Claude Fable 5: What Anthropic's New Flagship Can Do — and What It Means for Your Business

The launch in numbers

Anthropic released Claude Fable 5 on June 9, 2026, alongside the restricted Claude Mythos 5. Fable 5 is generally available on the API at $10 per million input tokens and $50 per million output tokens, holds a 1M-token context window, and tops Anthropic's published benchmark table on agentic coding and knowledge work.

The release made mainstream news because it came days after Anthropic publicly warned that frontier AI is becoming dangerous; TechCrunch framed it as the public version of the same technology the company is keeping partly restricted. For a business reader the more useful question is narrower: what does this model actually change in how much software work, document work and back-office work costs? We went through the launch coverage and the API documentation so you don't have to. Here is the business view, including the caveats the headlines skip.

What is Claude Fable 5?

Claude Fable 5 is Anthropic's new flagship model, released June 9, 2026. It is the same underlying model as the restricted Claude Mythos 5, deployed with safety classifiers active. It writes and migrates code at state-of-the-art level, operates desktop software through the screen, reads up to 1M tokens of context, and runs autonomous agent tasks for days.

Fable vs Mythos: one model, two deployments

Anthropic trained one model and shipped it two ways. Mythos 5, the version with some safeguards lifted, is available only to vetted cyberdefenders, infrastructure providers and select biology researchers through Project Glasswing, a program run in collaboration with the US government. Fable 5 is the generally available version: it ships with safety classifiers in three domains — cybersecurity, biology/chemistry, and attempts to distill the model itself. When a classifier fires, the request is answered by Claude Opus 4.8 instead. Anthropic says this happens in less than 5% of sessions on average (announcement).

Before release, external red-teaming and a bug bounty found no universal jailbreaks in over 1,000 hours of testing, per the same announcement. That is reassuring for procurement checklists. The practical consequence for production systems is different and often missed: your application can silently receive an Opus 4.8 answer instead of a Fable 5 answer in a small share of sessions. If your workflow depends on consistent quality or latency, you need to design for that. We come back to it in the adoption section.

Benchmarks: where it leads, and where the asterisks are

Fable 5 leads every row of Anthropic's published launch table against GPT-5.5 and Gemini 3.1 Pro. The headline numbers: 80.3% on SWE-bench Pro for agentic coding, 1932 ELO on GDPval-AA knowledge work, and 85.0% on OSWorld-Verified computer use. Some rows carry an asterisk that matters, explained below.

Two rows deserve a business translation. FrontierCode Diamondmeasures whether agent-written code is maintainable, the kind a team can keep working on. Fable 5's 29.3% vs GPT-5.5's 5.7% is the widest gap in the table and the one that matters most if AI-written code ends up in your production systems. GDPval-AA covers professional knowledge work beyond coding; the 1932 ELO comes with best-in-class results on legal (Legal Agent Benchmark: 13.3% vs 2.1% for GPT-5.5) and finance evals per digitalapplied's launch analysis.

One thing the table does not measure at all is language. If your workload runs in Hungarian, Polish or Czech rather than English, read the next section before you trust any of these rows.

Does Claude Fable 5 speak Hungarian (and other languages)?

For practical business use, yes. Claude models are documented as strong across languages, and our experience with Hungarian business text — proposals, contracts, customer correspondence — has been consistently good, including our first Fable 5 tests. What does not exist is a published Hungarian benchmark for the model, so any specific number you see quoted is extrapolation.

The gap is structural. The standard multilingual suite, MMMLU, covers 14 languages, and Hungarian is not among them; neither Anthropic's announcement nor the launch coverage published Hungarian results. The same holds for most mid-sized European languages. In our client work the failure mode is rarely grammar — earlier Claude generations already handled Hungarian morphology and formal business register reliably — but narrow domain terminology: legal, accounting and manufacturing vocabulary, where errors are subtle and expensive.

So the reliable answer comes from a pilot on your own material: 20-30 real documents in the language your business runs on, a defined notion of a correct output, and a measured error rate. The staged rollout makes this nearly free right now — Fable 5 is included on paid Claude plans at no extra cost until mid/late June 2026, after which usage credits apply. If your team is new to working with AI tools, pairing that pilot with structured AI training shortens the evaluation loop and keeps the results comparable across departments.

