Claude Fable 5.2 Is Out: $8/$40, Fast Mode, When to Switch

Fable 5.2 is cheaper than 5.1 and, on Anthropic’s charts, stronger at the long jobs. SandBase lists it as anthropic/claude-fable-5.2, and the card shows $8 / $40 with no discount on that row.

I opened Anthropic’s announcement on 10 November 2026 because the last Fable story on this blog was still a leak log. That log was about a 5.5 that never received a row. What shipped on 9 November is Claude Fable 5.2. The Claude API id is claude-fable-5-2. List price is $8 per million input tokens and $40 per million output tokens. Fast mode is $16 and $80.

If a job has to run this week, the Fable id on SandBase is 5.2. Claude Fable 5.2 returned HTTP 200 on 10 November 2026. Earlier the same day that path was Model Not Found. That 404 is not the live card later on this page.

Key takeaway

• Anthropic published Claude Fable 5.2 and Claude Mythos 5.2 on 9 November 2026. List price is $8 / $40 per million tokens. Fast mode is up to about 2.5× token generation at $16 / $80.
• The model page lists 1M context, 128K max output, adaptive thinking always on, default effort high, and a June 2026 knowledge cutoff. The Claude API id is claude-fable-5-2. The Bedrock id is anthropic.claude-fable-5-2.
• Anthropic says Fable and Mythos are the same underlying model. Mythos is for organizations in the Life Sciences or Cyber verification programs. Fable is the general release, with those safeguards on.
• Their routing line: start most work on Opus 5.5. Move to Fable 5.2 for long-horizon agentic work, hard coding, or science, or when Opus at a higher effort still falls short.
• SandBase lists anthropic/claude-fable-5.2. On 10 November 2026 the card showed $8 / $40 per million tokens, cache read $0.20, cache write $10, 1M context, 128K max output, status Available. No discount is printed on that price row. Chat Completions and Messages are both on the card, at api.sandbase.ai. Earlier the same day the path was 404.

Where the Fable line is today

The October leak log is still the right record of what did not ship. Claude Fable 5.5 leaks graded the Azure row, the Onigiri codename, and the second-checkpoint claim. None of those became a 5.5 launch. Prompts people pinned to that name were run on models that exist, in viral prompts, tested on Fable 5.1 and Opus 5.5.

The generation you can treat as a product starts at Fable 5.1. What Fable 5.1 changed for coding and cache pricing and why a Fable account is not a Mythos account are the baseline. Fable 5.1 long-running agents is the evaluation shape I would reuse before trusting a new id. Opus, the model Anthropic wants as the daily driver, is in the Opus 5 deep dive and the later Opus 5.5 pricing note.

This page is the release. Cost and a routing table against Opus 5.5 and GPT-6 Astra are in the comparison. The parameters that return 400, and the curl you can send today, are in the agentic path.

What the 9 November post says

Anthropic’s announcement, captured 10 November 2026. The page is dated November 9, 2026. Source: anthropic.com/claude-fable-and-mythos-5-2.

The post says testers saw outputs up to 30% faster than Fable 5.1 and Mythos 5.1, at a lower price than 5.1’s $10 / $50. That 30% is Anthropic’s figure for generation speed. It is not a latency SLA, and it is not a stopwatch from this desk.

The same announcement, scrolled to the price. Anthropic writes that Fable and Mythos are the same underlying model, that the price is $8 / $40 per million input and output tokens, and that fast mode is up to 2.5× token generation at $16 / $80.

I am not going to upgrade “the same underlying model” into “the same weights.” That is the sentence they published. Fable carries cybersecurity and biology safeguards so it can be offered generally. Organizations doing beneficial work in those areas apply through the Life Sciences and Cyber Verification Programs. The docs say Mythos 5.2 offers the same capabilities only to verified organizations and shares Fable 5.2’s specifications and pricing. A Fable key is not Mythos access. The 5.1 version of that split is already written up in the access-tier note.

Confirmed vs unknown

Confirmed means a primary page said it, and I retrieved that page on 10 November 2026 (Asia/Shanghai). Unknown means it was not on those pages.

Fable 5.2 overview, captured 10 November 2026. Header: claude-fable-5-2, 1M context, 128K max output, $8 input and $40 output. The comparison table sets default effort to high on Fable 5.2 and medium on Opus 5.5. Both list a June 2026 knowledge cutoff.

