Claude Fable 5 for Business: Anthropic’s Mythos-Class Model, the Benchmarks, Pricing and How to Roll It Out (2026)

Claude Fable 5 is Anthropic’s first generally available Mythos-class model — a tier above Opus — and the first model your company can buy that behaves less like a chat assistant and more like an autonomous senior colleague. It stays with a problem for days, plans and validates its own work, and posts benchmark numbers that more than double Opus 4.8 on the hardest coding evaluations. It also costs twice as much per token, ships with safety classifiers that can reroute sensitive requests to Opus 4.8, and requires 30-day data retention. This is the business guide: what Fable 5 actually is, the benchmarks in plain English, pricing and access (Claude, GitHub Copilot, Google Vertex AI), where it pays for itself, the trade-offs you have to govern, and how to roll it out without burning budget.

By Toni Dos Santos, Co-Founder, Spicy Advisory — we help mid-market and enterprise teams actually use the AI tools they’ve bought, tool-agnostically, across the UK and EU.

What is Claude Fable 5?

Claude Fable 5 is Anthropic’s most capable generally available model, released in June 2026 as the first model of its new “Mythos” class — a capability tier that sits above the Opus line. Under the hood it is the same model as Claude Mythos 5, which Anthropic restricts to vetted cyber-defence and scientific partners through its Project Glasswing trusted-access programme. Fable 5 is the version the rest of us get: identical capability, wrapped in expanded safety classifiers that detect high-risk requests — offensive cybersecurity, biology and chemistry, and model distillation — and route them to Claude Opus 4.8 instead, with a notification to the user (Anthropic announcement).

If your team standardised on Opus 4.8 after the May release, the mental model is simple: Opus 4.8 remains the careful, controllable daily flagship — we covered its effort selector and honesty upgrade in our Opus 4.8 business guide — while Fable 5 is the specialist you bring in for the work you’d otherwise staff with a senior hire: multi-day projects, whole-codebase changes, deep analysis with real stakes.

For a video walkthrough of the launch and what the Mythos class means in practice, this overview is a good 101 to circulate internally:

New to Claude as a company? Start with our Claude for companies playbook and the getting-started guide for teams, then come back here for the Fable-5-specific decisions.

How good is Claude Fable 5? The benchmarks, in business terms

Fable 5 is the strongest model Anthropic has tested on software engineering, knowledge work, vision and long-context tasks — and on the hardest evaluations the gap over the previous generation is unusually large. Three numbers tell the story. On SWE-Bench Pro (real-world agentic coding), Fable 5 scores 80.3% against 69.2% for Opus 4.8 and 58.6% for GPT-5.5. On FrontierCode (Diamond) — the hardest 50 tasks of Cognition’s benchmark — it scores 29.3% where Opus 4.8 manages 13.4% and GPT-5.5 just 5.7%: more than double the previous frontier. And on GDPval-AA, which measures real economically valuable knowledge work, it leads with 1932 against 1890 for Opus 4.8 and 1769 for GPT-5.5.

Anthropic reports Mythos 5 and Fable 5 scores within 1–3 percentage points of each other and shows the higher of the two. Starred (*) benchmarks show a larger gap because Fable 5’s safety fallbacks trigger on cybersecurity- and biology-adjacent questions, pulling its score closer to Opus 4.8 there. Source: Anthropic, June 2026.

The most strategically interesting chart from the launch isn’t a leaderboard, though — it’s accuracy versus cost. On FrontierCode Diamond, Fable 5 keeps converting extra compute into extra capability all the way up its effort range, while Opus 4.8 plateaus around 13% and GPT-5.5 stays flat near 5–6% regardless of spend:

Why this matters to a business reader: for the first time, budget is a capability dial on genuinely hard problems. At roughly the same per-task spend (≈$8–10), Fable 5 at high effort scores around 24% where Opus 4.8 peaks at 13.4%. If a task is worth a senior person’s day, paying $20 of compute for a 31% solve rate on frontier-difficulty work is a trade most CFOs will take. And don’t misread the “low” scores on evaluations like the Legal Agent Benchmark: 13.3% sounds modest until you notice GPT-5.5 scores 2.1% and Gemini 3.1 Pro 0.0% — these are tasks designed to be barely solvable, and Fable 5 is six times further along than the next non-Claude frontier model.