Pricing: think cost per task, not cost per token

Claude Fable 5 costs $10 per million input tokens and $50 per million output tokens on the API. That is exactly double Claude Opus 4.8 ($5/$25) and less than half of the invitation-only Claude Mythos Preview it replaces (Anthropic).

Two planning caveats before you extrapolate from the price list. First, Fable 5 uses a new tokenizer that produces roughly 30% more tokens for the same content than Opus-tier models, per the API documentation. The effective per-task cost is therefore higher than the per-token ratio suggests. Second, adaptive thinking is always on, with an "effort" parameter from low to max; where you set it changes both quality and cost per task class.

The more useful frame is what a completed task costs. A large run with 100K input and 20K output tokens lands around $2.00 on Fable 5, against roughly $1.10 on GPT-5.5, per digitalapplied's illustration. If that $2 run completes work a senior developer or analyst would spend 2-4 hours on, the comparison is no longer between two API bills. At Budapest senior rates of 25-30 thousand forints (about €63-75) per hour, the labor cost it displaces is 30-50× the model cost. Our AI ROI calculator runs this math on your own numbers, and our AI development cost guide covers what the integration around the model costs.

Worth knowing on billing: requests Fable 5 refuses before producing output are not billed, and a beta "fallbacks" API parameter can automatically retry a refused request on another Claude model, with a "fallback credit" refunding the prompt-cache switching cost. In the Claude app, Fable 5 is included at no extra cost on Pro, Max, Team and Enterprise plans during the staged rollout, from June 9 until mid/late June 2026, after which usage credits apply. That window is the cheapest way to test it on your real work.

Claude Fable 5 vs ChatGPT for business

The short version: Fable 5 is measurably ahead on agentic coding, knowledge work and long-running autonomous tasks, while OpenAI's GPT-5.5 costs half as much per token and stays competitive on routine, high-volume work. For most companies the decision is not a vendor choice; it is an architecture that routes each task type to the model that earns its price.

A fairness note: this is Anthropic's table, and OpenAI measures its own strengths elsewhere. OpenAI reports GPT-5.5 at 85.0% on ARC-AGI-2 and 93.6% on GPQA Diamond, and no Fable 5 numbers were published for either, so the pure reasoning comparison stays incomplete. On cost per task, digitalapplied's illustration puts a 100K-input / 20K-output run at about $2.00 on Fable 5 against roughly $1.10 on GPT-5.5 — on high-volume routine work that gap compounds into a serious monthly line item.

When does each win? Fable 5 earns its double token price where the task is long and autonomous: coding agents, legacy migrations, large document packs, multi-day workflows. Its SWE-bench Pro lead is 21.7 points, and on maintainable code (FrontierCode Diamond) the gap is fivefold, 29.3% vs 5.7%. GPT-5.5 handles classification, short replies and summaries just as well at half the price. Designing and building that per-task routing is a typical AI development engagement rather than a settings toggle. For the ChatGPT side's enterprise detail — plans, GDPR, Hungarian language quality — see our ChatGPT for business guide.

Five places a business will feel the difference

The capability jump shows up in five concrete areas: legacy code modernization, automating software that has no API, whole-archive document analysis, multi-day autonomous back-office agents, and the AI coding tools your developers already use. Each of the five below is grounded in a verified launch-day fact, with the source attached.

1. Legacy modernization drops from quarters to days

That was a codebase-wide migration completed in one day. The pattern transfers: many Hungarian companies run aging PHP, ASP.NET or Java systems where modernization has been postponed for years because it was priced as a multi-month project. With an agent of this class doing the mechanical bulk under engineering supervision, the same work scopes as a few-week sprint. This is the strongest single proof point of the launch, and it is the kind of project our system integration team gets asked about most.

2. Automating software that has no API

Fable 5 scores 85.0% on OSWorld-Verified, the computer-use benchmark where the model operates real desktop software through the screen, the way a person does. No API, no integration endpoint required. A large share of SMEs run invoicing, warehouse or HR tools that expose no interface at all; "our software is too old to automate" has been the standard blocker. A screen-operating agent removes it, opening a service category between classic RPA and full custom integration. If you have a process stuck behind a GUI, this is what process automation looks like in 2026.

3. Whole-business document analysis in one request

The 1M-token default context window fits an entire ERP codebase, a year of supplier contracts, or a complete public-tender documentation pack in a single prompt, with up to 128K tokens of output. For one-off analyses — contract portfolio review, tender preparation, an ERP data-migration audit — you no longer need to build a retrieval pipeline first just to get the material in front of the model. That removes a whole project phase from the budget for this class of work.

