Page added July 2, 2026. Sonnet 5 launched June 30 at intro pricing ($2/$10 through Aug 31, then $3/$15). Fable 5 is included on Pro/Max/Team at 50% of weekly limits only through July 7, then usage credits. What belongs on which →
FableGuide › Fable 5 vs Sonnet 5
The numbers
| Fable 5 | Sonnet 5 | |
|---|---|---|
| Model ID | claude-fable-5 | claude-sonnet-5 |
| Price (in/out, per MTok) | $10 / $50 | $2 / $10 intro through Aug 31, 2026 → $3 / $15 |
| Launched | June 9, 2026 (returned July 1) | June 30, 2026 |
| Context window | 1M tokens | — |
| Plan availability | Window through Jul 7, then credits | Default model on Free & Pro; all plans, Claude Code, API |
| Class | Mythos-class frontier | Workhorse / agent runner |
Note on the intro price: Sonnet 5 ships a new tokenizer that processes text at roughly 1.0–1.35× the token count of earlier Sonnets — Anthropic says the $2/$10 intro window is designed to make the transition cost-neutral. Budget on the post-August $3/$15 rate.
What Sonnet 5 actually changes
- Agent work got 5× cheaper. The pitch (and TechCrunch's headline read) is "a cheaper way to run agents": planning, tool use, browsers, terminals, autonomous runs at a level that recently needed bigger models. On agentic-search and computer-use benchmarks Anthropic reports it matching Opus 4.8's capability at medium-to-high effort — for a fraction of the cost.
- It finishes what it starts. The consistent early-tester theme: it completes complex tasks where earlier Sonnets stopped short, and checks its own output unprompted. One developer: "Unprompted, it wrote a reproducing test, implemented the fix, then stashed it to confirm the bug came back without the change."
- On raw agentic coding it slots below Opus. Reported numbers put Sonnet 5 at 63.2% vs Opus 4.8's 69.2% and Sonnet 4.6's 58.1% on agentic coding — a real jump over its predecessor, still short of Opus, well short of Fable.
- Calmer safety profile. Lower hallucination and sycophancy rates than Sonnet 4.6, better prompt-injection resistance — and substantially weaker offensive-cyber capability than Opus 4.8, with standard cyber safeguards on by default (lighter than Fable's post-return classifier).
Where Fable 5 is still alone
- The hardest problems. Multi-day autonomous runs, first-shot builds of complex systems, whole-codebase bug hunts, frontier reasoning. Nothing in the lineup — Sonnet 5 included — replaces it there. The full triage →
- 1M-token context with top-end reasoning. The combination, not just the window.
- Fleet management. Fable reliably orchestrates dozens of parallel subagents — and Sonnet 5 is exactly what you staff those subagent fleets with.
The routing table, post-return
| The task | Route to | Why |
|---|---|---|
| Hardest, longest, most ambiguous | claude-fable-5 | Frontier capability; spend the window on these. |
| Serious daily engineering | claude-opus-4-8 | The dependable heavyweight at $5/$25. |
| Agents, pipelines, apps, subagent fleets | claude-sonnet-5 | Near-Opus agent behavior at $2/$10 intro. |
| Instant, tiny, high-volume | claude-haiku-4-5 | Speed and pennies. |
Moral: Sonnet 5 didn't replace Fable — it replaced the excuse for running expensive models on cheap work. Pair them: Fable to lead, Sonnet 5 to staff the crew.
Sources
Anthropic — Introducing Claude Sonnet 5 · TechCrunch · VentureBeat · Benchmark figures are the launch-week reported numbers; independent testing is still arriving.