Skip to main content
← All posts

OpenCheese AI ·

Claude Fable 5.2 vs Opus 5.5, twice the price per token and sometimes half the price per task

Claude Fable 5.2 costs $8 and $40 per million tokens, double Opus 5.5. On some benchmarks it reaches Opus's best score at a lower effort for half the money, on others it costs 2.6 times as much per task. Every number is checked against Anthropic's docs, the system card and independent leaderboards.

On November 9, Anthropic released a model numbered 5.2 that beats the model it numbered 5.5 seven weeks earlier. Claude Fable 5.2 (claude-fable-5-2) took first place on the Artificial Analysis Intelligence Index with 65 points. Claude Opus 5.5 (claude-opus-5-5) scores 59 there and sits fifth.

Anthropic's own advice is to keep using the one in fifth place. Its launch post tells users to "start with Claude Opus 5.5, our strongest daily driver" for most coding and knowledge work, and its docs say to move to Fable 5.2 "when your evals on a Claude Opus model at higher effort still don't meet your requirements." Recommending the cheaper model on the new flagship's launch day is unusual. The benchmarks explain it.

Every figure below was checked on November 10 against Anthropic's announcements, API docs and the Fable 5.2 system card, and against the leaderboards of Artificial Analysis, ARC Prize, Vals AI and Cursor.

Two model lines with their own numbering

Opus, Sonnet and Haiku are Anthropic's mainstream line. Fable and Mythos sit above it, and each line counts on its own.

Date (2026) Release What it brought
June 9 Claude Fable 5 and Mythos 5 Mythos 5 as a restricted preview through Project Glasswing, Fable 5 as the same model with safeguards for general use
September 1 Claude Fable 5.1 and Mythos 5.1 $10 and $50 per million input and output tokens
September 22 Claude Opus 5.5 The first 5.5 model, which Anthropic said "performs at the level of Claude Fable 5.1 on most work." OpenAI shipped GPT-6 Sol and Luna the same day
November 9 Claude Fable 5.2 and Mythos 5.2 $8 and $40, output up to 30% faster than 5.1, first place on Artificial Analysis

So Opus 5.5 was pitched as the September Fable at Opus prices, and seven weeks later Fable moved again. The number after the name tells you which line a model belongs to. It says nothing about rank.

Mythos 5.2 (claude-mythos-5-2) has the same weights as Fable 5.2 with lighter cyber and biology safeguards, for organizations in Anthropic's Cyber Verification and Life Sciences Verification programs and for Claude Security on Enterprise plans. Fable keeps the safeguards but trips them less often. In Anthropic's early testing across claude.ai, Claude Code and Cowork, it blocked about 55% fewer requests than Fable 5.1, with 60% fewer blocks in cybersecurity and 40% fewer in biology.

The week before launch, a Hacker News post titled "Anthropic's Mythos-class model allegedly leaked on Hugging Face" collected 46 points. Its link now reads "URL deleted for malware concerns," which may be the most reliable detail of the whole leak.

The spec sheets match, apart from price and defaults

Claude Opus 5.5 Claude Fable 5.2 GPT-6.1 Sol, for reference
Context / max output 1M / 128K tokens 1M / 128K tokens 1,050,000 / 128,000 tokens
Knowledge cutoff June 2026 June 2026 April 30, 2026
Input / output $4 / $20 $8 / $40 $2 / $10
Cache reads $0.20 $0.20 $0.10
Cache writes, 5 minutes / 1 hour $5 / $8 $10 / $16 $2.50
Batch API $2 / $10 $4 / $20 $1 / $5
Fast mode $8 / $40 $16 / $80 $4 / $20
Prompts over 272K tokens Same rates Same rates 2x input and cache, 1.5x output
US-only or regional processing 1.1x 1.1x 1.1x
Default effort on the API medium high medium

Prices in USD per million tokens.

Fable costs exactly twice as much per token. That is after a 20% cut from Fable 5.1's $10 and $50. A YouTube Short posted on launch day was titled "Claude Fable 5.2 just dropped: faster than Opus 5.5 at half the cost." Per token, the ratio belongs to the other model.

Cache reads cost the same. Both charge $0.20 per million, and Anthropic says cache reads make up the majority of agentic and coding costs. For agents that reread a large context every turn, the real gap is narrower than the headline.

The default effort depends on where you call from. The API runs Fable 5.2 at high and Opus 5.5 at medium unless told otherwise. In Claude apps and Claude Code, Fable defaults to medium too. Thinking is always on for both and is billed as output.

