two models but two packagings of one model. The official wording: "Fable and Mythos are the same underlying model" - Fable carries full cybersecurity and biosafety protections, so it opens to everyone; Mythos strips part of that protection, can probe vulnerabilities and do dual-use research, and is only issued to vetted institutions through the Cyber or Life Sciences Verification Programs.

The timing is delicate: Anthropic is sprinting toward an IPO (S-1 filed Oct 26). This is the first new frontier model since CEO Dario Amodei publicly called for "slowing down frontier AI" in September - slowing down in words, shipping in deeds.

Key figures: coding score up from 32.4% to 49.1% (a 16.7-point jump); API price down 20%.

Take: the "one model, two gates" packaging is a new answer to the capability-vs-safety tension - sell the full model only to those who pass a verification bar, sell the protected version to everyone. Expect this to become a template for frontier releases.

Other releases (6)

Model / ToolDateTypeHighlights
Meta Muse Ember 1.011-9Light execution model / closed APIMeta's smallest model, built for the thousand-times-repeated agent work: tool calls, small code, routing, extraction, long-doc parsing. 1M context, up to 512K output, native text/image/audio/video/PDF. Price is the whole weapon: $0.10 in / $0.50 out / $0.004 cache per million. Third-party AA index 36, tying Claude Opus 5.5\u2019s top tier, but $0.14/task vs $0.41. Vendor: AutomationBench-AA 37.8% (Haiku 13.9%, GPT-6 Luna 20.2%); but FrontierCode 2.0 only 27.2%, losing to Haiku's 34.9% - don't use it for code. First model trained with Muse Wonder Full. Weights closed.
ByteDance Seedream 5.5 (image)11-9Image generation / editingThree things: 100-round edits on one image without degradation (pixel-level preservation of non-edited regions); native alpha-channel transparent output, layerable assets; severe text corruption down to 0.4%, minor 1.7%. Vendor MagicArena: film and office usability more than doubled; photoreal image-to-image GSB +55.62%. Price: 2K square edit ~0.28 yuan, under half of 5.0 Pro, ~40% faster. All vendor-reported.
NVIDIA GR00T 2 / Turbo11-9Robot world-action model / open in DecUnlike "image to action" models, it predicts future frames and actions together - imagine the physics, then move. Data: 402,730 hours of first-person video + 10,184 hours handheld-gripper demos + 23,500 hours teleop. Zero-shot 40% average on 50 unseen tasks; 71% with one demo, 95% with five; adaptation is 30 minutes of LoRA on one RTX 5090. Turbo is 2B params, 38ms on a 5090, 124ms on Jetson Thor, runs on the Unitree G1. Foxconn and Siemens evaluating. Weights open in December.
Amazon HOPE (open method)11-9MoE expert pruning / Apache 2.0Solves a real money problem: MoE activates few experts per token but all experts must stay in VRAM (95% of Qwen3.5-122B's 250GB is expert weights). HOPE scores expert pairs and solves a quadratic program to drop half optimally. Cutting 50% of experts saves ~115GB on Qwen3.5-122B; pruned Qwen3.5-35B-A3B hits SWE-Bench Pro 42.3, beating the unpruned 41.8. Limits self-stated: intra-layer only, pairwise, advantage fades below ~40% pruning.
LingBot-Touch (Ant Lingbo, open source)11-9Tactile foundation model / Apache 2.0Gives robot hands "feel": three incompatible tactile sensors get per-sensor tokenizers sharing one Transformer mapped to "standard fingertip coordinates." Pretrained on 4,983 hours of real tactile data. Tiny 5.45M / Small 21.76M params, only +0.2% on downstream. Real-robot 26 tasks x 20: lifts \u03c00.5 48.1% to 69.9%, LingBot-VLA 48.3% to 73.1%. Scope note: only inference code + frozen encoder released; the 5,000-hour corpus and training code aren't public.
Zyphra ZONOS3-Mini (open TTS)11-9Speech synthesis / Apache 2.0Sparse MoE, 800M total / 100M active, an order of magnitude below the previous 8B. Where it runs is the point: browser WebGPU, Pixel 9 first audio under 570ms, M3 MacBook Air under 150ms. Trained on 112 H100s for 2.5 days. Zero-shot cloning with age/gender/accent control. No standalone Chinese score published - test Mandarin yourself; also mind deep-synthesis filing and voice authorization.

Industry news

The trend in one line

Frontier releases are being split by "who's allowed to use the full capability" - the verification gate is now part of the product.

(Data from official announcements and public reports, 2026-11-10; vendor claims and third-party results are marked separately, with unverified items noted in the text. Model capability affects selection, but actual pricing follows the model plaza in real time.)

相关文章

GPT-6.2 Cuts Prices 20%: What the AI Price War Means

GPT-6.2 cuts list prices 20%, yet platform discounts run deeper: gpt-6-astra trades at under 1% of list. What this week's AI price war means for buyer

What Is Prompt Caching? How Cache Tokens Cut API Bills

Prompt caching stores a repeated prompt prefix and bills cache reads at 5-12% of normal input. Real prices, hit conditions, and how to cut your API bi

All 12 Claude Models Compared: Price and Best Use

Twelve Claude versions are on sale. The newest opus-5-5 costs $8.91 a month; the oldest opus-4-6 costs $49.50. Full table and a pick for each workload

Which AI Model Writes Code Best? Speed and Price Tested

gpt-5.6-terra and deepseek-v4.1-flash win on price and reliability; gemini-3.8-flash is fastest. Effective prices, measured speed, and a three-tier pi

One API Key, Every Model: How to Integrate a Multi-Model Marketplace

Point your base URL at the platform and one key covers all 138 on-sale models across 51 vendors, in both OpenAI and Anthropic request formats. Copy-pa