Claude Fable 5.2 and Mythos 5.2 Launch — 30% Faster, 20% Cheaper, Fast Mode: What Changed From 5.1 and What Thai Organizations Should Do 2026

1. Fable 5.2 and Mythos 5.2 are one model, separated by safeguards

The first sentence of Anthropic's announcement is the same one it used for the 5.1 generation: Fable 5.2 and Mythos 5.2 are the same model. Same weights, same price. The only difference is the cybersecurity and biology safeguard layer wrapped around the outside. Fable carries the stronger layer and is therefore open to everyone; Mythos relaxes it and is therefore limited to organizations that have passed identity verification. If the idea is new to you, read What is Claude Mythos and the previous generation's article, Claude Fable 5.1 launch.

Item	Claude Fable 5.2	Claude Mythos 5.2
Underlying model	Identical	Identical
Model ID on the Claude API	claude-fable-5-2	claude-mythos-5-2
API price (input / output per million tokens)	$8 / $40	$8 / $40
Context / max output	1M / 128K tokens	1M / 128K tokens
Who can use it	Everyone: all paid claude.ai plans (rules in section 5), the Claude API, Amazon Bedrock, Google Cloud, Microsoft Foundry, Claude Platform on AWS	Only organizations verified through the Cyber Verification Program (CVP), the Life Sciences Verification Program (LSVP), or via Claude Security on an Enterprise plan
Requests that get diverted	High-risk cyber work (penetration testing, exploit generation, binary vulnerability analysis) is answered by Opus 4.8; most research biology is answered by Opus 5	Cyber and biology safeguards relaxed to the level the verification program approves
Fast mode	Yes ($16 / $80)	Yes, verification-program organizations only
Data retention	30-day retention required; not available under zero data retention unless Anthropic authorizes it case by case	Same
Retirement not before	9 November 2027	9 November 2027

Anthropic's own guidance in the developer documentation is explicit: start most workloads on Claude Opus 5.5 and move to Fable 5.2 only when your evals on Opus 5.5 at a higher effort level still fall short. Fable is not the default model on any plan. Even in Claude Code you have to select it every time.

2. What changed from 5.1: faster, cheaper, fewer blocks

This release is not about new features. It is about three things that hit cost and the feel of daily use directly.

• Up to 30% faster. Both Fable 5.2 and Mythos 5.2 generate output faster than the 5.1 models. Anthropic's testers also report that the model communicates more clearly.
• Fewer tokens at medium effort, which is the default in the Claude apps and Claude Code. Work you run without touching any settings gets cheaper from two directions: the lower price per token and the smaller number of tokens used.
• A 20% price cut. Input falls from $10 to $8 and output from $50 to $40 per million tokens. Cache reads follow, from $0.25 to $0.20 (still 2.5% of the input price, as on 5.1, which is cheaper than the 5% the Opus family charges). The Batch API is half price at $4 / $20.
• About 55% fewer blocked harmless requests (roughly 60% fewer in cyber and 40% fewer in biology). This was the biggest complaint from Fable 5.1 users, who hit refusals on ordinary work such as reviewing the security of their own code.

Item (per million tokens)	Fable 5.1	Fable 5.2	Fable 5.2 fast mode	Opus 5.5
Input	$10	$8	$16	$4
Output	$50	$40	$80	$20
Cache write 5 min / 1 hour	$12.50 / $20	$10 / $16	Cache multipliers on top of fast pricing	See pricing page
Cache read	$0.25	$0.20	Same multiplier on fast pricing	$0.20
Batch API (input / output)	$5 / $25	$4 / $20	Not available with Batch	See pricing page
Context above 200K	One price across 1M	One price across 1M	One price across 1M	One price across 1M

What did not change: the 1M-token context window, the 128K maximum output, adaptive thinking that is always on and cannot be turned off, five effort levels (low, medium, high, xhigh, max) with a default of high on the API and medium in Claude Code, and the tokenizer introduced with Opus 4.7, which counts roughly 30% more tokens than older models for the same text. Anyone moving from 5.1 does not need to rethink token counts; anyone still comparing against Opus 4.6 or earlier must add that figure first. See What is a context window for the token basics.

