Claude Fable 5 is free for 4 more days. Here is the full guide
Anthropic’s strongest model ever is included in paid plans until July 12. The best practices, the copy-paste prompts, and the two patterns that keep 90%+ of its power once the bill starts.
Anthropic shipped its most intelligent model ever, a government switched it off, and now it is free for everyone on a paid plan.
Until Sunday.
Claude Fable 5 is the first model in the new Mythos class, a tier above Opus. It is included on Pro, Max, and Team plans through July 12 at 11:59 PM PT, up to half your weekly limit. After that, it moves to usage credits at $10 per million input tokens and $50 per million output, double Opus 4.8.
Simon Willison called it “something of a beast” and said “the challenge is finding tasks that it can’t do.” He also burned $110 of API credit in a day, which tells you why this guide has two halves: how to use the free window, and how to keep the power once it costs money.
I read the docs, the prompting guide, and the cookbooks so you can skip them. Here is everything, free.
What Fable 5 actually is
▫️ A tier jump, over an increment. On SWE-bench Pro it scores 80.3%, with Opus 4.8 at 69.2% and GPT-5.5 at 58.6%. On SWE-bench Verified it holds the #1 slot near 95%. On Terminal-Bench, 88%. The lead grows as tasks get longer and harder, which is the whole point of the Mythos class: built for agentic, long-horizon work with low error compounding.
▫️ 1M-token context, 128K output, zero long-context surcharge. A 900K-token request bills at the same per-token rate as a 9K one. Feed it the whole data room, the whole codebase, the whole contract stack.
▫️ The same model as Mythos 5, with guardrails. Mythos 5 runs unblocked for vetted organizations; Fable 5 ships with safety classifiers. That difference is why a US export order switched both off worldwide on June 12, the first frontier model shut down by regulators, before returning July 1 with a stronger classifier. Around 1 in 20 sessions gets silently rerouted to Opus 4.8 when the classifier fires, worth knowing before you blame your prompt.
▫️ The proof is public. Stripe ran a codebase-wide migration inside a 50-million-line Ruby codebase in one day, work scoped at two months for a team. One builder had it hand-code a GPU kernel in 2.5 hours for an 18.7x speedup. Andrej Karpathy called a community demo “incredible.”
How to prompt it (this part changed)
Fable 5 punishes Opus-era habits. Anthropic’s own guide frames everything as a delta, and the best-practices playbook needs three updates:
Describe the outcome, skip the steps. Instruction-following is now strong enough that step-lists become handcuffs the model obeys even when they hurt. Old CLAUDE.md files full of “always do X before Y” now work against you. Rewrite them.
Thinking is always on, and you steer with effort. There is zero off switch, temperature is gone, and asking it to “show your reasoning step by step” can trigger a refusal category. Set
/effortinstead: high by default, xhigh for the hardest work, low for routine. Lower effort on Fable 5 often beats max effort on older models.Front-load everything. It skips clarifying questions by default. Give it the goal, the audience, the constraints, and the definition of done in message one, the core move of context engineering. My favorite framing from the official guide:
“I’m working on [X] for [who]. They need [what it enables]. With that in mind: [request].”
Four more copy-paste lines that earn their place in every long run:
“Before reporting progress, audit each claim against a tool result from this session.”
That one nearly eliminated fabricated status reports in Anthropic’s testing, the failure the reliability playbook exists for.
“When the user is describing a problem or thinking out loud rather than requesting a change, the deliverable is your assessment. Report findings and stop.”
“When you have enough information to act, act. Give a recommendation, rather than an exhaustive survey.”
“Don’t add features, refactor, or introduce abstractions beyond what the task requires.”
Boundaries, decisiveness, and restraint: the three behaviors a stronger model needs spelled out.
What to run during the free window
Four days. Spend them on work where a wrong answer is expensive, since that is where the tier jump pays:
▫️ The refactor you have postponed. Multi-file, cross-cutting, the one that scares you. In Claude Code, switch with /model fable (v2.1.170+), and pair it with a /goal loop so it grinds until tests pass.
▫️ Whole-document analysis. A contract plus its exhibits, an earnings narrative against its own tables, a full competitor teardown. The 1M window is the feature.
▫️ A head-to-head test. Same hard task, two chats, Fable 5 versus Opus 4.8. You learn in twenty minutes whether the delta matters for your work, before you ever pay for it.
