AI Model & Benchmark Watch — November 13, 2026

Claude Fable 5.2 landed at 65 on the Artificial Analysis index, five points clear of Gemini 4 Carbon, while Google's Gemini 4 Hydrogen and Meta's Muse Ember pushed the cheap end of the market and the open-weight side had a quiet week.

Overview

Anthropic released Claude Fable 5.2 on November 9 at $8 / $40 per million tokens, down from Fable 5.1's $10 / $50. Artificial Analysis (AA) puts it first at 65 on its Intelligence Index, up 16 points from the 49 we recorded for Fable 5.1 last week. That is the biggest single jump this beat has logged, and it also carries the highest cost per index task on AA's list ($26.54). Below the top, Google shipped Gemini 4 Hydrogen (AA 54 at $1.50 / $7.50) and Meta shipped Muse Ember at $0.10 / $0.50. No new open-weight model reached the leaderboard; Kimi K3.1 Pro (46) is still the best downloadable score.

New & Updated Models (this week)

Commercial / closed

Claude Fable 5.2 (and Mythos 5.2) — Anthropic, November 9, 2026

Who / License: Anthropic; closed. Mythos 5.2 is reported as the same model with looser cyber and bio safeguards for verified organizations.

What's notable: $8 input / $40 output per million tokens, 1M context, 128K max output; batch is $4 / $20 and a fast mode is $16 / $80 (per llm-stats' summary of the launch). AA measured 65 at max effort, 62 at high and xhigh, with an output speed of 97 tokens/s and 275M output tokens to run the index, against a median of 72.8M. So the model is slow and wordy at max effort, and AA's cost-per-task figure reflects that. Anthropic's own comparison table, which I could only see through llm-stats because Anthropic's announcement URL returned a 404 for me, has Fable 5.2 at 93.0% on SWE-bench Pro (Opus 5.5: 89.9%), 67.6% on Terminal-Bench 4.0 (Opus 5.5: 66.4%), and 41.3% on Remote Labor Index (Opus 5.5: 21.3%). Opus 5.5 beats it in the same table on ProgramBench (91.2% vs 85.6%) and on two AA Elo evals, GDPval-AA and AA-Briefcase. Anthropic chose the harnesses for most of these, and llm-stats notes the GPT-6 Astra and Gemini 4 Carbon numbers in that table can differ from those labs' own posts.

Source: Artificial Analysis — Claude Fable 5.2 · llm-stats — Fable 5.2 launch · Artificial Analysis — leaderboard

Gemini 4 Hydrogen — Google DeepMind, November 12, 2026

Who / License: Google; proprietary, no public weights.

What's notable: Successor to Gemini 3.8 Flash and the first model in that tier to drop the "Flash" name. 1,048,576-token context, text output, with image, audio and video input; Google Cloud lists it as generally available as a short-term release with a minimum 45-day retirement notice. AA lists $1.50 / $7.50, 188 tokens/s, and an index score of 54 at xhigh (51 at high). Google says it is up to 2.5x faster than Gemini 3.8 Flash and cuts fabricated claims more than fivefold; those are Google's claims, relayed by secondary coverage. Reports on pricing disagree (one says Google kept 3.8 Flash's price), so the $1.50 / $7.50 here is AA's figure.

Source: Artificial Analysis — Gemini 4 Hydrogen · Google Cloud docs — Gemini 4 Hydrogen · Digital Trends

Muse Ember 1.0 — Meta, November 9, 2026

Who / License: Meta; hosted API and Muse Code. Meta hasn't said it will release weights.

What's notable: Meta's smallest and cheapest Muse model: $0.10 / $0.50 per million tokens, 1M context, up to 512K output, text/image/audio/video input. AA scores it 36 at max effort at $0.14 per index task. Meta's own comparison puts that next to Claude Haiku 5.5 at 36.2 (but $0.411 per task), and says Ember trails Haiku on FrontierCode 2.0-AA (27.2 vs 34.9). The comparison is Meta's, as is the claim that it beats both rivals on AutomationBench-AA. Google Cloud lists it as Preview.

Source: TestingCatalog · Google Cloud docs — Muse Ember · Artificial Analysis — leaderboard

Open-weight and unclear

Qwen4-Max-Preview — Alibaba, November 11, 2026

Who / License: Alibaba Cloud / Qwen team. Not open as far as I can tell: it is a time-limited preview behind a Token Plan subscription (Singapore region, interactive coding and agent tools only), ending November 25. Alibaba has not said whether Qwen4 weights will be public.

What's notable: Per Alibaba documentation as reported by DataNorth, 1M context and up to 393,216 output tokens, with text, image and video input. Alibaba has published no benchmark scores, no parameter count and no price. That settles last week's question of whether Qwen 4 had shipped: a preview did, a real release didn't. I'm not repeating the small third-party test scores floating around because I couldn't trace them to a primary source.