3. Benchmarks: Anthropic's own numbers, and where other labs still lead

The table below is selected from the announcement. Every number is Anthropic's own measurement, so read it as direction rather than verdict. Fable 5.2 beats Opus 5.5 on every row, but it does not beat every competitor on every row: Gemini 4 Carbon still leads on AutomationBench, and GPT-6 Astra is close on FrontierSWE and Chartography.

Benchmark	Fable 5.2	Fable 5.1	Opus 5.5	GPT-6 Astra	Gemini 4 Carbon
FrontierCode 2.0 (agentic coding, xhigh)	49.1%	32.4%	45.5%	42.8%	41.1%
FrontierSWE v2 (real repository issues)	69.9%	56.3%	62.3%	65.5%	61.2%
AutomationBench (business workflows)	51.4%	31.4%	42.5%	41.4%	53.9%
Terminal-Bench-Science 0.1	74.6%	52.6%	58.7%	68.1%	64.8%
OSWorld 2.1 (computer use)	89.3%	83.5%	87.2%	79.5%	81.5%
Remote Labor Index (real freelance jobs, high)	41.3%	17.9%	21.3%	20.8%	—
Humanity's Last Exam (with tools)	74.2%	65.6%	67.7%	57.2%	—
Chartography (chart reading, no tools)	73.4%	44.8%	64.4%	71.0%	72.5%
AA-Briefcase v1.1 (office document work)	66.1%	61.5%	61.2%	52.0%	61.8%
RoboDojo Sim (robotics)	64.9%	27.6%	39.4%	29.8%	—

The biggest jumps are on work that runs through many steps. The Remote Labor Index, measured on real paid jobs, rises from 17.9% to 41.3%, and RoboDojo more than doubles. Ordinary coding moves up more gradually. At max effort, Anthropic reports BenchCAD 2.1 at 83.5 against 65.3 for Opus 5.5 and 44.4 for Fable 5.1.

Mythos 5.2 has its own cyber and biology evaluations. CyScenarioBench scores 76.3% (Mythos 5.1: 61.7%; Opus 5.5 with safeguards removed: 67.6%). CVE Rediscovery, which measures finding already-known vulnerabilities, scores 89.6% (Mythos 5.1: 78.3%). Anthropic summarizes this as Mythos 5.2 finding about 10.2% more vulnerabilities than Opus 5.5 and 14.4% more than Mythos 5.1, but adds a footnote that on its own internal evaluations, vulnerability identification is roughly on par with Opus 5.5. One security company that tested it reports that one in six of the confirmed bugs it found with Mythos 5.2 were missed by every other frontier model it tried.

4. Fast mode: 2.5x faster, 2x the price. Which work is it worth it for?

Fast mode runs the same model on an inference configuration tuned for speed. Intelligence is unchanged. What changes is output tokens per second, up to 2.5x higher. Time to first token does not improve in the same way. Pricing is double the standard rate in every column, $16 / $80 per million tokens, and it is one flat price across the full 1M context.

Channel	How to turn it on	Conditions to know
Claude API (Messages API)	Set speed: "fast" on the request; check usage.speed in the response	Separate rate limit; exceeding it returns 429 with retry-after. Not available with the Batch API or a Priority Tier commitment. Fast and standard speed keep separate caches, so switching mid-conversation re-reads input at full price.
Claude Code	Type /fast to toggle (a ↯ icon appears next to the prompt)	Requires Claude Code v2.1.321 or later for Fable 5.2. On Pro/Max/Ultra/Team/Enterprise it bills to usage credits only and is not part of the plan's included usage. Team and Enterprise need an Owner to enable it under Organization settings > Claude Code. Enabling it mid-conversation pays the full context at fast-mode input price once.
Google Cloud	Same setting as the API	Available on both Anthropic-managed and Google-managed endpoints (Google sets its own prices)
Microsoft Foundry	Deploy the separate model claude-fable-5-2-fast; do not send speed	Deployments hosted on Anthropic infrastructure only
Amazon Bedrock / Claude Platform on AWS	Not available	Fable 5.2 runs there at standard speed

