Claude Fable 5
Anthropic released 2026-06-09 proprietary
Claude Opus 5 is both better and cheaper.
It scores +1 higher and costs 50% less ($10.00/M against $20.00/M) on this workload — and it does everything this model does.
The same model via batch is 50% cheaper ($10.00/M) — same weights, different latency.
Our take
editorial — not a measurementThe current top of the LMArena text board (1508) and the most expensive non-Pro model in the catalogue at $10/$50. It is priced 2x Opus 5 for roughly one index point of measured intelligence (AA 62 vs 63 for Opus 5 at max effort), which makes it a bad default and a defensible choice only for long-running agents where a single avoided failure is worth more than the token bill. Two things make the real cost worse than the sticker: it inherits the Opus 4.7-generation tokenizer, so the same text produces about 30% more tokens than a Sonnet 4.6-era model, and its measured time-to-first-token is roughly 98 seconds, which rules it out of anything interactive. Buy it for unattended overnight work, not for a chat box.
- Unverified in this take: This note cites a time-to-first-token figure that the data layer withheld as implausible (it appears to be total reasoning time mislabelled at source). Treat it as unverified.
- Unverified in this take: Anthropic does publish a token-inflation figure, but it compares newer Claude models against EARLIER CLAUDE models — not against another vendor. The cross-vendor version of this claim is unsupported, so read any comparison to a non-Claude price with that in mind.
Strengths
- Highest LMArena text Arena Score observed (1508)
- Full 1M context at flat pricing — no long-context tier penalty
- 128k max output, extendable to 300k via the batch API beta header
- 50% batch discount and 0.1x cache reads apply across the whole window
Weaknesses
- $10/$50 per MTok — 2x Opus 5 for a marginal intelligence delta
- ~98s time-to-first-token makes it unusable interactively
- ~30% token inflation vs pre-4.7 Claude models (Anthropic's own figure) is a hidden real price rise
- Scores one AA index point below Opus 5 despite costing double
Reach for it when
- Long-horizon unattended agents
- Work where a retry costs more than the tokens
- Batch jobs that can absorb the latency
Avoid it if
- You need interactive latency
- You are cost-sensitive
- Opus 5 already passes your evals
Sources: platform.claude.complatform.claude.comarena.aiartificialanalysis.ai
Every price dimension
| Input | $10.00/M |
|---|---|
| Output | $50.00/M |
| Cached input | $1.00/M |
| Cache write | $12.50/M |
| Cache write 1h | $20.00/M |
| Web search | $0.01/call |
| Batch discount | 50% |
Independent scores
| Intelligence | 62.1 |
|---|---|
| Coding | 76.5 |
| Agentic | 56.6 |
| LMArena Elo | 1494.7default |
Sources are listed separately rather than averaged. Across the 62 models both have scored they correlate at r = 0.806 — close agreement overall, but four models rank very differently between them, and a blended score would hide exactly those.
Best rankings by task
- asciiart #1of 83 1367
- godotgamedev #1of 43 1344
- agenticgamedev #1of 34 1295
- htmlslides #1of 34 1270
- agentichtmlslides #1of 16 1254
- agenticslides(html) #1of 16 1252
- gamedev #2of 144 1383
- androidnative #2of 54 1303
Capability
| Context window | 1M |
|---|---|
| Max output | 128K |
| Input modes | text, image, file |
| Tool use | yes |
| Reasoning | always on |
| Open weights | no |
Provenance
| Price source | openrouter.ai |
|---|---|
| Fetched | 2026-08-24 |
| Quality data | verified |
| Cross-checked | vendor page |
A published time-to-first-token figure for this model was implausible (98s) and has been withheld rather than displayed.
Claude Fable 5 is a Mythos-class model from Anthropic, built for autonomous knowledge work and coding. It supports text, image, and file inputs with text output, with reasoning support and...