Reference — updated July 27, 2026
Claude Fable 5 glossary
Anthropic's Mythos-class launch brought a vocabulary that did not exist eight weeks ago: covered models, safeguard fallback, effort levels, refusal categories. Each entry below is self-contained, dated, and written for someone who has to make a decision, not pass a quiz.
The short version
Five terms carry most of the weight. Mythos-class is the capability tier Fable 5 belongs to, and it is what triggers the mandatory 30-day data retention. Safeguard fallback is why a Fable 5 request can quietly come back answered by Opus 4.8. Effort is the one tuning knob the API still gives you. Usage credits are how a flat subscription turns metered. Zero data retention is the compliance line Fable 5 cannot cross. Everything else in this list is detail hanging off those five.
- Adaptive thinking
- Fable 5 reasons before answering on every request and you cannot turn that off — sending thinking with type disabled returns a 400, as does the old budget_tokens form. The raw chain of thought is never returned; by default thinking blocks come back empty, and you have to ask for a summarized display to show users anything. Two practical consequences: thinking tokens count against max_tokens alongside the visible answer, so budget headroom generously, and a streaming UI with the default omitted display looks frozen during the thinking phase.
- Batch API
- An asynchronous endpoint that takes 50% off both input and output tokens in exchange for giving up latency guarantees, with results returned within 24 hours. The discount stacks with prompt caching, so cached input inside a batch lands around 95% below list. The rule of thumb is simple: if no human is sitting there waiting for the answer — evaluation runs, backfills, nightly summarization, bulk classification — routing it through batch halves the bill for no work beyond changing the call site.
- Context window
- How much text the model can consider in a single request. Fable 5 accepts up to 1M input tokens and produces up to 128K output tokens per response. The subtlety worth internalizing is that max_tokens caps thinking and visible output together, so a run at xhigh or max effort can exhaust its budget reasoning before it writes anything — Anthropic's guidance is to start around 64K at those levels. A large window also is not free: every token you resend is billed unless it comes from a cache hit.
- CursorBench
- Cursor's own coding-agent evaluation, cited by Anthropic at version 3.2 in the Opus 5 launch materials. The reported result is that Opus 5 at max effort lands within 0.5% of Fable 5's peak score at roughly half the cost per task. Because it measures work inside a real editor loop rather than a research harness, it is the benchmark closest to how most developers actually experience these models — and here Fable 5's remaining lead is small enough that price usually decides.
- DeepSeek V4
- DeepSeek's open-weight flagship, released with the weights freely downloadable. The Pro tier is priced at $0.435 per million input tokens and $0.87 per million output tokens, with Flash at $0.14/$0.28. DeepSeek reports 55.4% on SWE-bench Pro — a vendor-reported figure, not independently confirmed. It is one of the main open-weight contrasts to Fable 5's closed API.
- The effort parameter
- A request field, output_config.effort, taking low, medium, high, xhigh or max, defaulting to high. It is the primary control for trading intelligence against latency and cost on Fable 5, and it largely replaces the sampling knobs — passing a non-default temperature, top_p or top_k returns a 400 error. Anthropic's own guidance is to start at high and reserve xhigh and max for capability-sensitive coding and agentic runs; lower settings on Fable 5 often beat xhigh on earlier models, which means the expensive model at low effort can be the cheap option.
- Export control
- U.S. export-control restrictions that affect how Chinese open-weight models are distributed and accessed. This is directly relevant to this site's differentiator timeline: Fable 5 was offline from June 12 to July 1 due to a hosting-related disruption and was restored on July 1. Chinese models such as DeepSeek, Qwen and GLM remain broadly available, but the regulatory picture is in flux and availability can change without notice.
- Claude Fable 5
- Anthropic's publicly available Mythos-class model, API ID claude-fable-5, released June 9, 2026. It prices at $10 per million input tokens and $50 per million output tokens, carries a 1M-token context window and up to 128K output tokens per response, and scores 80.3% on SWE-Bench Pro. It is the most expensive model on Anthropic's public list — roughly double Opus 5 and five times Sonnet 5's introductory rate — so the practical question is rarely whether it is capable but whether the task justifies the multiplier.