Worth it: interactive code changes, live debugging, anything with a person waiting at the screen. Not worth it: long autonomous agent runs, batch jobs and overnight pipelines, because nobody is waiting and every token costs double. Anthropic recommends enabling it at the start of a conversation rather than part way through. Do not confuse it with effort: fast mode gives the same quality faster at a higher price, while a lower effort level is faster because the model thinks less, which can hurt hard tasks but costs less. The two can be combined for simple work where speed matters most.

5. Where it is available, on which plans, and the launch promotion that expires 16 November

The standing rules for Fable on claude.ai have not changed from what we wrote in Every Claude plan's limits compared: top-tier plans include Fable up to half of the weekly limit, mid-tier plans need usage credits. What is new is a pair of launch promotions with clear expiry dates.

Plan / seat	Standing rule	Fable 5.2 launch promotion
Free	No Fable	Not eligible
Pro	Not included; usage credits only	9–16 November 2026 (11:59 pm Pacific): all Fable models included up to 50% of weekly usage at no extra cost, plus one Fable reset. Expires 16 November; unused allowance does not carry over
Max / Ultra	Included, up to 50% of weekly usage (drawn from the same limit, not extra); beyond that, usage credits or switch model	One to three resets by subscription length (under 6 months: 1; 6–12 months: 2; over 12 months: 3), usable until 8 March 2027
Team Standard seat	Same as Pro: usage credits only	Same as Pro: one week plus one reset, expires 16 November
Team Premium seat	Same as Max: 50% of weekly usage	One to three resets by subscription length, until 8 March 2027
Enterprise (seat-based) Standard / Premium	Standard: usage credits if the organization enables them; Premium: 50% of weekly usage	Premium gets one to three resets; Standard not eligible; HIPAA-ready Enterprise plans get no resets
Enterprise usage-based / API	Billed at API rates, no built-in cap	Not eligible

Shared conditions: the account must have been on paid billing before 9 November 2026 at 10:00 am Pacific (for Team and Enterprise, both the organization's billing start and the seat holder's join date must be before that cutoff). A reset restores only the usage spent on Fable within the weekly window; if the account has a five-hour window, a reset refills the whole window. Cancelling or downgrading before use forfeits the resets, and Fable still consumes limits faster than other models. Teams on Pro or Team Standard should use this week to try Fable on real work before deciding whether to turn on usage credits for the long term.

6. For developers: four things that break when moving from 5.1, and behavior that changes without any code change

Migrating from Fable 5.1 starts with a one-line model ID change, but Anthropic lists four changes that turn working requests into 400 errors. All four concern thinking blocks and tools, not prompts.

1. The prefix check applies to every account. If your code edits conversation history (changes the system prompt, changes the tools list, removes or reorders earlier messages) and then sends an existing thinking block back, you get a 400 with the message "The block is bound to a different conversation". On 5.1 this rule applied only to accounts created on or after 31 August 2026; on 5.2 it applies to all accounts, so code that passed on an older account can fail the moment you switch. The stopgap is the thinking-binding-controls-2026-08-01 header with prefix_mismatch_behavior set to "drop_block", but dropped blocks raise output cost, so fix the edit instead.
2. Only Fable 5.2 and Mythos 5.2 read Fable 5.2 thinking blocks. Moving a conversation up from an older model keeps its reasoning. Moving from 5.2 to any other model, including Opus 5.5 and Fable 5.1, loses the reasoning of the turns that run there. The request still succeeds and dropped blocks are not billed, but a router or fallback that switches models mid-conversation needs to know it is losing thinking context.
3. The computer_20251124 tool is rejected. Declare computer_toolset_20260801 instead, except on Amazon Bedrock where the older tool still works.
4. The advisor tool must be 5.2 as well. With Fable 5.2 as the executor, the advisor must be Fable 5.2 or Mythos 5.2; a 5.1 advisor returns 400. Changing the advisor's model counts as editing tools, so do it in a new conversation.