How much does Claude Fable 5 cost — and where do you get it?

Fable 5 is priced at $10 per million input tokens and $50 per million output tokens on the Claude API — roughly double Opus 4.8’s $5 / $25. Anthropic is cushioning the landing for subscribers: Pro, Max, Team and seat-based Enterprise plans include Fable 5 at no extra cost until June 22, 2026, after which usage moves to credit-based billing, with Anthropic stating it intends to fold it back into standard subscriptions as capacity allows. If your team is already rationing Opus usage, our guide to protecting Claude usage limits just became more relevant, not less.

Three access routes matter for companies:

• Claude apps and API. Available immediately as claude-fable-5 for API and consumption-based enterprise customers; in the apps it appears in the model picker once enabled for your workspace, and admins can allow or restrict it per workspace or seat.
• GitHub Copilot. Available on Pro+, Max, Business and Enterprise tiers across VS Code, Visual Studio, JetBrains, Xcode, Eclipse, GitHub.com, the Copilot CLI and the cloud agent. Crucially for governance: it’s an admin policy that is off by default, and switching it on is an explicit acknowledgement of Anthropic’s 30-day data-retention requirement. GitHub’s own benchmarks show Fable 5 completing autonomous coding workflows with fewer tool calls and lower token consumption than prior Opus-tier models.
• Google Gemini Enterprise Agent Platform / Vertex AI. Fable 5 ships as a partner model in Model Garden. Enabling it in a GCP project requires consenting to the Advanced AI Safety Addendum, accepting Anthropic’s marketplace terms, enabling the model per region, and turning on prompt-response sharing with Anthropic — then it’s available for agents, computer use, prompt caching, function calling and batch workloads.

The pattern to notice: every enterprise surface gates Fable 5 behind an explicit administrative decision. That’s unusual, and it’s your opening to do the governance work before the model arrives, not after. Our CISO’s guide to enterprise AI security covers the checklist.

What actually changes for day-to-day work

Anthropic’s framing is that Fable 5 can “stay with a problem far longer than any model before it” — operating autonomously for days, coordinating tools, and validating its own work at high reasoning effort. The early-customer evidence says this is not marketing copy.

Software engineering: the migration machine

The headline case study is Stripe, which used Fable 5 to complete a codebase-wide migration across a 50-million-line Ruby codebase in a single day — work that would otherwise have taken a team months. Other early partners report that Fable 5 “one-shots” full applications that previously needed dozens or hundreds of prompts, and anticipates edge cases instead of waiting to be told about them. Combined with the Terminal-Bench (88.0%) and SWE-Bench Pro (80.3%) numbers, the practical translation is: framework upgrades, language migrations and cross-cutting refactors that never made it off the backlog are now projects you can scope in days. For where this sits next to Copilot and plain coding, see when to use Claude, Copilot or code.

Finance, analytics and knowledge work

On Hebbia’s Finance Benchmark for senior-level reasoning, Fable 5 ranks highest among tested models, with the biggest gains in document-based reasoning, chart and table interpretation, and expected-value analysis. Trading firm IMC found it performing at or above their internal senior-analyst benchmarks on factual lookup, conceptual reasoning and root-cause analysis. And analytics platform Hex reports Fable 5 is the first model to clear 90% on its core analytics benchmark of long-running analytical tasks — roughly ten points ahead of Opus 4.8 — while finishing everyday spreadsheet workflows 25–30% faster and in fewer turns. In plain terms: more of your exploratory analysis, reporting and dashboard-building can be delegated to agentic workflows with less babysitting.

Vision: from screenshot to system

Fable 5 extracts precise quantitative data from dense scientific figures and can reconstruct working applications from screenshots alone — capabilities earlier Claude models needed scaffolding and extra tools to approximate. Two office workflows fall out of this immediately: screenshot a legacy dashboard or PDF report and have Fable 5 reverse-engineer the underlying logic into SQL or analytics code; or capture a legacy UI and have it generate a modern front-end and API layer as the start of a modernisation project.

