Claude Fable 5 (Anthropic)

Status (2026-06-16): Very new — released 2026-06-09, one week old at time of writing. Core specs (model ID, pricing, context, safeguards) are confirmed from Anthropic's own announcement + API docs. Several capability claims are first-party or third-party and are flagged inline. Numbers in this article are quoted, not invented — where a figure is missing or contested, it says so.

What it is

Claude Fable 5 (model id claude-fable-5) is Anthropic's most capable publicly available model as of June 2026. It is the first generally-released member of Anthropic's new top capability tier, internally called the Mythos class — a tier above the Opus class.

There are two models in the launch, sharing the same weights/capabilities:

Model	API model ID	Who gets it
Claude Fable 5	`claude-fable-5`	Generally available. Ships with safety classifiers that can refuse requests.
Claude Mythos 5	`claude-mythos-5`	Limited release only, via Project Glasswing (approved enterprise / select biology researchers). No safety classifiers. Successor to Claude Mythos Preview.

The key framing from Anthropic: Fable 5 is Mythos 5, but made safe for the public by bolting on refusal classifiers in high-risk domains. Where Fable refuses, the request falls back to Claude Opus 4.8 to produce a safe answer. So Fable 5 is best understood as "Mythos-class intelligence behind a safety gate, with Opus 4.8 as the fallback responder."

This sits alongside — not replacing — the existing Claude 4.X family (Opus 4.8, plus Sonnet / Haiku at the 4.x generation). Opus/Sonnet/Haiku Messages-API behaviour is explicitly unchanged; the new behaviours below are Fable/Mythos-only.

Why it matters

A new capability tier went public. Anthropic had been signalling that its most powerful models were too dangerous to release broadly; Fable 5 is the compromise — Mythos-class capability shipped to everyone, gated by classifiers. TechCrunch framed the release as coming "days after warning AI is getting too dangerous."
It topped independent rankings on day one. Per Artificial Analysis (analysis dated 2026-06-10), Fable 5 launched at #1 on the Intelligence Index with a score of 64.9, ~5 points ahead of the closest non-Anthropic model (GPT-5.5), and set the highest score on 5 of the 10 underlying benchmarks. (See caveats — this is one aggregator's methodology, one day after launch.)
It's built for long-horizon agentic work. Anthropic positions it for "large migrations, complex implementations, and multi-day autonomous sessions" — exactly the agentic-RL / long-horizon coding regime. Relevant to OpenAlice's autonomous repo→PR ambitions.
It introduced a refusal-as-success API contract. Refusals return stop_reason: "refusal" as an HTTP 200, not an error, with server-side / client-side / manual fallback paths and "fallback credit" billing. This is a real integration change, not just a bigger model — see model routing.

Capabilities & positioning (confirmed specs)

Shared specs (Fable 5 == Mythos 5):

Context window: 1M tokens (default).
Max output: up to 128k tokens per request.
Pricing: $10 / M input tokens, $50 / M output tokens. 90% input-token discount with prompt caching. (Anthropic notes this is "less than half the price of Claude Mythos Preview"; TechCrunch notes it is roughly double Opus 4.8's rate.)
Data retention: 30-day retention, no zero-data-retention option — both are designated "Covered Models."
Thinking: Adaptive thinking is always on — it's the only mode; thinking: {"type":"disabled"} is unsupported. Depth is controlled via the effort parameter (cf. test-time compute & reasoning).
Raw chain-of-thought is never returned — only "summarized" or "omitted" (default) thinking blocks.
Supported at launch: effort, task budgets (beta), the memory tool, code execution, programmatic tool calling, context editing / tool-result clearing (beta), compaction, vision.

Safeguards / fallback architecture (Fable 5 only): Three classifier domains can decline a request, falling back to Opus 4.8:

Cybersecurity (exploitation / offensive tasks)
Biology / chemistry (e.g. bioweapon, toxin synthesis)
Distillation (attempts to extract the model's capabilities — cf. distillation)

Anthropic states >95% of Fable sessions involve no fallback at all, and that a 1,000+ hour external bug-bounty found no universal jailbreak (though the UK AI Security Institute reportedly "made progress toward one" in early testing). See AI safety, jailbreaks & red-teaming.

