§1What actually shipped
On 9 June Anthropic released Claude Fable 5, the first publicly available Mythos-class model. If you followed the "Mythos Preview" announcements earlier this year, this is that model with safeguards bolted on: Mythos 5 itself stays behind trusted-access controls, and Fable 5 is the version the rest of us get, with safety classifiers wrapped round offensive cyber, life-sciences content and attempts to extract its own reasoning. Declined requests come back as a refusal stop reason, and Anthropic's own advice is to configure automatic fallback to Opus.
The commercial shape matters more than the benchmark chart. API pricing is $10 input and $50 output per million tokens, roughly double Opus. Subscription plans include Fable 5 at no extra cost only until 22 June, counting at 2x usage weighting; from 23 June it moves to prepaid usage credits. That makes Fable 5 the first frontier model to launch effectively outside the subscription bundle, and the next fortnight a free evaluation window with a hard close.
§2How to actually run it
In Claude Code it is claude --model claude-fable-5, or /model claude-fable-5 mid-session if it has not reached your picker yet. The context window is 1M tokens in, 128K out, with adaptive thinking on by default.
The single most useful operational fact, straight from Anthropic's prompting guide: effort is the primary dial, not the model picker. Default to high, reserve xhigh for the genuinely capability-bound work, and note that low and medium on Fable 5 often beat xhigh on prior models. The second most useful fact: turns get long. Single requests can run for many minutes and autonomous runs for hours, so harnesses built around snappy request-response will need their timeouts, streaming and progress reporting rethought before you migrate. We learnt this one the hard way: our model router was silently losing long generations to client timeouts that had never mattered on faster models, and the failover logic never fired because a timeout was not the error class it was watching for.
§3The prompting shifts worth stealing
We read the official guidance, the early write-ups and the inevitable YouTube explainers so you do not have to. Stripped of hype, six changes earn their place:
- Give goals, not steps. The model is built for end-to-end work. Hand it the outcome and the success criteria, and pick a task harder than what you would have given the previous model. Micro-stepping it wastes exactly the capability you are paying double for.
- Tell it when to stop planning. At high effort it can over-gather and over-deliberate. One line fixes it: "when you have enough information to act, act."
- Ground its progress claims. On long runs, instruct it to audit every status claim against an actual tool result before reporting. Anthropic says this nearly eliminated fabricated progress reports in testing. Any builder who has had an agent declare victory over work it never did should tattoo this one somewhere visible.
- Use fresh-context verifier subagents. Separate verification agents outperform self-critique. Build the check into the run, on an interval, against the spec.
- Give it a memory file. One lesson per file, corrections and confirmed approaches alike. The model performs measurably better when it can consult its own past mistakes.
- Delete your old prompts. Skills and system prompts written for prior models are often too prescriptive now and actively degrade output. And never ask it to echo its reasoning in the response: that trips the reasoning-extraction classifier and bounces the request entirely.
§4We benchmarked it before writing this
We ran Fable 5 through our own router on one code task with 12 hidden assertions and one agentic loop of dependent tool calls (ledger lookups, FX rates, a calculator, a late-fee policy), against two baselines. Against Opus 4.7, our production default, quality was a dead heat: both passed everything, both landed the identical exact figure, and Opus was substantially faster and roughly half the cost. On that evidence alone you would shrug and keep your money.
Then we re-ran the baseline Anthropic itself compares against: Opus 4.8. Both models still pass the code task. The agentic loop is where it splits. Fable 5 returned the exact total, £6,118.90. Opus 4.8 returned £6,149.90, a wrong number delivered without hesitation, with one malformed tool call per run, and it reproduced that same wrong answer across three repeat runs. Opus stayed faster (8.5s vs 33.1s on the loop) and about half the cost. So the trade, on our small sample: Fable 5 buys correctness on multi-step tool work, at double the price and roughly four times the latency.
That is a handful of tasks, not a benchmark suite, so treat it as a smoke test rather than a verdict. But the shape of the result matches Anthropic's own framing with unusual precision: the gap does not show up in single-shot code, it shows up in the agentic loop, which is exactly the workload the model is sold on. If your evaluation consists of re-running tasks your current model already solves, you will conclude Fable 5 is a slower, dearer Opus, and you will have measured the wrong thing.
§5The 30-day catch
Here is the part the launch coverage and the twelve-minute hype videos do not lead with. With Fable 5 and Mythos 5, Anthropic mandates 30-day retention on all Mythos-class traffic, on every surface, first- and third-party alike, and the policy applies even where an enterprise previously held a zero-retention agreement. The data is not used for training, human access is logged, and deletion after 30 days is the stated norm. The rationale is genuinely defensible: attacks like best-of-N jailbreaking only become visible across many requests, and you cannot detect a campaign you refuse to remember.
Defensible is not the same as free. If you serve clients under contracts that promise zero retention, or you operate in a regulated setting where your processor terms say prompts are not stored, then Fable 5 is off your menu regardless of what the benchmarks say, and so is every "99% cheaper" frontier alternative whose endpoints fail the same test. We verified that last part ourselves this week and could not reach a single no-retention route to the cheapest new competitor either. The sovereignty gap between what frontier models can do and what regulated workloads are allowed to touch widened this week, on both ends of the price range. Our hedge, unchanged: a local, self-hosted tier for the workloads where the prompts must never leave the building.
§6What builders will get wrong
Three predictable mistakes. First, making it the default model: at double the price and slower turns, undirected Fable 5 is an expensive way to do work Opus does in half the time. Second, keeping the old micro-step prompting style, which buys the premium and then refuses the product. Third, and worst, wiring client workloads into it without reading the retention page, then discovering the data-processing agreement they signed says something their new model provider no longer honours.
The stealable step for this week: while it still costs nothing extra on a subscription, point Fable 5 at the hardest unsolved item in your backlog, as a goal with success criteria and a verifier loop, and judge it on that. And before 23 June, read your own processing agreements next to the retention policy. The capability question answers itself in an afternoon. The contractual one is the one that bites in month three. Which of your workloads could you actually move?