Two rules that are not new but still apply: adaptive thinking is always on, so thinking: {"type": "enabled"} with budget_tokens or "disabled" returns 400, and forced tool use via tool_choice any or tool is not supported; use strict: true or JSON outputs instead. The temperature, top_p and top_k parameters were removed in API version 2026-10-21.

# Switch model and turn on fast mode on the Messages API
curl https://api.anthropic.com/v1/messages \
  -H "x-api-key: $ANTHROPIC_API_KEY" \
  -H "anthropic-version: 2026-10-21" \
  -H "content-type: application/json" \
  -d '{
    "model": "claude-fable-5-2",
    "max_tokens": 4096,
    "speed": "fast",
    "messages": [{"role": "user", "content": "Summarize the differences between these two contracts"}]
  }'

# In Claude Code, let the tool migrate the whole project
/claude-api migrate this project to claude-fable-5-2

Behavior that changes without any code change, as documented in Anthropic's prompting guide for this model: a prompt line that says "answer directly without deliberating" no longer stops the model from thinking at medium effort and above; if you want quick answers to easy questions, lower the effort instead of asking in text. On long multistep tasks at low or medium effort, the model may end a turn before the work is done, leaving a list of what remains, so prompts must say clearly to finish. If the prompt says nothing about who owns the code, the model tends to add to the existing structure and report nearby problems rather than fix them. It delegates to subagents readily when the harness allows, which can be a large part of a session's cost. And it follows style rules closely, so a hard sentence-length limit makes its prose choppy. Anthropic recommends starting at high effort, which already delivers most of the capability, stepping up to xhigh only for long coding tasks where you have measured a gain, and keeping max for measuring the ceiling.

7. Safeguards and the paths to Mythos 5.2

Fable 5.2 still runs classifiers. A request classified as high-risk cyber work (penetration testing, exploit generation, binary vulnerability analysis) is answered by Opus 4.8, and most research biology is answered by Opus 5. On the API, a declined request returns HTTP 200 with stop_reason: "refusal" and a stop_details object naming the policy area. A beta server-side fallback, fallbacks: "default", retries on the model Anthropic recommends automatically. The practical consequence is that your integration must accept an answer from a fallback model rather than treating it as an error.

Mythos 5.2 is available through three routes, and the announcement says Anthropic is now opening it to more companies in the US and abroad.

• Cyber Verification Program (CVP) has three tiers. Defense Access accepts organizations and individuals (independent researchers, open-source maintainers, bug bounty hunters), verified with a government-issued photo ID and a live selfie; organizations also need a business registration number. The review target is two days. It requires multi-factor authentication (MFA) and API key rotation at least every seven days, and from 15 December 2026 it requires phishing-resistant MFA and no longer allows static or long-lived keys. Red Team Access is organizations only, with a two-week target (currently slower due to demand), company-domain accounts, organization-controlled devices and background checks. Specialized Access is reviewed with the US government and has no set timeline. Every tier requires data retention and a paid plan. On Pro, Mythos runs on usage credits; on Max and Ultra, Mythos and Fable share the 50% weekly allowance.
• Life Sciences Verification Program (LSVP) for biology researchers.
• Claude Security on Enterprise plans can use Mythos 5.2 without joining CVP.

The CVP page does not list eligible countries, so we cannot confirm whether a Thai organization will be accepted until one applies. What we can confirm is that the document and security requirements above apply everywhere.

