LLM Metrics Explained: Claude Fable 5.2 Benchmarks for Beginners

LLM metrics are the numbers that tell you what a large language model can do, how fast it does it and what it costs. When Anthropic released Claude Fable 5.2 on 9 November 2026, the launch came with more than a dozen benchmark scores, two price tiers and a set of safety figures. This guide explains those LLM metrics in plain language, so a trader who wants to use AI for research, coding or automation can read any model launch and decide what actually matters.

All Fable 5.2 numbers below are taken from Anthropic’s Fable 5.2 announcement. They are vendor-reported, which is the first thing to keep in mind when you read any benchmark table.

Key takeaways

• LLM metrics fall into six groups: capability, agentic ability, cost, speed, effort and safety. A single headline score never tells the full story.
• Claude Fable 5.2 leads its own comparison table on most benchmarks, but Gemini 4 Carbon scores higher on AutomationBench and Opus 5.5 has a slightly higher AA-Briefcase Elo.
• Fable 5.2 costs $8 per million input tokens and $40 per million output tokens, with a fast mode at double the price.
• Always check the effort level, whether safeguards were on, and who ran the test before comparing two scores.
• For traders, the useful questions are cost per task, coding and tool-use scores, and chart reading, not the biggest number on the page.

What are LLM metrics?

An LLM metric is a measurement of a model on a defined test. Some tests are like exams with right and wrong answers. Others drop the model into a real task, such as fixing a bug in a code repository or completing a spreadsheet workflow, and check whether the job got done. Price and speed are metrics too, because a model that is 2% better but costs three times as much may be the wrong choice for a daily scanner or a backtesting agent.

AI FOR TRADERS

AI Bootcamp 2.0 for traders and investors

Use ChatGPT, Claude and AI agents for research, coding and automation in your trading workflow.

• Prompting and agent workflows for markets
• Build tools without being a full-time coder
• 6 courses in one bootcamp

See the bootcamp →by Marketcalls

1. Capability LLM metrics

These measure whether the model can solve a problem at all: maths, science questions, code, reading a chart. The score is usually the percentage of tasks solved. Humanity’s Last Exam and Chartography are capability benchmarks.

2. Agentic LLM metrics

An agent is a model that plans, uses tools, runs code and keeps going until a multi-step job is finished. Agentic benchmarks such as FrontierSWE, Terminal-Bench and OSWorld score the end result of that whole loop. These matter most if you want an AI to build a strategy, run a backtest and fix its own errors.

3. Cost LLM metrics

Models are billed per token, a chunk of text roughly three quarters of a word. Prices are quoted per million input tokens (what you send) and per million output tokens (what the model writes). Output is almost always more expensive. The more useful number is cost per task, which combines price with how many tokens the model needs to finish.

4. Speed LLM metrics

Two speed metrics show up in most comparisons. Time to first token (TTFT) is how long you wait before the answer starts appearing. Tokens per second is how fast the rest of the answer streams out. For a live market dashboard both matter; for an overnight research job, neither matters much.

5. Effort settings in LLM metrics

Newer models let you choose how long they think before answering, with levels such as Medium, High, Xhigh and Max. Higher effort usually means a better score and a bigger bill. Two scores for the same model can differ by several points only because the effort level changed.

6. Safety and alignment LLM metrics

These measure how often a model refuses harmful requests, how often it blocks harmless ones, and how it behaves in adversarial tests. They do not tell you how smart the model is, but they tell you how predictable it is when it runs unattended.

Claude Fable 5.2 benchmark scores at a glance

Fable 5.2 and Mythos 5.2 are the same underlying model. Fable has extra safeguards for cybersecurity and biology and is available to everyone; Mythos has fewer safeguards and is limited to vetted organisations. Here are the main reported scores against Fable 5.1, Opus 5.5 and two rival models.

Fable 5.1 scored lower on every row, for example 56.3% on FrontierSWE v2 and 17.9% on the Remote Labor Index, so the jump from 5.1 to 5.2 is large on agentic work.