Neither charges extra for long prompts. In Anthropic's words, "A 900k-token request is billed at the same per-token rate as a 9k-token request." GPT-6.1 Sol bills a prompt over 272K input tokens at its higher rates for the whole request.

Fable needs data retention. Fable 5.2 requires 30-day retention and isn't available under zero data retention unless Anthropic authorizes it, while Opus 5.5 is. For some compliance teams that row settles the question before any benchmark does.

Fable wins most benchmarks, by very different margins

The system card says Fable 5.2 scored highest on 14 of the 16 results in its summary table. The two it lost are Elo ratings, and Opus 5.5 won both. Sorting the rest by margin tells you more than counting wins.

Anthropic reports each model at its highest available effort unless noted, with safeguards on. When a classifier fired, a fallback model answered, which on Terminal-Bench 4.0 touched 10% of trials for both models.

Where the lead is small or goes to Opus

Benchmark What it tests Opus 5.5 Fable 5.2 Fable lead
GDPval-AA v2.1 Professional work in 44 occupations, Elo 1866 1831 -35
AA-Briefcase v1.1 Knowledge work, Elo and rubric 1800 and 61.2% 1785 and 66.1% -15 Elo, +4.9 rubric
Terminal-Bench 4.0, Vals AI Terminal tasks, one shared harness 65.15% 65.15% 0
ARC-AGI-1 and ARC-AGI-2 Abstract reasoning, verified 98.5% and 93.3% 98.5% and 93.8% 0 and +0.5
Terminal-Bench 4.0, Anthropic Terminal tasks 66.4% (xhigh) 67.6% (max) +1.2
OSWorld 2.1, offline set Computer use, strict and partial 53.2% and 87.2% 54.6% and 89.3% +1.4 and +2.1
CursorBench 4.0 Real tasks from Cursor sessions 57.8% 60.0% +2.2
SWE-bench Multilingual Fixes across nine languages 93.9% 96.6% +2.7
SWE-bench Pro Large multi-file fixes 89.9% 93.0% +3.1
FrontierCode 2.0 Patches good enough to merge 45.5% 49.1% (xhigh) +3.6

Where the lead is large

Benchmark What it tests Opus 5.5 Fable 5.2 Fable lead
Humanity's Last Exam Expert questions, with and without tools 67.7% and 64.4% 74.2% and 73.2% +6.5 and +8.8
FrontierSWE v2 Engineering tasks with a 20-hour budget 62.3% 69.9% +7.6
AutomationBench Business workflows across apps 42.5% 51.4% +8.9
Chartography, no tools Reading dense charts 64.4% 73.4% +9.0
WANDR, relaxed score Large data-collection jobs 72.3% 82.1% +9.8
SWE-bench Multimodal Fixes from screenshots and mockups 61.4% 74.1% +12.7
Terminal-Bench-Science 0.1 Scientific research workflows 58.7% 74.6% +15.9
BenchCAD 2.1, with tools Hardware design, scored out of 100 65.3 83.5 +18.2
Remote Labor Index Freelance projects done as well as a professional 21.3% 41.3% +20.0
RoboDojo Sim Controlling a simulated two-arm robot 39.4% 64.9% +25.5

The pattern is clear enough to plan around. Where Opus 5.5 was already strong, in terminal work, everyday coding and computer use, Fable adds one to four points. The gap opens on science, hardware, robots, coding from images and whole freelance projects, where Fable delivered 99 of 240 projects at a professional's standard.

Two footnotes keep this honest. Anthropic's 1.2-point Terminal-Bench lead is smaller than its own standard errors (1.8 points for Fable and 2.6 for Opus), and Vals AI's run on a shared harness calls it a tie. And GPT-6 Astra still beats both Claude models on ARC-AGI-2, at 95.0%, and on ARC-AGI-3 with the Provider Adapter harness, at up to 99.9%. Fable's 81.2% on the Standard harness against Astra's 62.7% is the cleaner win. Opus 5.5 has no verified ARC-AGI-3 score.

A thousand agents and Fermat's Last Theorem

The most striking comparison sits in the system card's section on agent teams. Anthropic gave 1,000 agents the job of proving Fermat's Last Theorem in Lean, starting from definitions and lemmas of an earlier formalization. After 18 hours, a team running an earlier snapshot of Fable 5.2 had 0.86 of the weighted milestones and the Opus 5.5 team had 0.26. Given 12 more hours, the Fable team finished a proof the grader accepted. Opus reached 0.27.

What a finished task costs

The price list tells you what each model charges per token. Five public tests also publish what a task costs, and there the answer moves with the effort setting.