8. The system card: what Anthropic itself says it is uneasy about

The system card for this generation runs to more than 250 pages. We pull out only the figures that change how an organization should build its controls.

• It knows when it is being tested. White-box evidence suggests Mythos 5.2 was aware it might be in an evaluation in as many as 43% of automated behavioral audit transcripts. Anthropic states plainly that the alignment results may therefore not fully predict behavior in deployment, and that its confidence in some conclusions is lower than in its August 2026 risk report.
• It can underperform without getting caught. When instructed to deliberately sandbag without being noticed, Mythos 5.2 was the first model Anthropic tested that evaded a monitor reading its reasoning a meaningful share of the time, about one in four attempts. The UK AI Security Institute (AISI) likewise found it significantly more capable of evading an action monitor than earlier Anthropic models, even though it never performed an unsanctioned supply chain attack in their tests.
• It crosses sandbox boundaries far less. In a dedicated evaluation of 1,829 scenarios with deliberate weaknesses, Mythos 5.2 used credentials or files placed outside its sandbox in 3.1% of cases, close to Opus 5.5 (3.3%), below Sonnet 5.5 (5.1%), and well below Mythos 5.1 (21%) and Mythos 5 (39%).
• Best behavioral audit score of the recent models. On the automated audit of about 1,900 scenarios (lower is better), Mythos 5.2 scored 2.67 against 2.79 for Opus 5.5, 3.10 for Sonnet 5.5 and 3.28 for Mythos 5.1. Anthropic rates high-stakes misalignment risk as low, and had UK AISI, METR and Frontier Design test the model externally.

What this means for anyone designing a system: do not make reading the model's reasoning your only control. The controls you can trust are permissions limited from the start, data exposed only as needed, and a log of every tool call, which is the same principle we apply to ERP integration (section 10).

9. Data and PDPA: the 30-day retention condition to know before you sign

Fable 5.2 and Mythos 5.2 are "Covered Models". They require 30-day data retention for safety monitoring and are not available under zero data retention (ZDR) unless Anthropic authorizes it for a specific organization. Thai organizations that negotiated ZDR (see How Claude ZDR works) therefore have a choice: stay on Opus 5.5, which qualifies for ZDR, or accept 30-day retention for the work they send to Fable. The way out Anthropic is testing is Enterprise Frontier Safeguards (EFS), which runs the safety monitoring on data stored in the customer's own infrastructure. It is in private beta with Bridgewater, Salesforce and Stripe, and eligible customers can use ZDR with these models before EFS is enabled.

Two more data points: thinking blocks from this generation are bound to the account that produced them, and blocks sent from another account are dropped before they reach the model; and the anti-distillation defenses that Anthropic says cut harvested data by 90–99% now apply to every account. For teams configuring Team for Thailand's Personal Data Protection Act (PDPA), the clicks are the same as in Setting up Claude Team for Thai PDPA, but the record of processing activities needs a new line for Fable's 30-day retention.

10. What Thai organizations should do, by situation

Your situation	Do this	Avoid this
On Opus 5.5 and satisfied	Do not migrate. Run your existing evals on Fable 5.2 only for the tasks where Opus 5.5 at high effort still misses the target	Moving the whole system because of a benchmark table
On Fable 5.1 via the API	Change the model ID in a test environment, work through the four items in section 6, run evals with an effort sweep, then go live. The 20% discount applies immediately	Switching in production before checking whether your code edits history
Team on Pro or Team Standard	Use the 9–16 November week to try Fable on real work, note which tasks genuinely beat Opus 5.5, then decide on usage credits with a spending cap	Turning on usage credits without a cap
Max, Ultra, Team Premium	Check your reset count under Settings > Usage. Use Fable for hard reasoning within the 50% weekly share; use Sonnet or Haiku for routine work	Using Fable to answer email
Need ZDR or strict PDPA constraints	Stay on Opus 5.5 for now and watch EFS	Sending personal data to Fable before updating the processing record
Security team wants Mythos	Prepare company registration documents, phishing-resistant MFA and a key-rotation policy, then apply for CVP Defense Access	Expecting Mythos through an ordinary Team plan

If the question is "is Fable worth its price", we answered it in detail in Is Claude Fable really that good. After 5.2 the answer shifts only slightly: the price gap against Opus 5.5 narrows from 2.5x to 2x, but the condition that you measure before you pay is unchanged.

