Seven of November's 22 models matter: the leader is 34 points stronger but costs 80 times more
Twenty-two models from 13 labs shipped in the first 17 days of November, and the author of an X thread picked out seven worth knowing. On Artificial Analysis's composite index, Claude Fable 5.2 leads with 65 points, while open-weight Inkling 2 Small sits at the bottom of the list with 31, a 34-point gap. Prices differ by 80 times: $8 versus $0.10 per million input tokens for Meta's Muse Ember. That makes it sensible to send simple requests to the cheap model and reserve the expensive one for hard tasks.
AI-processed from X: @replay_0x; edited by Hamidun News
Choosing a model for a product has become harder than writing code for it. According to the count of the author of an X post (@replay_0x), 22 models from 13 labs shipped in the first 17 days of the month, and OpenAI released five of them. Tracking every release is pointless. The author picked seven that they consider significant and broke them down by quality and price. One caveat: Claude compiled the list, and the source is a single social media post, so check the prices against official price sheets.
How to read the scores and prices
The scores come from the Artificial Analysis index. That is an independent service that runs models through a battery of tests and condenses the results into one number. It is a convenient yardstick, but an averaged one: it knows nothing about your specific task.
Prices are written as "$8 / $40". The first number is the cost of a million input tokens (a token is roughly a fragment of a word from your prompt), and the second is the cost of a million output tokens, meaning the model's reply. Replies cost more because generating text is harder than reading it.
The seven models
Claude Fable 5.2 — 65 points. First place on the index at maximum "effort", the mode in which the model thinks for longer. Priced at $8 / $40, which is 20% below Fable 5.1 (by this arithmetic, the previous version cost roughly $10 / $50).
GPT-6.2 Astra — 62. OpenAI's answer, released 8 days after the leader. Same price, $8 / $40, trailing by 3 points.
GPT-6.2 Sol — 57. The author calls it nearly frontier (that is, close to the strongest models) at $2 / $10. That is four times cheaper than the two top models, at a cost of 8 points relative to the leader.
Gemini 4 Hydrogen — 51. Google's new mid-tier model at $1.50 / $7.50.
Qwen4-Max — 43. Alibaba's new flagship at $1.50 / $4.50.
Muse Ember — 36. Meta's small model at $0.10 / $0.50. It suits high-volume simple operations: labeling, classification, short answers.
Inkling 2 Small — 31. Open weights with 594 billion parameters under the Apache 2.0 license. You don't pay for API access, but you have to run it on your own hardware, and that hardware costs money too.
How to choose in five minutes
The author proposes four scenarios:
- The hardest work: agents, large codebases. Take Fable 5.2 or GPT-6.2 Astra, where quality matters more than the per-token price.
- A workhorse on a limited budget. GPT-6.2 Sol or Gemini 4 Hydrogen will handle most internal tools and chatbots.
- Lots of cheap requests. Muse Ember: at $0.10 per million input tokens, you can process enormous streams of text.
- Your own stack. Inkling 2 Small: data never leaves for an external API, and Apache 2.0 permits commercial use.
What to do with this
The key pair of numbers: from first place to seventh is 34 points, while on price the gap is roughly 80 times ($8 versus $0.10 per million input tokens). In other words, the last few quality points at the top cost a lot. That is why routing makes sense. Say you run a support service. Muse Ember can identify the topic of a ticket and sort emails. A reply to a complicated complaint that requires reading a long correspondence is better handed to Astra or Fable. That noticeably cuts the token bill, and quality suffers only where it isn't critical.
There are limitations too. The index is an average: a model with 51 points can beat the leader on a narrow task, and on long agentic scenarios the picture can differ. Maximum "effort" mode also burns more output tokens, so the real bill may come in higher than the price sheet suggests. Finally, all prices and scores come from one post, not from primary sources.
The practical takeaway is simple: don't chase every release. Pick 2–3 candidates in your price tier, run them on 50–100 real requests from your product, and compare quality and total cost. Choose a model based on your own data, not its place in a ranking.
- Nov 21, 2026Google Adds a Chat History Sidebar to Search's Homepage
- Nov 21, 2026METR: The Length of Tasks an AI Agent Completes with 50% Reliability Doubles Every 3.6 Months
- Nov 21, 2026Google moves AI Mode chat history onto its worldwide search homepage
- Nov 21, 2026Калифорния подала в суд на OpenAI: агенты вышли из-под контроля 80 раз, публично — 27
Key Questions Answered
How to read the scores and prices?
The scores come from the Artificial Analysis index. That is an independent service that runs models through a battery of tests and condenses the results into one number. It is a convenient yardstick, but an averaged one: it knows nothing about your specific task. Prices are written as "$8 / $40". The first number is the cost of a million input tokens (a token is roughly a fragment of a word from your prompt), and the second is the cost of a million output tokens, meaning the model's reply.
What are the key details on The seven models?
Claude Fable 5.2 — 65 points. First place on the index at maximum "effort", the mode in which the model thinks for longer. Priced at $8 / $40, which is 20% below Fable 5.1 (by this arithmetic, the previous version cost roughly $10 / $50). GPT-6.2 Astra — 62. OpenAI's answer, released 8 days after the leader. Same price, $8 / $40, trailing by 3 points. GPT-6.2 Sol — 57. The author calls it nearly frontier (that is, close to the strongest models) at $2 / $10.
How to choose in five minutes?
The author proposes four scenarios: The hardest work: agents, large codebases. Take Fable 5.2 or GPT-6.2 Astra, where quality matters more than the per-token price. A workhorse on a limited budget. GPT-6.2 Sol or Gemini 4 Hydrogen will handle most internal tools and chatbots. Lots of cheap requests. Muse Ember: at $0.10 per million input tokens, you can process enormous streams of text. Your own stack. Inkling 2 Small: data never leaves for an external API, and Apache 2.0 permits commercial use.
Test Frontier LLMs & Autonomous Agents in One Workspace
Run Claude Opus 5.5, GPT-6, Gemini 3.1, and DeepSeek side-by-side in Nomad AI (nomadnet.ai) — unified model routing, deep research, and multimodal workflows under one subscription.
The AI world, distilled — once a week
Seven stories that actually mattered, hand-picked. No noise, no reposts, no press releases.
Done! Check your inbox for a confirmation.