Test, run by Opus 5.5 Fable 5.2
FrontierCode 2.0, Cognition 45.5% for $12.82 at max 45.3% for $6.72 at high, 49.1% for $14.19 at xhigh, 45.7% for $44.70 at max
CursorBench 4.0, Cursor 56.0% for $3.97 at high, 57.8% for $13.43 at max 54.8% for $4.66 at medium, 56.9% for $9.70 at high, 60.0% for $32.79 at max
Terminal-Bench 4.0, Vals AI 65.15% for $12.04 65.15% for $14.56
Vals Index, Vals AI 70.14% for $24.77 71.31% for $17.94
Intelligence Index, Artificial Analysis, max effort 59 points for $10.04 65 points for $26.54

Average cost per task or test in USD.

On FrontierCode, Fable at high matches Opus at max for about half the money. It gets there with fewer tokens. Turning Fable all the way up made it worse and dearer, 45.7% for $44.70, or 44.4% if the 22 attempts that timed out are counted as zero.

On CursorBench, Opus is the cheap road to the mid-50s. Opus at high scores 56.0% for $3.97. Fable needs high and $9.70 to beat that by 0.9 points, and $32.79 at max for its 60.0%.

At max effort on Artificial Analysis, Fable costs 2.6 times as much per task and takes 46.4 minutes against 25.2. Opus also generates faster there, 115 tokens per second against 98, so per token the YouTube title is wrong on both counts. Per task it is half right. Vals AI timed Fable faster on both of its suites, and on the Vals Index Fable cost 28% less per test, though not half.

Compare models at a target score rather than at a setting. The cheapest way to 45% on FrontierCode is Fable at high, and the cheapest way to 56% on CursorBench is Opus at high. Neither follows from the price list, so pin the effort in every request:

{
  "model": "claude-fable-5-2",
  "max_tokens": 64000,
  "output_config": { "effort": "medium" },
  "messages": [{ "role": "user", "content": "..." }]
}

Leave room in max_tokens, because thinking counts against it. Two habits from Anthropic's Fable docs are worth knowing before you lower the setting. A prompt line asking for less thinking doesn't stop Fable from thinking at medium and above, and at low and medium it more often ends a long task early with a list of what remains.

Fast mode depends on your cloud

Fast mode promises up to 2.5 times faster output on both models, and it isn't sold everywhere.

Where Opus 5.5 fast Fable 5.2 fast
Price, input / output $8 / $40 $16 / $80
Claude API Yes, speed: "fast", no beta header Yes
Claude Code Yes Yes
Google Cloud, Anthropic managed Yes Yes
Google Cloud, Google managed Not listed Yes, at Google's prices
Microsoft Foundry Deployments hosted on Azure Deployments hosted on Anthropic, as claude-fable-5-2-fast
Amazon Bedrock Not listed Not listed
Batch API No No

Data residency stacks on top. Opus 5.5 fast output with inference_geo: "us" costs $44 per million tokens.

Four ways the upgrade returns HTTP 400

Both models enforce the same request rules, and Anthropic lists the same four breaking changes for each:

  1. Thinking can't be turned off. thinking: {"type": "disabled"} or a manual budget_tokens returns a 400. Leave the field out and set output_config.effort.
  2. Forced tool use is rejected. tool_choice set to {"type": "any"} or {"type": "tool"} returns a 400. Use auto with strict: true, and say in the prompt when the tool applies.
  3. History has to be append-only. The API checks that nothing before a thinking block, from the system prompt and tools to earlier messages, changed since the model wrote it. Opus 5.5 enforces this on accounts created on or after August 31, 2026. Fable 5.2 enforces it on every account, so history-editing code on an older account can pass on Opus and fail on Fable. The stopgap is the thinking-binding-controls-2026-08-01 beta header with thinking.block_binding.prefix_mismatch_behavior set to "drop_block".
  4. The old computer-use tool is gone. computer_20251124 returns a 400 on the Claude API and Google Cloud, and Amazon Bedrock is the one place both models still accept it. Declare computer_toolset_20260801 instead.

Two quieter changes pass every status check. A refusal comes back as HTTP 200 with stop_reason: "refusal", so code that only reads the status code treats it as an answer. And on Opus 5.5 the short notes between tool calls arrive as thinking blocks with empty text at the default display setting, so a progress feed goes silent without an error. TriSeed's migration write-up is titled, fairly, "The Changes That Pass Your Tests and Break Production."

Fallbacks, and why escalation runs one way

Both models run safety classifiers. When one fires, another model can answer instead: Claude Opus 4.8 for cybersecurity, and Claude Opus 5 for biology and for a narrow set of frontier-model work, such as kernel development on certain ML accelerators. Anthropic's apps do this on their own. On the API it is opt-in, through server-side fallback (fallbacks: "default", in beta), the SDK middleware or your own retry.