Fable 5.2 LLM metrics explained for beginners

FrontierCode and FrontierSWE: can it write and fix code?

These are the benchmarks to watch if you use AI to write Python, AFL or Pine Script. FrontierSWE gives the model real software issues and checks whether its fix passes the project’s tests. Fable 5.2 scores 69.9% against 62.3% for Opus 5.5. FrontierCode is harder; Anthropic also reports that at High effort Fable 5.2 scores 45.3%, close to Opus 5.5’s 45.5%, at roughly half the cost per task. That last detail is a cost metric hiding inside a capability chart.

Terminal-Bench and OSWorld: can it use a computer?

Terminal-Bench checks work done through the command line, the same way you would install packages, run a backtest script or read logs. OSWorld checks whether the model can operate a desktop: click, type, open files. A high OSWorld score (89.3% for Fable 5.2) matters if you want an agent to drive a broker terminal or download reports for you.

AutomationBench and Remote Labor Index: can it do a real job?

AutomationBench scores business workflows across apps, and here Gemini 4 Carbon leads with 53.9% against 51.4%. The Remote Labor Index uses real freelance tasks that people were paid for. Fable 5.2 more than doubles Opus 5.5 here (41.3% vs 21.3%), but note it was run at High effort.

Humanity’s Last Exam and Chartography: can it reason and read charts?

Humanity’s Last Exam is a set of very hard expert questions; “with tools” means the model could search or run code. Chartography tests reading charts and figures without tools. For traders this is the closest benchmark to “look at this price chart and tell me what you see”. Fable 5.2 scores 73.4%, narrowly ahead of Gemini 4 Carbon (72.5%) and GPT-6 Astra (71.0%).

Elo ratings: how often people prefer it

An Elo score comes from head-to-head comparisons, the same system used in chess. It is relative, so 1785 only means something next to another model’s Elo. On AA-Briefcase, Opus 5.5 (1800) edges Fable 5.2 (1785) even though Fable 5.2 has the higher rubric score. Different metrics on the same test can disagree.

Safety metrics: how it behaves unattended

On Anthropic’s automated behavioural audit, where lower is better, Mythos 5.2 scores 2.67 against 2.79 for Opus 5.5. In an adversarial sandbox test it used outside credentials or files in 3.1% of scenarios, down from 21% for Mythos 5.1. Fable 5.2 also blocks about 55% fewer legitimate requests than Fable 5.1. For a trading agent that runs on its own, fewer false refusals and fewer rule breaks both count.

Fable 5.2 price and what a task really costs

Worked example: you ask Fable 5.2 to review a strategy repository. You send 200,000 tokens of code and instructions and it writes 20,000 tokens of analysis and fixes. In standard mode that is 0.2 x $8 + 0.02 x $40 = $1.60 + $0.80 = $2.40. In fast mode the same job costs $4.80. Run it 20 times while iterating on a strategy and the gap becomes $48 against $96. Cost per task, not price per token, is the number to compare across models.

How to read LLM metrics and benchmark scores

• What task? A coding score says little about chart reading. Match the benchmark to your use case.
• What effort level? Fable 5.2’s 49.1% on FrontierCode was at Xhigh effort; at High it was 45.3%.
• Safeguards on or off? Several cyber and biology rows are Mythos scores, or Opus 5.5 without safeguards.
• Who ran the test? Vendor-reported numbers are a starting point. Some rows, such as RoboDojo Sim, were run internally, and CyScenarioBench uses different tasks for each lab.
• What does it cost per task? A small gain at double the cost may not be worth it for a daily job.

LLM leaderboards: Arena and Artificial Analysis

Vendor tables are one view. Two independent leaderboards are worth bookmarking because they rank many models side by side with the same method.