Source: DataNorth — Qwen4-Max Preview · Gadget Pilipinas

ERNIE 5.5 Flash Preview — Baidu (reported November 10–11, 2026)

Who / License: Baidu; the preview is a paid API. A separate ERNIE-5.5-Flash-Base checkpoint (reported 280B total, 14B active) was released with open weights on November 5, but that is a pretraining checkpoint without post-training.

What's notable: OpenRouter lists the preview at $0.28 / $1.13 per million tokens with 1M context. I did not find Baidu's own announcement or an AA score, so no benchmark claims here.

Source: OpenRouter — ERNIE 5.5 Flash Preview · LetsDataScience — ERNIE 5.5 Flash Base

Head-to-Head — Current Frontier

AA index, context and (where marked) prices were fetched from AA pages on November 13; the AA leaderboard lists the highest-effort variant of each model, and the table uses that. Prices marked † and the DeepSeek, GLM and Mistral rows are carried from the November 6 edition and were not re-fetched. SWE-bench Pro for Fable 5.2 and Opus 5.5 is Anthropic's reported figure via llm-stats; the Sonnet 5.5 figure is carried from October 30. I dropped the LMArena and GPQA columns this week because I could not pull current figures (arena.ai blocks direct fetches); last week's values are in the November 6 edition. "—" means no sourced figure.

State of play: Fable 5.2 leads AA by five points over a Carbon that outsiders can't call, so for the first time in a while the top score is also a model you can pay for. It isn't cheap to reach: AA's cost per index task is $26.54 at max effort versus $9.08 at high effort (score 62), and $10.04 for Opus 5.5 (59). If you take the high setting, Fable 5.2 is still ahead of Opus 5.5 max for less. Hydrogen is the more interesting price story. At 54 it scores within a point of Sonnet 5.5 and GPT-6.1 Sol at max and costs $1.50 input, though AA's cost-per-task column favors GPT-6.1 Sol ($1.27 at max, $0.59 at high) over Hydrogen ($3.16 at xhigh, $1.87 at high), because Hydrogen spends more tokens. Best open model is still Kimi K3.1 Pro at 46, 19 points behind Fable 5.2 and 13 behind Opus 5.5.

Benchmark & Leaderboard Movement

• New #1 on AA. Fable 5.2 (max) at 65 replaces Gemini 4 Carbon (60). Last week's order was Carbon 60, Opus 5.5 59, Astra 57; Opus 5.5, Astra and Sonnet 5.5 are unchanged at 59, 57 and 55.
• Anthropic now holds three of the top three slots on AA's list (Fable 5.2 at max, xhigh and high), and Opus 5.5 holds the next one down at 59 behind Carbon.
• Fable 5.1's 49 is stale. The AA leaderboard top 25 no longer shows Fable 5.1; I can't say whether it was dropped or just fell below the cutoff. Fable 5.2's price cut and score jump make it the one to compare against.
• Self-reported scoreboard. Nearly every Fable 5.2 benchmark in launch coverage is Anthropic's, run on harnesses Anthropic or its partners picked. Where Anthropic's table and AA overlap in spirit (Opus 5.5 still ahead on AA-Briefcase and GDPval-AA Elo), Fable 5.2 doesn't sweep. Read the 93.0% SWE-bench Pro number as a claim.
• Verbosity is the hidden variable. Fable 5.2 used 275M output tokens to complete the AA index. A model that thinks that long scores well and bills accordingly, so compare on cost per task, not on list price.
• Open weights: no movement. Kimi K3.1 Pro stays the top open score at 46; the Qwen4 preview has no scores and no weights.

Analysis

For hard agentic and long-horizon coding work, Fable 5.2 has the best number available, but run it at high effort first; the jump from 62 to 65 at max costs nearly three times as much per task on AA. For cheap and fast, Hydrogen at $1.50 / $7.50 and 188 tokens/s is the one to benchmark against your own prompts, and GPT-6.1 Sol remains cheaper per index task by AA's measure. Muse Ember is worth a look for routing and extraction at $0.10 / $0.50, with the caveat that its 36 is Haiku-class, not frontier. The open-versus-closed gap on AA widened this week, from 13 to 19 points between Kimi K3.1 Pro and the best model you can call, and it will stay wide until Qwen4 ships weights with scores attached.

Sources

• Artificial Analysis — Leaderboard
• Artificial Analysis — Claude Fable 5.2
• Artificial Analysis — Gemini 4 Hydrogen
• Artificial Analysis — Kimi K3.1 Pro
• llm-stats — Claude Fable 5.2 launch
• llm-stats — AI news
• LLM Gateway — model timeline
• Google Cloud docs — Gemini 4 Hydrogen
• Digital Trends — Gemini 4 Hydrogen
• TestingCatalog — Meta launches Muse Ember
• Google Cloud docs — Muse Ember 1.0
• DataNorth — Qwen4-Max Preview
• Gadget Pilipinas — Qwen4-Max Preview
• OpenRouter — ERNIE 5.5 Flash Preview
• LetsDataScience — Baidu ERNIE 5.5 Flash Base
• My-Library — AI Model & Benchmark Watch, November 6, 2026 edition