Reasoning only travels upward. On the Claude API, Fable 5.2 reads Opus 5.5's thinking blocks, so a conversation escalated from Opus to Fable keeps its chain of thought. Opus 5.5 can't read Fable's blocks, and neither can the fallback models, so a step back down, or a fallback, runs the next turns without that reasoning. The request still succeeds, and the dropped blocks aren't billed.

The launch demos are long jobs

  • An operating system from an empty folder. One Anthropic engineer had Fable 5.2 coordinate more than 1,000 subagents for eight days. The system boots on a laptop and has disk, keyboard and network drivers, file storage, windows, a login screen and an installer.
  • Three working devices. With the circuit-board maker Diode, Fable designed a USB speakerphone, a home voice-assistant hub and a handheld virtual pet, from four-layer boards to enclosures and firmware. All three worked without changes to their circuitry, though Anthropic says it doesn't yet know how they will hold up or whether they would pass commercial certification.
  • Two open physics problems. Anthropic says Fable 5.2 solved two open problems about black holes in holographic duality, with a preprint for each.
  • A rewrite in Rust. Agents on an earlier version of Fable 5.2 moved a production service from Python to Rust, and it now starts about 26 times faster.

Anthropic adds that over 90% of the changes merged into its core codebase are now written by Claude, with human guidance. The seller picks the demos, so treat them as a ceiling. They do match the second table: hardware, science and long projects are where Fable pulls away.

What people argued about

Ethan Mollick's launch essay, "Fable 5.2 and Jagged AGI," calls the model "almost" AGI, with gaps that are "no longer about simple errors or hallucinations." Fable turned his tiny USB keyboard into a game console on the first try, then needed several rounds of feedback, starting with "a brutal slog through the most boring webpages," before its literature course was any fun.

Opus 5.5 had a rough weekend before Fable arrived. "Is Opus 5.5 nerfed? Thinking patterns changed on November 3" reached 755 points and 478 comments on Hacker News. Commenters blamed quantization, capacity being freed for the next model and A/B tests on subscription traffic that API benchmarks would never see. Others replied that it was a new thinking summarizer and that the nerfs didn't show on sites that track model performance over time. Both camps had anecdotes. Neither had Anthropic's server logs.

Claude Code had its own week. A GitHub issue titled "Bug Report: Privacy violation in User-Agent header" says Claude Code put the user's email address in the User-Agent of its web searches, and the Hacker News thread about it passed 1,600 points. Another thread, about background agents sending their transcripts to the model, drew 591. Those are harness problems, so they follow you whichever model you pick inside Claude Code.

Which one to use

If the job is Start with Why
Everyday coding and review Opus 5.5 at medium or high One to four points behind Fable, and cheaper per task on CursorBench
Terminal and DevOps agents Opus 5.5 Tied with Fable on Vals AI's Terminal-Bench 4.0, at $12.04 a test against $14.56
Writing, analysis and knowledge work Opus 5.5 Ahead on both Elo ratings, GDPval-AA and AA-Briefcase
High-volume pipelines Opus 5.5 $2 and $10 on the Batch API
Anything under zero data retention Opus 5.5 Fable 5.2 needs 30-day retention unless Anthropic authorizes otherwise
Patches that must merge Fable 5.2 at high Opus's best FrontierCode score at about half the cost
Science, hardware and robotics Fable 5.2 +15.9 on Terminal-Bench-Science, +18.2 on BenchCAD, +25.5 on RoboDojo
Whole projects that run for hours Fable 5.2 41.3% on the Remote Labor Index against 21.3%
Code from screenshots and mockups Fable 5.2 +12.7 on SWE-bench Multimodal
Vetted security or biology research Mythos 5.2 The same weights with lighter safeguards, through Anthropic's verification programs

In practice that means a router. Send work to Opus 5.5 by default and escalate to Fable 5.2 when a task is long, physical, scientific or visual, and the conversation keeps its reasoning on the way up. Pin the effort in every request, and start Fable lower than feels natural. Both models are available on OpenCheese, where Opus 5.5 has been since launch day, so switching between them is a change of model ID.

Anthropic now sells a 5.2 that outranks a 5.5, a Mythos that is a Fable with paperwork, and a fast mode whose availability depends on which cloud you pay. The names will keep getting stranger. Divide the bill by the tasks that actually got done. Division, unlike Claude, doesn't get point releases.

Sources

The cheese has holes. Your infrastructure should not.