Claude Fable 5.2 Is Live: 20% Cheaper and Faster

Key Highlights

Prices went down 20%: input $8 and output $40 per million tokens, down from $10/$50 on Fable 5.1, with $0.20 cache reads and $10 cache writes (5-minute)
Across-the-board gains: FrontierCode 2.0 rises from 32.4% to 49.1%, AutomationBench from 31.4% to 51.4%, and RoboDojo Sim from 27.6% to 64.9%
Faster and cheaper per task: generation is up to 30% faster than 5.1, and at High effort it matches Opus 5.5’s best score for roughly half the cost per task
Same envelope: a 1M-token context window (default and maximum), 128,000 max output tokens, and adaptive thinking that is always on
⚠️ Four breaking changes: the prefix check now applies to every account, thinking blocks are tied to the model that produced them, computer_20251124 returns an error, and advisor tools must be same-generation — read before migrating
Live on APIYI: claude-fable-5-2 and claude-fable-5-2-thinking through both OpenAI-compatible and native Anthropic endpoints, with all four billing items matching official pricing

Background

Anthropic released Claude Fable 5.2 and Claude Mythos 5.2 on November 9, 2026 — less than three months after Fable 5.1 shipped on September 1. Where 5.1 was a same-price upgrade, 5.2 cuts prices and speeds things up: input drops from $10 to $8, output from $50 to $40, and generation is up to 30% faster. For a model positioned for demanding reasoning and long-horizon agentic work, that combination matters more than another leaderboard row — the same budget buys more agent loops. Fable 5.2 and Mythos 5.2 remain the same underlying model with different safeguard profiles: Fable is available to everyone with cybersecurity and biology safeguards, while Mythos is limited to organizations verified through Anthropic’s programs at identical specs and pricing. Anthropic’s own guidance is to start with Opus 5.5 for most coding and knowledge work, and move up to Fable 5.2 only when Opus at higher effort still falls short — or down to Sonnet 5.5 or Haiku 5.5 for simpler work. APIYI now serves claude-fable-5-2 and claude-fable-5-2-thinking, with input, output, cache read, and cache write all matching the official list.

Deep Dive

Pricing: this release cuts it

Billing item	Fable 5.2	Fable 5.1	Change
Input	$8.00 / 1M	$10.00 / 1M	down 20%
Output	$40.00 / 1M	$50.00 / 1M	down 20%
Cache read	$0.20 / 1M	$0.25 / 1M	down 20%
Cache write (5-minute)	$10.00 / 1M	$12.50 / 1M	down 20%

There is also an optional fast mode at $16 input / $80 output, which speeds up token generation by up to 2.5x. Per Anthropic’s terms it is twice the standard rate, aimed at latency-sensitive work that can absorb the premium. APIYI currently bills at the standard rate and does not offer a separate fast variant.

Performance: ahead of the previous generation almost everywhere

Official comparison figures (Fable 5.2 / Fable 5.1 / Opus 5.5, each at its highest available effort with safeguards on):

Benchmark	Fable 5.2	Fable 5.1	Opus 5.5	What it measures
FrontierCode 2.0	49.1%	32.4%	45.5%	Agentic coding (would maintainers merge it)
FrontierSWE v2	69.9%	56.3%	62.3%	Agentic software engineering
AutomationBench	51.4%	31.4%	42.5%	Business workflow automation
Terminal-Bench-Science 0.1	74.6%	52.6%	58.7%	Agentic scientific research
WANDR	82.1%	68.7%	72.3%	Agentic search
OSWorld 2.1	89.3%	83.5%	87.2%	Computer use
RoboDojo Sim	64.9%	27.6%	39.4%	Robotics simulation
AA-Briefcase v1.1	66.1%	61.5%	61.2%	Knowledge work (rubric)
Remote Labor Index	41.3%	17.9%	21.3%	Remote work tasks
Humanity’s Last Exam (with tools)	74.2%	65.6%	67.7%	Multidisciplinary reasoning
Chartography	73.4%	44.8%	64.4%	Visual chart recognition

Among competitors, GPT-6 Astra takes FrontierSWE v2 at 65.5% and Gemini 4 Carbon takes AutomationBench at 53.9%; overall Fable 5.2 leads on the most rows in this table. RoboDojo Sim (27.6% to 64.9%) and Chartography (44.8% to 73.4%) are the largest jumps, pointing at robotics and chart vision as the two areas 5.2 focused on.