Benchmark claims (attribution matters — read the caveats):

Anthropic (first-party): state-of-the-art on "nearly all tested benchmarks"; first to break 90% on Anthropic's core analytics benchmark ("a 10-point jump over Opus"); highest score among frontier models on Cognition's FrontierCode and Hebbia's finance benchmark.
Artificial Analysis (third-party, 2026-06-10): Intelligence Index 64.9 (#1); AA-Omniscience 40; Humanity's Last Exam 53% (>7 pts ahead of next-best); GDPval-AA Elo 1932.
Press / vendor blogs (lower confidence): SWE-bench Verified ~95.0%; SWE-bench Pro 80.3% (vs Opus 4.8 69.2%, GPT-5.5 58.6%, Gemini 3.1 Pro 54.2%). At least one outlet (TechJack) reports the SWE-bench Pro figure is contested — treat it as unverified.

Honest caveats & unknowns

It's ~1 week old. Most numbers are from launch-window first-party material or single third-party aggregators. Expect revisions; independent replications are thin.
SWE-bench Pro 80.3% is disputed. Quoted from vendor/news blogs and explicitly flagged as "contested" by at least one independent write-up. Do not cite it as settled.
It was temporarily pulled offline. Within ~3 days of launch, a U.S. government export directive temporarily forced Fable 5 offline after Amazon's security team flagged a jailbreak vulnerability to the White House. White House AI adviser David Sacks suggested the block could be lifted once Anthropic remediates the safety issue. As of the sources read (mid-June 2026), no confirmed reinstatement date — current live availability is uncertain and should be re-verified before relying on the API.
No public architecture / parameter / training details. Mythos-class internals (size, training recipe, knowledge cutoff) are not disclosed in the read sources. The API docs do not state a knowledge cutoff for Fable/Mythos 5.
Anthropic's own benchmark table referenced in the announcement was not fully itemized in the page text we could fetch — granular per-benchmark first-party scores beyond the analytics/HLE/coding headlines were not captured.
"Mythos" vs "Mythos Preview" — Mythos 5 supersedes an earlier "Mythos Preview." The exact lineage/timeline of the preview is not detailed here.

How it connects to OpenAlice

Alice harness / sub-agent providers today are NOT Anthropic Fable. Per shared memory, OpenAlice's default provider is Codex OAuth, default model `gpt-5.5` on mvp (prod still gpt-5.4), with the gpt-5.x family + Lyria-3 for music. So Fable 5 is not currently wired as Alice's primary brain.
However, this very sub-agent runs on Claude (Opus 4.8 / Sonnet via Claude Code) for the orchestration/research/coding layer — the deep-research + delegate pattern. Fable 5's profile (1M context, always-on adaptive thinking, multi-day agentic sessions, programmatic tool calling, memory tool) maps directly onto that delegation substrate, and onto OpenAlice's stated goal that autonomous repo→PR coding is a base AGI capability of Alice (measured via SWE-bench).
If/when OpenAlice evaluates Fable 5, the integration cost is non-trivial and worth flagging up front: - Refusal handling: stop_reason:"refusal" arrives as HTTP 200 — any client that treats non-text completions as errors will mis-handle it. Plan the fallback path (→ Opus 4.8 or a non-Anthropic model) explicitly. Connects to model routing & cost-optimal serving. - Cost: at $10/$50 per M, Fable is roughly 2× Opus 4.8 — fine for hard, sparse, long-horizon tasks (migrations, multi-file RCA), wasteful for routine turns. A cost-ladder / router that reserves Fable for the top rung fits OpenAlice's existing cost-ladder research rig. - Data retention: the mandatory 30-day, no-ZDR "Covered Model" status conflicts with any zero-retention posture — relevant to OpenAlice's GDPR-compatible data-handling bar.
Until live availability re-stabilizes (see the export-directive caveat), treat Fable 5 as a candidate upgrade for the hardest agentic-coding tier, not a drop-in — and verify the API is actually serving before planning around it.

Claude Fable 5 (Anthropic) — the first public Mythos-class model