This article was originally published on aicoderscope.com
TL;DR: Claude Fable 5 is the strongest coding model you can buy today — 80.3% on SWE-bench Pro, well ahead of Opus 4.8 — but at $10/$50 per million tokens it is exactly double Opus 4.8. For most agentic coding it is overkill. Use the free Pro/Max window (ends June 22) to test it, then route to it only for the hard 20% of tasks.
| Claude Fable 5 | Claude Opus 4.8 | GPT-5.5 | |
|---|---|---|---|
| Best for | The hardest agentic tasks, long-horizon refactors | Daily agentic coding, best price/performance | Cheaper agent runs, Copilot/Codex users |
| Price (in / out per M) | $10 / $50 | $5 / $25 | ~half Fable's input |
| SWE-bench Pro | 80.3% | 69.2% | 58.6% |
| The catch | 2× the token bill; reroutes to Opus 4.8 on restricted topics | Not as strong on the gnarliest tasks | Trails both on agentic coding |
Honest take: Fable 5 is the best, and you should not run it as your default. Keep Opus 4.8 (or a cheaper model) on the wheel and reach for Fable 5 only when a task actually defeats it. The June 22 free window is the one time to use it freely — burn through your hardest backlog before then.
What Fable 5 actually is
Anthropic shipped Claude Fable 5 on June 9, 2026 — the first publicly available "Mythos-class" model, a tier above Claude Opus 4.8. It shares the same underlying model as Claude Mythos 5, the frontier system Anthropic keeps gated. The difference is safety: Fable 5 runs with production safeguards across cybersecurity, biology, chemistry, model distillation, and frontier-AI-development topics. Mythos 5 itself stays restricted to Glasswing partners and a small set of biology researchers who get those guardrails lifted.
The safeguard mechanism matters for developers in one specific way: when a request trips one of those restricted categories, Fable 5 hands the response off to Opus 4.8 instead. For ordinary application code this never fires. If you write security tooling, exploit research, or anything that brushes the cyber category, you may silently get an Opus 4.8 answer rather than a Fable 5 one. It is not an error — it is the design — but it is worth knowing before you wonder why a "Fable 5" run felt like Opus.
The headline specs: a 1M-token context window by default and up to 128k output tokens per request. The API model ID is claude-fable-5, and it launched simultaneously on the Claude API, Amazon Bedrock, Google Vertex AI, Microsoft Foundry, and inside GitHub Copilot and Claude Code.
The benchmark numbers that actually matter for coding
Anthropic positioned Fable 5 as a coding and science leap, and the agentic-coding numbers back that up. On Anthropic's own SWE-bench Pro table, the model lands at 80.3%, with Opus 4.8 at 69.2%, GPT-5.5 at 58.6%, and Gemini 3.1 Pro at 54.2%. That is the widest gap between an Anthropic flagship and the field in a couple of release cycles.
| Benchmark | Claude Fable 5 | Claude Opus 4.8 | Notes |
|---|---|---|---|
| SWE-bench Pro | 80.3% | 69.2% | Anthropic-reported; the harder, contamination-resistant set |
| SWE-bench Verified | ~95% | — | Reported by independent trackers; treat as ceiling, not floor |
| Terminal-Bench 2.1 | 88.0% | — | Long-horizon terminal/agentic tasks |
| Context window | 1M tokens | 1M tokens | Same ceiling as Opus 4.8 |
| Max output | 128k tokens | — | Useful for whole-file rewrites |
A caveat on SWE-bench Verified: the ~95% figure circulating from independent trackers is impressive, but Verified is a saturating benchmark and several teams have flagged contamination concerns across the 2026 model wave. SWE-bench Pro — newer, harder, and built to resist memorized solutions — is the more honest signal, and that is where the 11-point lead over Opus 4.8 lives. Eleven points on Pro is real; do not anchor on the 95%.
What does that translate to in practice? On a multi-file refactor where Opus 4.8 gets stuck re-reading the same files or proposes a change that breaks a downstream import, Fable 5 is materially more likely to hold the whole dependency graph in its head and land the change in one pass. On a 40-line bug fix, you will not notice the difference — and you will pay double for the privilege.
Pricing: the part that hurts
Fable 5 costs $10 per million input tokens and $50 per million output tokens. That is exactly 2× Claude Opus 4.8, which sits at $5 / $25 and has not moved since Opus 4.7. It is also roughly double GPT-5.5 on input.
Agentic coding is output-heavy and context-heavy — agents re-read files, generate diffs, run tools, and re-read again. A single non-trivial agent task on a real repo can chew through 200k–500k input tokens and 30k–80k output tokens once you count the back-and-forth. Run the math at Fable 5 rates:
Example: one medium agentic task
Input: 400,000 tokens × $10 / 1M = $4.00
Output: 60,000 tokens × $50 / 1M = $3.00
----------------------------------------
Fable 5 total per task ≈ $7.00
Same task on Opus 4.8 ($5 / $25):
Input: 400,000 × $5 / 1M = $2.00
Output: 60,000 × $25 / 1M = $1.50
----------------------------------------
Opus 4.8 total per task ≈ $3.50
Ten tasks a day on Fable 5 is ~$70/day, ~$1,400/month, before you cache anything. That is the number that should make you pause. Two things blunt it: prompt caching reads run at roughly 0.1× base input, and the Batch API takes 50% off both directions for non-interactive jobs. Neither helps an interactive Cursor session much, where you are paying full freight on fresh context every turn.
The June 22 free window — what to do with it
From launch on June 9 through June 22, Fable 5 is included at no extra cost on Pro, Max, Team, and seat-based Enterprise plans. On June 23 it is removed from those plans, and continued use requires credits or API billing.
This is the single most useful fact in this article. For roughly two weeks you can run the best coding model in the world against your hardest problems without watching a token meter. Concrete plan:
- Pull your "too hard for the agent" backlog forward. The refactors you keep deferring because Opus 4.8 fumbles them — do them now, on Fable 5, while it is free.
-
Run it inside Claude Code, where the model picker and
--model claude-fable-5flag let you switch per-session without touching API keys. - Benchmark it on your own repo, not on Twitter screenshots. Give it the same gnarly task you gave Opus 4.8 last week and compare the diffs.
- Decide before June 23 whether any of your work genuinely needs it at $10/$50, or whether routing covers you.
After June 22, the calculus flips hard toward "route, don't default."
Setting it up as your coding backend
Claude Code is the path of least resistance. Select Fable 5 from the model picker, or pin it for a session:
# One-off session on Fable 5
claude --model claude-fable-5
# Inside an existing session, switch with /model
> /model claude-fable-5
If you run a Claude Code power-user setup with CLAUDE.md, slash commands, and subagents, you can scope Fable 5 to a single subagent that handles only the hard architectural work and leave the cheaper model on everything else.
Cursor: add Anthropic as a custom model in Settings → Models, supply your Anthropic API key, and register the model ID claude-fable-5. Note that Cursor's own Pro request quota and its Anthropic passthrough are billed separately — using your own key bills you at Anthropic's $10/$50 directly, which can be cheaper or more expensive than Cursor's bundled pricing depending on volume. If you are weighing the two, our Cursor vs Claude Code comparison breaks down where each one's billing model wins.
Cline (VS Code): open the model dropdown, pick the Anthropic provider, paste your API key, and select claude-fable-5. Cline shows a running token/cost estimate per task in the sidebar — leave it visible. On Fable 5 it is the difference bet