Efficiency: the same score at half the cost

Anthropic’s efficiency claims matter more than the raw scores:

At FrontierCode 2.0 High effort, Fable 5.2 scores 45.3%, essentially level with Opus 5.5’s best of 45.5%, at roughly half the cost per task
At Medium effort (the default in Claude apps and Claude Code), it finishes comparably hard tasks with fewer tokens than 5.1
At Xhigh effort, it reaches 49.1% — the highest score of any released model Anthropic has tested

For agent workloads, the cost advantage is not just the 20% price cut; it is also fewer tokens per task.

Specs

Spec	Value
Model IDs	`claude-fable-5-2` / `claude-fable-5-2-thinking`
Context window	1,000,000 tokens (default and maximum, single price across the window)
Max output	128,000 tokens
Thinking	Adaptive thinking always on; depth controlled by the `effort` parameter
Effort levels	`low` / `medium` / `high` / `xhigh` / `max`
Tokenizer	Same as the one introduced with Opus 4.7
API formats	OpenAI-compatible `chat/completions` and native Anthropic `messages`
APIYI groups	`ClaudeCode`, `Claude_Enterprise`, `Default`, `SVIP`

Compared with 5.1, claude-fable-5-2 is not yet available in the Claude_Reverse group on APIYI; the other groups are unchanged.

⚠️ Four Breaking Changes

If you already call claude-fable-5-1, check these four items before switching the model ID. The first two return 400 errors outright.

1. The prefix check now applies to every account

The API’s prefix check verifies that nothing before a Fable 5.2 thinking block has changed since it was produced: the system prompt, the tools array, or any earlier message. On Fable 5.1 this check only enforced for accounts created on or after August 31, 2026; older accounts logged without blocking. On Fable 5.2 it applies to every account, so a request carrying a stale thinking block without prefix_mismatch_behavior set returns:

The block is bound to a different conversation

As a stopgap, send the thinking-binding-controls-2026-08-01 beta header and set thinking.block_binding.prefix_mismatch_behavior to "drop_block" so the API drops the affected blocks and continues; dropped blocks appear in the top-level input_transformations array. Treat this as temporary — dropping blocks raises output tokens — and fix the append-only discipline instead.

2. Thinking blocks are tied to the model that produced them

Each thinking block records the model that produced it, and only certain models can read it. Fable 5.2 and Mythos 5.2 read each other’s blocks, plus those from Opus 5.5, Sonnet 5.5, Haiku 5.5, Fable 5.1, Mythos 5.1, Opus 5, Fable 5, Mythos 5, and earlier Opus / Sonnet / Haiku models. The reverse does not hold: moving from Fable 5.2 to any other model — including Opus 5.5 and Fable 5.1 — loses their reasoning for those turns. When a request carries a block the target model cannot read, the API silently drops it before the model sees it, and dropped blocks are not billed; the beta header surfaces the drops in input_transformations.

3. `computer_20251124` is no longer available

The older computer use tool computer_20251124 returns an error on every platform except Amazon Bedrock. Migrating to 5.2 means moving to the new computer use toolset.

4. Advisor tools must be same-generation

With the advisor tool, a Fable 5.2 executor needs a Fable 5.2 or Mythos 5.2 advisor; earlier advisor models cannot be paired with it.

Practical Use

Code example

Call claude-fable-5-2 through the OpenAI-compatible format. Set APIYI_API_KEY in your environment first:

import os
from openai import OpenAI

client = OpenAI(
    api_key=os.environ["APIYI_API_KEY"],
    base_url="https://api.apiyi.com/v1"
)

response = client.chat.completions.create(
    model="claude-fable-5-2",
    messages=[
        {"role": "user", "content": "Extract the three duplicated auth paths in this repository into one module, list the files you change, and explain how to regression-test it."}
    ]
)

print(response.choices[0].message.content)

The native Anthropic messages format works too, with the same model field. For long-running agent work, stream the response and check stop_reason and intermediate results at each step.

Where it fits

Long-horizon agentic coding

Leads both FrontierCode 2.0 and FrontierSWE v2 — a fit for multi-file migrations, large refactors, and coding tasks that run for hours

Business workflow automation

AutomationBench at 51.4% suits multi-step, cross-system automation and data pipelines

Deep research and search

WANDR at 82.1% fits multi-round web research, follow-up questions, and synthesis

Computer use

OSWorld 2.1 at 89.3% suits browser and desktop automation; note that the old computer use tool must be replaced

Usage notes

Test the prefix check first: keep conversations append-only and never edit, reorder, or delete earlier turns. Use drop_block as a stopgap, not a long-term fix.
Set effort per step: use high or xhigh for hard steps and drop to medium for routine ones to cut cost without hurting quality.
Do not reuse thinking blocks across generations: moving from 5.2 back to 5.1 or Opus 5.5 loses the earlier reasoning, so rebuild the context.
Use fast mode deliberately: for latency-sensitive work that can absorb 2x pricing, Anthropic’s fast mode is an option; APIYI currently bills at the standard rate.