SandBase’s Claude Fable 5.2 card on 10 November 2026, the live page, not an earlier 404. The id is anthropic/claude-fable-5.2. The price row shows $8.00 input and $40.00 output per million tokens, cache read $0.20, cache write $10.00, context 1M, max output 128K, status Available. No discount is printed on that row.

Input on the overview is text and images. Output is text. Adaptive thinking is always on and cannot be turned off. Retirement is “not sooner than November 9, 2027.” Those are lifecycle facts, not a promise that the behavior stays fixed.

Price, against 5.1 and against Opus

Fable 5.1’s SandBase card still shows list $10 / $50, with a 15% discount displayed as $8.50 input and $42.50 output per million tokens. Cache read on that card was $0.21 against a $0.25 list, and cache write $10.63 against $12.50. Opus 5.5’s card showed list $4 / $20 and the same 15% discount, at $3.40 / $17, cache read $0.17, cache write $4.25. Those discounts are a SandBase bill. They are not Anthropic’s rate card, and they move. Dated to the cards opened on 10 November 2026.

A request with 100,000 input tokens and 20,000 output tokens, at Anthropic list prices, before cache:

$8 is 20% under $10, and $40 is 20% under $50, on the list rate, not on a cached bill. Cache read is $0.20, which the docs call 2.5% of base input for this model. Batch is half off. I did not meter a session, so I will not invent a cache-hit rate.

Fast mode doubles the token price for up to 2.5× generation speed. If the session is waiting on tests, a browser, or a person, you pay double for a slice of the timeline. I would leave it off until a trace shows token generation as the wait. The agentic note is where that trade sits next to effort.

When to leave Opus 5.5

Anthropic’s models overview says to start with Opus 5.5 for most workloads, and to use Fable 5.2 for demanding reasoning and long-horizon agentic work, or when your evals on Opus at higher effort still fall short. That is a better rule than “newest id wins.”

Opus 5.5’s API default effort is medium. Fable 5.2’s is high. A swap that only changes the model id runs one effort level hotter, and thinking tokens are billed as output. The migration guide says the request succeeds, and the change shows up as cost, latency, and results. Set effort on purpose.

Two cells in Anthropic’s own table stop me from calling 5.2 a sweep. On AA-Briefcase v1.1 they list Fable 5.2 at 66.1% rubric and 1785 Elo, and Opus 5.5 at 61.2% rubric and 1800 Elo. Rubric up, Elo down. On AutomationBench they list Fable 5.2 at 51.4% and Gemini 4 Carbon at 53.9%. The grid, including GPT-6 Astra, is in the comparison. Every score there is Anthropic’s, with the effort notes they printed. I did not re-run them.

Other cells they highlight, still as vendor results: FrontierCode 2.0 at 49.1% with Xhigh effort (High effort is about 45.3%, level with Opus 5.5’s 45.5%, at roughly half the cost per task on that eval); FrontierSWE v2 69.9%; Terminal-Bench-Science 74.6% (Mythos 5.2 at 75.3% on that row); WANDR relaxed 82.1%; OSWorld 2.1 89.3% on a partial offline set; RoboDojo Sim 64.9% on an internal run; Humanity’s Last Exam with tools 74.2%; Chartography without tools 73.4%. The system card is the methodology.

The same post describes an eight-day operating-system experiment, a 30-hour Fermat formalization, two black-hole calculations, and three working devices. Those are Anthropic’s accounts of prerelease use. I would not put them in a buying doc. The Fable 5.1 science note is the older standard for how thin a research anecdote is allowed to be.

Try today on SandBase

The call that matches the live card is Chat Completions, with the id the card prints: anthropic/claude-fable-5.2. The same page lists Messages at /v1/messages. This sample does not set effort or speed. Those fields are not in the card’s Chat Completions schema, and I did not capture a 5.2 completion for this page.

curl --fail-with-body --silent \
  -X POST "https://api.sandbase.ai/v1/chat/completions" \
  -H "Authorization: Bearer $SANDBASE_API_KEY" \
  -H "Content-Type: application/json" \
  -d '{
    "model": "anthropic/claude-fable-5.2",
    "messages": [
      {"role": "user", "content": "Name the riskiest assumption in this migration plan, then the check that would falsify it."}
    ],
    "max_tokens": 4096
  }'

At the card’s displayed rates, that bill is $8 per million input and $40 per million output. The 100k / 20k example above is $0.80 + $0.80 = $1.60, the same arithmetic as Anthropic’s list, because this card does not show a discount. Arithmetic on the card, not a receipt. The hyphenated claude-fable-5-2 is the Claude API id, not the SandBase string.