- Frontier-Bench v0.1
- An agentic terminal-coding benchmark Anthropic used as the headline metric at the Opus 5 launch: Opus 5 at 43.3%, Fable 5 at 33.7%, Opus 4.8 at 18.7%. Anthropic reports these from an internal run on the mini-SWE-agent harness with a GKE backend, taking mean reward over five attempts per task. One footnote changes how you read it: Opus 4.8 served as the fallback on safety-classifier refusals during that run, so the figures describe a configured agent policy rather than a bare model endpoint.
- GDPval
- An OpenAI benchmark measuring models on real economically valuable knowledge work: 1,320 tasks across 44 occupations drawn from the top nine sectors of US GDP, each built from an actual work product by a professional with roughly 14 years of experience, with 220 tasks open-sourced as a gold subset. Anthropic reports an Elo-scored variant, GDPval-AA, in its launch tables. It is the closest thing the field has to asking whether a model can do a real job, which is also why it is graded by human comparison rather than a unit test.
- Project Glasswing
- Anthropic's defensive-cybersecurity program, announced April 7, 2026 with launch partners including AWS, Apple, Broadcom, Cisco, CrowdStrike, Google, JPMorganChase, the Linux Foundation, Microsoft, NVIDIA and Palo Alto Networks, and since expanded to roughly 150 organizations across about 15 countries. Participants use the unrestricted Mythos model against their own codebases and infrastructure. There is no public application form — the program is invitation-only and NDA-bound — so if you are evaluating whether to wait for Mythos access, the honest answer for almost every team is that it is not a path that opens.
- GLM 5.2
- Zhipu's open-weight model released under the MIT license, priced at roughly $1.40 per million input tokens and $4.40 per million output tokens. Zhipu reports 62.1% on SWE-bench Pro — a vendor-reported figure, not independently confirmed.
- Kimi K3
- Moonshot AI's 2.8T-parameter flagship, released July 27, 2026 with open weights. Pricing is $3 per million input tokens and $15 per million output tokens. It is one of the most prominent open-weight models competing with Fable 5.
- Claude Mythos 5
- The same underlying model as Fable 5 with its cybersecurity and biology safety restrictions removed. It is not a product you can buy at any price; access runs only through Project Glasswing, to vetted cyber defenders and critical-infrastructure operators. This matters when you read benchmark tables: rows attributed to Mythos 5 describe a model you cannot call, and on safeguarded topics Fable 5 behaves closer to Opus 4.8 than to the Mythos numbers.
- Mythos-class
- Anthropic's label for its frontier capability tier — models capable enough at cybersecurity and biology work that the company treats them differently from the rest of the Claude line. Claude Fable 5, released June 9, 2026, is the first Mythos-class model made publicly available. The label is not marketing: it is what pulls in always-on safety classifiers, a mandatory 30-day data retention window, and exclusion from zero-data-retention agreements. If a model is Mythos-class, assume the compliance rules around it are stricter than for Opus or Sonnet.
- Open weights
- Models whose trained weights are freely downloadable, letting anyone self-host or fine-tune them. Examples in this generation include GLM 5.2 (MIT), DeepSeek and Kimi K3. This contrasts with Fable 5, which is only available through Anthropic's closed API.
- OSWorld
- A computer-use benchmark that runs agents against a real Ubuntu desktop rather than a simulator: 369 tasks spanning browsers, office suites, file management, terminals and image editors, each with a scripted initial state and a programmatic success check against the file system. OSWorld 2.0 adds a set of longer, multi-step workflow tasks. If your plan involves an agent driving a GUI instead of an API, this is the benchmark family that predicts whether that works — and scores here remain far below what a human achieves on the same tasks.
