Claude Fable 5 explained: chat, Cowork, agents and code — and the 22 June deadline

What’s the news — and the four audiences it lands on

Anthropic launched Claude Fable 5 on 9 June 2026, a model it describes as exceeding every Claude it has ever made generally available. The company reports state-of-the-art scores on most capability benchmarks, with the lead over its prior models widening as tasks get longer. A new safety layer routes queries on high-risk topics — chiefly cybersecurity and research biology — to the previous-generation Claude Opus 4.8, triggering in fewer than 5% of sessions on average (Anthropic).

The release also has a hard commercial edge. Fable 5 is included at no extra cost on Claude Pro, Max, Team and seat-based Enterprise plans from 9 June to 22 June 2026. From 23 June, it leaves plan limits and is billed via usage credits — a top-up balance you switch on in Settings > Usage that lets you keep working at API rates once you hit your plan’s included limits, with a monthly spend cap and a $2,000 daily cap. After the window, list pricing is $10 per million input tokens and $50 per million output tokens, with batch calls at half price and a 90% discount on cached input (Anthropic; Claude Help Center; Abhishek Gautam).

80.3%Claude Fable 5’s score on SWE-Bench Pro, against 58.6% for GPT-5.5 — the largest agentic-coding gap of any frontier release this year.

What that means depends on which door you walk through:

• Standard chat subscribers (Pro/Max web). In Claude.ai, Fable 5 replaces whatever you’d have used on long, messy tasks — drafting a board paper, summarising a contract bundle, untangling a VAT query. Simon Willison reports the model has access to a full container environment that can install packages and clone repositories, and used it to upgrade his MicroPython-in-WASM sandbox into a working cpython_wasm wheel in a single session (Simon Willison).
• Cowork users. Claude Cowork is Anthropic’s desktop agentic workspace for non-developers — think of it as a research assistant that lives in a folder on your machine and can read, write and reorganise files across documents, spreadsheets and PDFs. Fable 5 is available in Cowork during the free window, which is the cheapest way to test it on a knowledge-work job that previously ate a Tuesday afternoon (Simon Willison).
• Agent builders and automation runners. If you wire Claude into n8n, Zapier, LangChain, CrewAI or your own scripts, Fable 5 is on the same day-one list as Bedrock, Vertex AI, Microsoft Foundry and GitHub Copilot. The Copilot policy is off by default for Business and Enterprise, and enabling it requires accepting a 30-day data retention clause tied to the safety classifier (Abhishek Gautam).
• Developers using Claude Code. Both the CLI and the web version of Claude Code ship with Fable 5. Willison used it to add a human-in-the-loop pause-and-resume feature to Datasette Agent in a day, with the model going on to ship supporting changes in his LLM library unprompted — work he says would otherwise have taken days (Simon Willison).

What to do before 22 June — and how to budget after

The free window is two weeks long, and the cheapest way through it is one real long-horizon task rather than twenty chat prompts. Pick something that already takes a person half a day — a vendor onboarding, a contract redline, a spreadsheet reconciliation — and let Fable 5 chew on it. Track whether the safety fallback fires (it tells you when it does), and measure tokens per completed task rather than per request: Anthropic claims Fable 5 finishes work in fewer turns, so a higher unit price can net out cheaper (Abhishek Gautam).

For API and agent-builder use, the number that changes the math is the 90% prompt-caching discount. Agentic workloads resend the same large system prompt, repository map and tool definitions on every turn; with caching, repeated input drops from $10 per million to roughly $1, which puts a long-running Fable 5 session close to uncached Opus pricing. Run your real traffic shape through the math before deciding the model is too expensive — see our usage-based pricing guide for the framework.

After 23 June, Fable 5 calls on Pro/Max/Team/seat-Enterprise consume usage credits. The pattern that works: enable credits, set a conservative monthly cap, log your first week, then raise or lower it. Most firms we’ve spoken to find the upper bound they actually need is far below what they feared — see how small teams track agent spend for the practical rig.

Two caveats worth your attention

Two asterisks. First, the headline cyber and biology scores in the launch materials were measured on Mythos 5, the restricted variant reserved for Project Glasswing partners and the US government; public Fable 5 routes those queries to Opus 4.8 and scored 0% on offensive cyber tasks in blocking mode (Abhishek Gautam). For a typical UK small firm — most of you are doing knowledge work, marketing, accounts and code review — the published coding and reasoning numbers are the ones you’ll actually see.

Second, this is a frontier model and the bill can move quickly. Simon Willison got through $110.42 worth of Fable 5 tokens in a single day — more than his entire $100/month Max subscription — all absorbed by the free window (Simon Willison). From 23 June, a day like that is a real bill. If you’re a sole trader on a £20 plan, set a spend cap before you fall in love.

The takeaway for a UK small firm

Use the fortnight. Pick one messy job, run Fable 5 on it, log what you actually spend, and set a usage-credit cap before 23 June. If it doesn’t earn its keep, you lose an afternoon; if it does, you’ve found a multiplier that pays for the rest of the year.