▫️ Durable assets over one-off outputs. Have it write reusable Skills, a memory file, a CLAUDE.md audit, or your agent architecture. Outputs you keep after the window closes.
▫️ An overnight run. The Mythos class was built for long-horizon autonomy, the pattern behind the Autoresearch Playbook and every serious loop.
Skip it for quick emails and simple lookups. It is slower to first token and counts roughly double against your usage limits, so routine work stays on Sonnet.
The two patterns that keep the power after July 12
Here is the part most people will miss, and it comes straight from Anthropic’s own numbers. You keep 90%+ of Fable 5 while paying half or less, by changing where it sits in your stack.
Pattern 1: Fable 5 as the advisor. A cheaper executor (Sonnet 5) does the work and calls Fable 5 for guidance when stuck, about once per task.
The advisor pattern: the cheap hands type, the expensive brain steers, about once per task. On SWE-bench Pro, this combo hits ~92% of Fable 5’s score at ~63% of the price, because most tokens bill at the executor’s rate. It ships as a first-class advisor tool in the API, and the pattern generalizes: the expensive brain steers, the cheap hands type.
The advisor math on SWE-bench Pro: most of Fable 5's accuracy, well under Fable 5's price. Pattern 2: Fable 5 as the orchestrator. Flip it. Fable 5 plans and delegates to Sonnet 5 workers through Claude Managed Agents, and the token-heavy grunt work bills at the worker rate.
The orchestrator pattern: Fable 5 plans once, Sonnet 5 workers grind in parallel. On BrowseComp, this hits 96% of Fable 5’s performance at 46% of the cost. Anthropic’s own cookbook example makes it concrete: verifying 20 facts across the largest US national parks cost $1.61 in 194 seconds with the split team, versus $4.00 and 608 seconds for Fable 5 alone. Cheaper and three times faster.
The rule of thumb: advisor for coding and agent tasks where a cheap model owns the workload, orchestrator for research and fan-out where the win is keeping bulk tokens away from the expensive model. Both stack with prompt caching (cache reads at a tenth of the input rate) and the Batch API (50% off), the levers from the token-cost playbook.
One warning for Claude Code users: subagents inherit the session model by default. A Fable 5 session with an unpinned subagent fleet bills everything at Fable rates. Pin your workers (model: sonnet or haiku) before you walk away.
The bigger picture
Anthropic is running the classic land-and-price play: give everyone a taste of the frontier, then charge for it, the same ladder logic as every pricing journey. And it is working. Anthropic already passed OpenAI in enterprise API revenue, the Mythos tier gives it a premium shelf, and the free window is the demand test, in a quarter where capital keeps flooding the AI build.
For you, the play is simpler. Test it on your hardest work this week while it costs zero. Decide with your own evidence whether the tier jump moves your metric. And if it does, run it as an advisor or an orchestrator rather than a default, because the one-person operations that win this era route intelligence like a resource, exactly the discipline behind all 100 agent ideas worth building.
Four days. The frontier is on the house.
TL;DR
Claude Fable 5, the first Mythos-class model above Opus, is included on paid plans through July 12 at 11:59 PM PT, then moves to credits at $10/$50 per million tokens. It leads SWE-bench Pro at 80.3% (Opus 4.8: 69.2%, GPT-5.5: 58.6%) with a 1M context window and 128K output. Prompt it differently: outcomes over step-lists, effort levels instead of temperature, full context up front, and the four copy-paste lines above for long runs. Spend the free window on hard, expensive-if-wrong work: the postponed refactor, whole-document analysis, an overnight loop, a head-to-head against Opus. After July 12, keep 90%+ of the power for half the price: Sonnet 5 executor with a Fable 5 advisor (~92% of the score at ~63% of the cost) or a Fable 5 orchestrator with Sonnet 5 workers (96% at 46%). Pin your subagent models, cache aggressively, and route intelligence like a resource.
Keep reading
Master Claude
▫️ Claude best practices: the power-user guide
▫️ The context engineering guide
▫️ The Loop Library: 12 Claude Code loop recipes
Build with agents
▫️ The Claude managed agents guide
▫️ The AI agent reliability playbook
▫️ The token-cost optimization playbook
The business of AI
▫️ Anthropic passed OpenAI at $30B ARR
This is all very nice but can you explain the risks to small biz and solopreneurs from Anthropic having virtually no humans dealing with customer queries. Billing problems or anything out of the ordinary and the robo 'customer service' tier are useless and will keep you busy but going nowhere until forever.