- Prompt caching
- Marking a stable prefix of your prompt with cache_control so repeat requests do not pay full price to reprocess it. Cache reads cost 0.1x the base input price — a 90% discount — while writes cost 1.25x for the default five-minute TTL or 2x for the one-hour TTL, so the five-minute tier pays for itself on the second hit. At Fable 5's $10 per million input tokens, a long system prompt or a large spec resent on every turn is exactly the workload where caching decides whether the project is affordable.
- Qwen 3.8 Max
- Alibaba's flagship, priced at roughly $2 per million input tokens and $6 per million output tokens, with a 1M-token context window. As of August 2026 it has been announced but not yet released.
- reasoning_extraction
- One of the refusal categories, alongside cyber and bio. It fires when a prompt tries to make the model reproduce its internal reasoning as response text — an anti-distillation defense applied per request. This catches a very common pattern by accident: agent harnesses that instruct the model to echo its thought process verbatim so a downstream step can read it. If you need process visibility, request summarized thinking blocks instead of asking the model to narrate its reasoning into the answer.
- stop_reason refusal
- The API signal for a classifier block. The Messages API returns HTTP 200 — a success, not an error — with stop_reason set to refusal and a stop_details object naming the category. Any error handling built purely on HTTP status codes will wave these through as normal completions, which in a pipeline means an empty or wrong link in the chain with no alarm attached. Branch on stop_reason explicitly on every Fable 5 call, and treat a null stop_details as a generic refusal rather than assuming a category.
- Safeguard fallback
- When one of Fable 5's safety classifiers fires, Anthropic does not hand you a flat refusal — the request is served by Claude Opus 4.8 instead. Anthropic describes this as affecting under 5% of conversations on average, and it is why a Fable 5 workload can return noticeably weaker answers on security or life-sciences topics without anything looking broken. In the Claude apps the switch is shown as a notice; on the API you have to detect it yourself. Log it as its own event or you will misread the quality drop as model drift.
- SWE-Bench Pro
- Scale AI's contamination-resistant successor to SWE-Bench Verified: 1,865 problems across 41 actively maintained repositories, split into public, held-out and commercial sets. The public repositories use strong copyleft licenses and the commercial ones are private, both chosen to keep the code out of training corpora. Tasks are long-horizon and typically span multiple files. Fable 5 scores 80.3%, the highest published public-model figure. Compare scores only within the same harness — the number moves substantially with scaffolding.
- Tokenizer shift
- Starting with Fable 5, Anthropic changed the tokenizer, and the same text now counts roughly 30% more tokens than it did on older models. This breaks the arithmetic people carry over from previous invoices: comparing list prices alone understates the real gap, because you are also buying more tokens for identical input. Before you sign off on a migration budget, run a representative sample through the new model and read the actual usage figures rather than extrapolating from last quarter's bill.
- Usage credits
- A pay-as-you-go balance that sits alongside a Claude subscription and bills at API rates once plan-included usage runs out. Since July 20, 2026 this is the only route to Fable 5 on Pro and Team Standard, cushioned by a one-time $100 credit; on Max and Team Premium, credits take over after Fable 5 consumes 50% of the weekly limit. Practically, credits are how a flat monthly subscription becomes a variable bill, so any heavy Fable 5 plan needs a dollar budget, not just a plan tier.
- Zero data retention (ZDR)
- A contractual arrangement under which Anthropic does not store your prompts or responses at rest after the API response is returned. It is not a toggle: it is negotiated with sales, approved per organization, and does not automatically extend to other organizations under the same account. Fable 5 is a covered model requiring 30-day retention for safety monitoring, so it cannot be used from a ZDR-enabled organization at all. For teams with a hard ZDR requirement, that is a blocker no amount of configuration removes.
Turn the vocabulary into a number
Knowing what effort and prompt caching mean is step one. Step two is seeing what your own workload costs at $10/$50 per million tokens, with caching and batch applied — and what the same job costs on Opus 5 or Sonnet 5.