For months, the question hanging over every Anthropic release was “when does the Mythos-class model go GA?” The May 28 Opus 4.8 announcement teased it for “the coming weeks.” The answer arrived today at Code with Claude Tokyo, and it isn’t called Mythos. It’s called Claude Fable 5 — and it’s not an Opus at all.
A new tier, not a new version#
Fable 5 sits above Opus in Anthropic’s lineup. That’s a structural change, not a marketing one. Since March 2024, the Claude family has had three rungs — Haiku, Sonnet, Opus, named in ascending order of poem length. Fable breaks the metaphor on purpose: it’s not a longer poem, it’s a different kind of writing. Anthropic’s positioning is blunt — “our most powerful, most intelligent model” — and the version number does the rest of the talking. This is the first Claude to carry a 5. The 4.x family (Opus 4.8, Sonnet 4.6, Haiku 4.5) continues unchanged underneath it.
The model ID is claude-fable-5. No date suffix, no tier prefix. Context window is 1M tokens, max output 128K — matching Opus 4.8 on both.
And about that Mythos question: Anthropic hasn’t drawn the line publicly, but the timing is hard to ignore. A “Mythos-class” model was promised for all customers within weeks of May 28; thirteen days later, a new top tier ships under a fresh name with no security baggage attached. Mythos spent 2026 becoming shorthand for “the model too dangerous to release” — autonomous zero-day discovery, Project Glasswing, restricted defense partners. Whatever capability lineage Fable 5 actually has, Fable is a name you can put on an enterprise procurement form. Mythos is not. If this is the commercial release of that tier, the rename was the easy call.
The pricing signal#
Fable 5 costs $10 per million input tokens and $50 per million output — exactly 2× Opus 4.8’s $5/$25.
This is worth pausing on, because Anthropic has spent two years holding flagship pricing flat. Opus 4.6, 4.7, and 4.8 all shipped at $5/$25; the 1M context window came at standard pricing with no long-context premium. The discipline was the message: capability compounds, price doesn’t.
Fable 5 breaks that pattern — but in the defensible direction. It doesn’t reprice Opus; it adds a premium rung above it. Opus 4.8 remains the workhorse at unchanged pricing, and 69.2% SWE-bench Pro at $5/$25 is still the best deal in frontier coding. Fable is for the work where model quality is the binding constraint and 2× token cost is noise next to the engineering hours at stake: codebase-scale migrations, long-horizon autonomous runs, the hardest debugging sessions. Anthropic is letting the market segment itself.
What changes at the API level#
If you’ve already migrated to Opus 4.7 or 4.8, Fable 5 is nearly a drop-in: same adaptive-thinking-only surface, sampling parameters (temperature, top_p, top_k) still removed, assistant-turn prefills still rejected, effort levels from low through xhigh and max, Task Budgets still in beta.
There is exactly one new breaking change, and it’s a telling one: an explicit thinking: {"type": "disabled"} now returns a 400. Omitting the thinking parameter still runs without thinking — but the explicit off-switch is gone from the schema. Opus 4.8 accepts it; Fable rejects it. The trajectory Anthropic has been on since adaptive thinking landed — deprecate budget_tokens, remove sampling knobs, make the model decide — continues. Each generation removes another dial humans were probably misusing.
One small gift for high-throughput users: Fable 5’s minimum cacheable prompt prefix is 2,048 tokens, half of Opus 4.8’s 4,096. Shorter system prompts that silently failed to cache on Opus will cache on Fable.
What Fable does not have: a Fast mode. The 2×-rate-for-2.5×-speed option introduced with Opus 4.8 remains Opus-only. If latency is your constraint, Anthropic’s answer is still Opus 4.8 Fast, not Fable.
Available in Claude Code today#
Fable 5 is selectable in Claude Code as of this morning — /model fable and you’re on it. That’s the rollout pattern Anthropic has standardized: the flagship lands in the agentic harness on day one, because the harness is where the capability actually cashes out. No waitlist, no research preview gating for the model itself.
For teams on Bedrock, Vertex, or Foundry: no availability announced yet. Following the Opus 4.8 precedent, expect a multi-cloud lag of days to weeks.
The benchmark silence#
Anthropic published no SWE-bench numbers at announcement. For a company that put 69.2% SWE-bench Pro at the top of the Opus 4.8 release thirteen days ago, the silence is conspicuous — and there are two readings.
The cynical one: the numbers aren’t ready, or aren’t differentiated enough above Opus 4.8 to lead with. The likelier one, given the Mythos lineage: Anthropic is deliberately not framing this model around scores. Mythos Preview’s announcement in April also contained zero coding benchmarks — entirely capability narrative. If Fable 5 is the step-change researchers described, a 2-point SWE-bench delta would undersell it, because the benchmark saturates exactly where this class of model is supposed to pull away: tasks that take hours, not minutes.
We’ll update the models reference the moment numbers land. Until then, the honest position is the one we always recommend: run it on your own backlog. The 2× pricing means the evaluation question is concrete — does Fable 5 finish work that Opus 4.8 leaves 80% done? For migrations and long autonomous runs, that’s not a benchmark question. It’s a Tuesday afternoon experiment.
The bigger picture#
Forty-one days separated Opus 4.7 from 4.8. Thirteen days separate 4.8 from Fable 5. Anthropic closed a $65B Series H two weeks ago, crossed $47B run-rate revenue, and is reportedly approaching its first operating profit. The cadence isn’t slowing — it’s compounding, and the new naming scheme creates headroom the 4.x ladder didn’t have. There is now a tier above Opus. Which means there can be a Fable 5.1.
The competitive field should read the structure, not just the model. OpenAI’s GPT-5.5 leads Terminal-Bench; Google’s Gemini 3.5 Pro goes GA this month with 2M context. Anthropic’s answer is a price-segmented capability ladder topped by a model it can iterate without disturbing the $5/$25 workhorse underneath. That’s not a release. That’s a product line.
Sources: Anthropic newsroom, Claude platform docs — models overview, Code with Claude Tokyo, Claude Opus 4.8 coverage