Arena leaderboard: what users prefer

The Arena leaderboard ranks models from blind head-to-head votes: a user sends one prompt, sees two anonymous answers and picks the better one. Thousands of these votes are turned into a rating, similar to the Elo idea above, with separate boards for text, vision, image and other arenas. It is the best quick check of how a model feels to use. It does not measure cost or speed, and a new model needs time to collect enough votes, so very recent releases can be missing or move around in their first weeks.

Artificial Analysis: intelligence, cost and time per task

Artificial Analysis runs its own tests and puts quality, price and speed on one page. Its Intelligence Index v5.0.1 combines nine evaluations, including AA-Briefcase v1.1, AutomationBench-AA, Terminal-Bench 4.0, FrontierCode 2.0, HLE-Diamond and Terminal-Bench-Science 0.1. Alongside it, the site reports time per task and cost per task, and compares API providers on speed, latency and price.

Artificial Analysis added Claude Fable 5.2 on 9 November in several effort settings. In its domain capability index, Fable 5.2 at Max effort ranks first in all six domains it shows:

Use the two together: Arena tells you which answers people like, Artificial Analysis tells you what that quality costs and how long it takes. Leaderboards update often, so check the live pages before you choose a model.

Which LLM metrics matter for traders?

If you build tools with AI, a few metrics carry most of the weight:

• Coding and agentic scores (FrontierSWE, Terminal-Bench) if you generate or fix strategy code, scanners and dashboards.
• Chart reading (Chartography) if you want commentary on charts or screenshots.
• Cost per task for anything that runs every day, such as end-of-day reports or alert summaries.
• Speed (time to first token, fast mode) for intraday tools where a reply has to arrive in seconds.
• Reliability and safety for agents that place orders. Even with a strong model, keep a human approval step and test in sandbox mode first.

If you are new to using AI for strategy work, our guide on prompting Claude to build trading strategies covers the workflow, and the LLM trading agent tutorial in Python shows how an agent is wired together. For a cheaper open model comparison, see how Gemma 4 narrows the gap.

Which Claude model should you use?

Anthropic still describes Claude Opus 5.5 as its strongest daily driver for most coding and knowledge work. Fable 5.2 is the step up for harder, longer jobs, and Sonnet 5.5 or Haiku 5.5 are the cheaper options for simple tasks. Fable 5.2 can also act as an advisor or orchestrator that hands simpler sub-tasks to smaller models, which keeps cost per task down. On the API the model ID is claude-fable-5-2, and it is available on AWS, Google Cloud and Microsoft Azure.

Frequently Asked Questions

What are LLM metrics?

LLM metrics are measurements of a language model on defined tests: capability (can it solve the task), agentic ability (can it finish multi-step work), cost, speed, effort settings and safety behaviour.

What is the Claude Fable 5.2 benchmark score on coding?

Anthropic reports 69.9% on FrontierSWE v2 and 49.1% on FrontierCode 2.0 at Xhigh effort, both ahead of Opus 5.5 (62.3% and 45.5%).

Is Claude Fable 5.2 better than Opus 5.5?

On most reported benchmarks, yes. Opus 5.5 still has a slightly higher AA-Briefcase Elo (1800 vs 1785) and remains Anthropic’s recommended daily model for most tasks.

What is the Claude Fable 5.2 price?

$8 per million input tokens and $40 per million output tokens. Fast mode, up to 2.5x faster, costs $16 and $80.

What is time to first token in LLM metrics?

Time to first token (TTFT) is the delay between sending a request and receiving the first word of the reply. It matters most for interactive tools where you are waiting on the answer.

Are LLM benchmark scores reliable?

They are useful but incomplete. Check the effort level, safeguard settings and who ran the test, and run a few of your own tasks before switching models.

MEMBERSHIP

Python Elite membership

15 Python, AI and algo trading programs in one plan, including QuantFlow and AI Bootcamp live mentorship.

• Python from fundamentals to trading apps
• Backtesting, broker APIs and OpenAlgo
• Live QuantFlow and AI Bootcamp sessions

View Python Elite →by Marketcalls