Pricing and Availability

Pricing

Billing item	Price per 1M tokens
Input	$8.00
Output	$40.00
Cache read	$0.20
Cache write (5-minute)	$10.00

Pricing is aligned with the official list and may change; the table is for reference only. The Model Pricing tab in the top navigation is authoritative: Model Pricing.

Available groups

All four groups — ClaudeCode, Claude_Enterprise, Default, and SVIP — serve the model, and claude-fable-5-2 and claude-fable-5-2-thinking behave the same across both endpoints. Claude_Enterprise is the more stable Claude official-passthrough group, built for enterprise support and emergency needs: pure official passthrough with no AWS Claude (not include Amazon Claude), at a 1.2x group multiplier. ClaudeCode is 0.95x. Default and SVIP also have the model available. The Claude_Reverse group is not available yet.

Stacking recharge promotions

The prices above are list rates and can be combined with recharge bonuses and group discounts. Actual charges follow the console. See recharge promotions for the rules.

Summary and Recommendations

Fable 5.2 is a rare release that cuts price, speeds up generation, and improves capability at the same time. For teams already running agent workloads, the question is not whether to move but when — a 20% lower unit price, fewer tokens per task, and roughly half the cost per task versus Opus 5.5 compound into a larger gap than the sticker price suggests.

Hard agentic coding and migrations: leads FrontierCode 2.0 and FrontierSWE v2, so replace long-running tasks first
Budget-sensitive long-horizon work: matches Opus 5.5 at High effort for half the cost, a natural upgrade when Opus falls short
Everyday coding and knowledge work: follow Anthropic’s guidance and stay on Opus 5.5 rather than paying Fable prices for simple tasks
Before migrating: work through the prefix check and thinking-block binding first; those are the two most likely traps

Sources and verification date:

Anthropic announcement: anthropic.com/claude-fable-and-mythos-5-2 (November 9, 2026)
Model overview: platform.claude.com/docs/en/models/fable-5-2/overview
What’s new: platform.claude.com/docs/en/models/fable-5-2/whats-new-fable-5-2
Pricing: platform.claude.com/docs/en/about-claude/pricing
APIYI live model API: api.apiyi.com/api/pricing
Research and pricing verified on November 10, 2026

Changelog

AI Radar

Claude Fable 5.2 Is Live: 20% Cheaper and Faster

Key Highlights

Background

Deep Dive

Pricing: this release cuts it

Performance: ahead of the previous generation almost everywhere

Efficiency: the same score at half the cost

Specs

⚠️ Four Breaking Changes

1. The prefix check now applies to every account

2. Thinking blocks are tied to the model that produced them

3. `computer_20251124` is no longer available

4. Advisor tools must be same-generation

Practical Use

Code example

Where it fits

Long-horizon agentic coding

Business workflow automation

Deep research and search

Computer use

Usage notes

Pricing and Availability

Pricing

Available groups

Stacking recharge promotions

Summary and Recommendations

​Key Highlights

​Background

​Deep Dive

​Pricing: this release cuts it

​Performance: ahead of the previous generation almost everywhere

​Efficiency: the same score at half the cost

​Specs

​⚠️ Four Breaking Changes

​1. The prefix check now applies to every account

​2. Thinking blocks are tied to the model that produced them

​3. computer_20251124 is no longer available

​4. Advisor tools must be same-generation

​Practical Use

​Code example

​Where it fits

Long-horizon agentic coding

Business workflow automation

Deep research and search

Computer use

​Usage notes

​Pricing and Availability

​Pricing

​Available groups

​Stacking recharge promotions

​Summary and Recommendations

Key Highlights

Background

Deep Dive

Pricing: this release cuts it

Performance: ahead of the previous generation almost everywhere

Efficiency: the same score at half the cost

Specs

⚠️ Four Breaking Changes

1. The prefix check now applies to every account

2. Thinking blocks are tied to the model that produced them

3. `computer_20251124` is no longer available

4. Advisor tools must be same-generation

Practical Use

Code example

Where it fits

Usage notes

Pricing and Availability

Pricing

Available groups

Stacking recharge promotions

Summary and Recommendations