The vision side matters here more than the demos suggest. Anthropic calls Fable 5 a new state of the art for vision tasks, demonstrated among other things by playing Pokémon FireRed through a vision-only interface; the business-relevant number is 29.8% on GDPpdf, the visual document reasoning benchmark, against 24.9% for GPT-5.5 (per digitalapplied's comparison). Real company archives are scanned contracts, stamped invoices and PDF exports with broken text layers, and that is exactly the material this benchmark tests.

4. Back-office agents that run for days

Anthropic positions Fable 5 for long-running agentic work: in harnesses like Claude Code it runs for days, plans across stages, delegates to sub-agents, checks its own work, and processes millions of tokens per task. In business terms that is quote generation, order processing and document workflows handled end to end rather than chat-assisted. We mapped the build-vs-buy decision for exactly this in our chatbot vs n8n vs custom agent comparison; Fable 5 raises the ceiling of the custom-agent column without changing the decision logic.

The launch also shipped the supporting infrastructure that multi-day runs need: a memory tool, context compaction and editing, task budgets, and server-side fallbacks to other Claude models. Task budgets in particular answer the first question every finance director asks about autonomous agents: yes, you can cap what a run is allowed to spend before it starts.

5. It is already inside your developers' tools

Claude Fable 5 became generally available in GitHub Copilot on launch day. Early-access vendors published their own results: Cursor called it state of the art on CursorBench and said it "opened up a class of long-horizon problems" previously out of reach; Replit measured it as the top model on its ViBench end-to-end benchmark, building apps in less time with fewer tokens; Figma called it "a clear step forward on agentic coding and prototyping"; Hex reported it as the first model to score 90% on its core analytics benchmark; the legal AI platform Harvey shipped it to customers on day one. If your team uses any of these, the upgrade arrives without a procurement decision — which makes it worth checking what data flows through those tools.

GDPR and EU compliance: the part to read before the pilot

Fable 5 is a "Covered Model" with mandatory 30-day retention of API inputs and outputs, and Anthropic does not offer it under zero-data-retention agreements, per the API documentation. For EU companies this belongs in the data-protection impact assessment before any production rollout.

On the procurement side, the news is better. Fable 5 is available from day one on the Claude API, Amazon Bedrock, Google Vertex AI and Microsoft Foundry. If your company already has a cloud contract and DPA with AWS, Google or Microsoft, you can buy frontier-model access through paperwork your legal team has already reviewed. One honest gap: whether Bedrock or Vertex serve Fable 5 from EU regions had not been verified when this article was published — confirm the region before making a data-residency promise to your own customers.

The EU AI Act adds a second layer: deployers of systems built on general-purpose models carry transparency duties, such as disclosing AI interaction to users where required. Treat this as a due-diligence checklist item alongside the DPIA rather than legal advice; our EU AI Act and GDPR compliance guide walks through the obligations by risk class.

A practical adoption path

Three steps, in order: pilot on your own documents before trusting any benchmark, route tasks by type instead of standardizing on one model, and design for refusals and fallback before production. The companies that skip the third step find out about it from confused users.

Pilot on your own material first

No Hungarian benchmark exists for Fable 5, and your domain language is its own dialect anyway. Pick 20-30 real documents, tickets or contracts, define what a correct output looks like, and run a one-to two-week pilot. The free window in the Claude app (until mid/late June 2026 on paid plans) makes this nearly costless. Measure before you commit; the result is also the evidence your management or your DPO will ask for.

Route tasks by type

The head-to-head above gives you the split. The architecture that wins on cost routes each task class to the cheapest model that clears the quality bar — Fable 5 for agents, migrations and complex document work, cheaper models for volume. Single-vendor setups are simpler to build and consistently more expensive to run.

Design for refusals and fallback

Fable 5 declines some requests and silently answers under 5% of sessions with Opus 4.8. Refused requests are free, and the beta fallbacks parameter automates retries, but someone has to decide what your application does when quality, latency or tone shifts mid-workflow. A DIY integration that ignores this works in the demo and surprises you in month two. This is precisely the unglamorous layer where an experienced integrator earns their fee — and if you want a scoped pilot with measurement built in, tell us about your use case.

Frequently asked questions