SandBase’s Claude Fable 5.1 card on 10 November 2026, kept as the previous id. The id is anthropic/claude-fable-5.1. Displayed prices are $8.50 input and $42.50 output per million tokens, against list $10 and $50, with a 15% off badge. Context 1M, max output 128K, status Available. This discount is 5.1’s. It is not on the 5.2 card.

Daily work can stay on Claude Opus 5.5, id anthropic/claude-opus-5.5. That card priced it at $3.40 / $17 that morning against a $4 / $20 list. The same 100k / 20k sketch is $0.68 at those displayed rates. Claude Fable 5.1 is the fallback when you already have a score on that id. Its displayed $8.50 / $42.50 is still above the 5.2 card’s $8 / $40.

The Anthropic catalog on SandBase shows the 5.2 tile as Available at $8 / $40. The crop is in the agentic note. I am not using an 18 October catalog snapshot to decide whether 5.2 is listed. The Fable 5.2 card earlier on this page is the check.

FAQ

Is Claude Fable 5.2 released?

Yes. The announcement is dated 9 November 2026. The model overview lists claude-fable-5-2 as Active (latest), released November 9, 2026, with retirement not sooner than November 9, 2027.

How much does Claude Fable 5.2 cost?

On Anthropic’s rate card: $8 per million input tokens and $40 per million output tokens. Cache read is $0.20, a five-minute cache write is $10, a one-hour cache write is $16, and the batch API is 50% off input and output. Fast mode is $16 and $80. SandBase’s card on 10 November 2026 shows the same $8 / $40, plus cache read $0.20 and cache write $10. It does not print a discount on that row. I am not quoting a SandBase fast-mode price, because that card did not show one.

Is Mythos 5.2 a different model?

Anthropic says they are the same underlying model. Fable has the cyber and biology safeguards and is the general release. Mythos is for organizations verified through the Life Sciences or Cyber programs. The docs say specs and price match. A Fable API key does not grant Mythos.

Should I switch from Opus 5.5?

Not as a default. Anthropic tells most workloads to start on Opus 5.5. Move a job to Fable 5.2 when it is the hard, long one, and only after you set effort, because Fable’s API default is high and Opus’s is medium. On SandBase the string for that move is anthropic/claude-fable-5.2. The Opus 5 deep dive is still the right background for the Opus line you may already run.

Can I call Fable 5.2 on SandBase?

Yes. The card linked above returned HTTP 200 on 10 November 2026. The model string is anthropic/claude-fable-5.2, with dots, not the hyphenated Claude API id claude-fable-5-2. Earlier the same day the path was Model Not Found. Fable 5.1 and Opus 5.5 stay listed when the job should remain on those ids.

Related reading

• Fable 5.2 vs Opus 5.5 vs GPT-6 Astra: list prices and a routing table. Anthropic’s benches stay labeled as theirs.
• Agentic path: effort, fast mode, and the call to send today.
• The Fable 5.5 leak timeline, kept as history, and the prompt tests that ran on shipped models.
• Fable 5.1 coding, research, and cache pricing and Opus 5.5 API pricing.

What I would ship this week

I would not point production at Fable 5.2 until effort, retention, and tool choice are explicit. The id itself is listed. Opus 5.5 remains the daily driver Anthropic describes. Fable 5.1 remains on SandBase for jobs you have already scored on that id. The long-running agent note is how I would freeze a baseline before the string changes.

Sign in and open Claude Fable 5.2. Ordinary work can still go through Claude Opus 5.5. Keep the model string in one place.

Sources retrieved 10 November 2026 (Asia/Shanghai): Anthropic announcement, Fable 5.2 overview, Fable 5.2 migration guide, Claude models overview, and the SandBase cards linked above, including the Fable 5.2 card at $8 / $40.