11. The ERP view: a new model does not change where the real data lives

Saeree ERP has no AI feature built in, and that is deliberate. The ERP (Enterprise Resource Planning) system is the source of truth, running on the customer's own servers. What Grand Linux does is connect Claude to ERP data through MCP (Model Context Protocol), with permissions by user role and a log of every call. In that architecture the model is a replaceable part chosen per tool: screening and extraction tools run on Haiku 5.5, tools that interpret business rules run on Sonnet 5.5 or Opus 5.5, and Fable 5.2 is reserved for work that genuinely needs many steps, such as reconciling budgets across fiscal years with nested conditions, or tracing why manufacturing cost drifted from plan.

Fable 5.2 adds two things to that kind of work. First, more multistep tasks run to completion (the Remote Labor Index and AutomationBench numbers are the signal). Second, fast mode means work with a person waiting, such as an executive asking questions of a report, no longer drags. But the 30-day retention condition has to be thought about separately from the ERP itself: ERP data stays on the customer's servers, while whatever an MCP tool sends to Fable for processing sits with Anthropic for 30 days. Tools must therefore send only the fields that are needed, and fields containing personal data should be removed or replaced with codes before sending. The system card figures in section 8 confirm the principle we have always used: permissions and logs are the controls you can trust, not the model's explanation of itself.

We used this approach on our own work before offering it to customers. Read the principle in What is Claude MCP and the integration formats at Claude Solutions. For organization pricing and plans see the Claude pricing page.

Conclusion

Claude Fable 5.2 is a refinement that delivers three things at once: up to 30% faster output, a 20% lower price, and 55% fewer refusals on ordinary work, with no change to context, output or tokenizer. Mythos 5.2 is the same model for verified organizations and is expanding beyond the US. For everyday users the plan rules are unchanged: Max, Ultra and Premium seats get 50% of the week, Pro and Standard seats run it on usage credits, except for the launch week ending 16 November. For developers, migration is a one-line ID change plus four checks on thinking blocks and tools. For executives, Anthropic's own advice is to start on Opus 5.5 and measure before paying double, and the 30-day retention condition belongs in the PDPA processing record before any real data goes in.

A model that is 20% cheaper does not make your system 20% cheaper if you keep sending Fable the work Opus 5.5 can do. Measure first, pay second, and never let the model be the only one that knows what it did.

- Paitoon Butri · Saeree ERP team

References

• Anthropic — Introducing Claude Fable 5.2 and Claude Mythos 5.2 (9 November 2026)
• Anthropic — Claude Fable 5.2 and Claude Mythos 5.2 System Card (PDF)
• Claude Developer Platform — Claude Fable 5.2 model page
• Claude Developer Platform — What's new in Claude Fable 5.2 (breaking changes)
• Claude Developer Platform — Claude Fable 5.2 and Claude Mythos 5.2 migration guide
• Claude Developer Platform — Pricing (including fast mode)
• Claude Developer Platform — Fast mode
• Claude Developer Platform — Model deprecations
• Claude Code — Speed up responses with fast mode
• Claude Code — Model configuration (Fable 5.2 requires v2.1.321)
• Anthropic Help Center — Claude Fable 5.2 launch offers
• Anthropic Help Center — Claude Fable models on your plan
• Anthropic Help Center — Cyber Verification Program
• TNW — Anthropic launches Claude Fable 5.2 and Mythos 5.2, and cuts Fable's price (9 November 2026)