Long-horizon work and memory

With up to 1 million tokens of input, Fable 5 ingests entire repositories or multi-quarter document dumps in one go. More importantly, Anthropic’s experiments show that when paired with persistent file-based memory, Fable 5 tripled its performance over Opus 4.8 on the long-horizon strategy game Slay the Spire, reaching late-game states three times more often — a proxy for the thing businesses actually care about: an agent that accumulates notes, refines hypotheses and keeps multi-day projects coherent instead of starting from zero each session. If your team builds reusable workflows, this is where Claude skills compound.

The trade-offs your rollout has to govern

An honest brief names the costs. Fable 5 has three, and all of them are manageable if you plan for them.

1. Safety fallbacks can interrupt legitimate work

The classifiers that make general availability possible are deliberately conservative. Anthropic acknowledges false positives that can affect legitimate security research and advanced scientific work, particularly in biology and chemistry — when triggered, the request silently downgrades to Opus 4.8 quality (with a notification). Red-team results show the same system blocks meaningful progress on offensive cyber tasks even under common jailbreaks, which is the point. Brief your security and R&D teams that a fallback is expected behaviour, not a bug — and that eligible organisations can apply for Mythos 5 trusted access through Anthropic’s partner programme if their work keeps tripping the filters.

2. The 30-day data-retention requirement

To run those safety systems, Anthropic requires 30-day retention of prompts and outputs for all Mythos-class usage — including Fable 5, on every surface, first- and third-party. The data is used only to defend against complex attacks (such as multi-turn jailbreaks) and to reduce false positives, and is deleted after 30 days in almost all cases — but it is a real departure from the zero-data-retention arrangements many enterprises negotiated for other Claude models. If you operate under strict data-sovereignty or matter-confidentiality rules (legal, health, financial services), classify which workloads can tolerate 30-day retention before enabling the model. Our AI data residency guide for UK enterprises walks through the framework.

3. Twice the token price — but count cost per outcome

At $10 / $50 per million tokens, Fable 5 is the wrong tool for routine summarisation and everyday Q&A — that work belongs on Opus 4.8, Sonnet or Haiku. But early adopters consistently report Fable 5 finishing complex tasks in fewer turns and fewer total tokens, sometimes compressing months of work into days. The metric that matters is cost per completed unit of work, not cost per token: a $40 agent run that replaces a week of senior engineering time is the cheapest labour you will buy this year. The discipline is routing — which is a governance and training problem, the kind we cover in measuring AI training ROI.

Fable 5 or Opus 4.8: which model for which job?

Most companies should run both. Here’s the routing table we’re recommending to clients:

For the wider vendor question — Claude vs ChatGPT vs Gemini as a platform decision — our Claude vs ChatGPT for business comparison still holds; Fable 5 simply raises the ceiling on the Claude side.

How to roll out Fable 5 without burning budget (or trust)

The model is the easy part. Here’s the five-step sequence we run with clients:

1. Pick two or three high-value, long-horizon workloads — a stalled migration, a recurring senior-analysis bottleneck, a modernisation project. Fable 5 is justified by compressing weeks into days, not by marginally better chat.
2. Enable it through a governed surface. Use the Copilot admin policy or workspace-level controls so access maps to the teams that own those workloads. The off-by-default posture is a feature — keep it.
3. Clear the retention question first. Classify which data can tolerate 30-day retention; route everything else to Opus 4.8. Write it down — one paragraph in your AI policy beats a quarter of Slack arguments.
4. Brief users on the two surprises: safety fallbacks to Opus 4.8 (expected, notified, not a bug) and the temporary inclusion in subscriptions until June 22, 2026 (after which usage is credit-based — budget accordingly).
5. Measure cost per outcome from week one. Tokens per completed task, interventions per agent run, calendar time saved. That’s the evidence that survives the CFO conversation.

Where Spicy Advisory fits

We help teams get from “we enabled the new model” to “it changed our throughput”: workload selection, the retention and governance policy, workflow redesign on real deliverables, and the enablement that closes the 91/21 gap. We’re tool-agnostic and work in English or French, in person and hybrid across the UK and EU — see our Claude training for teams, AI training in the UK and Spicy Advisory for Enterprise. Smaller team? The Claude for small business guide is the gentler on-ramp, and our 2026 launches roundup keeps the wider model race in perspective.