Anthropic shipped Claude Opus 5 (claude-opus-5) on 2026-07-24 at $5 / $25 per million input/output tokens - the same bill as Opus 4.8 - with agentic quality that comes close to Fable 5 at half Fable’s token rate. It is the new default on Claude Max, the strongest model on Claude Pro, and Anthropic’s recommended start for complex agentic coding. For agent workflows, use it as the orchestrator: replace claude-opus-4-8 with claude-opus-5, keep Sonnet 5 as the daily driver, and reserve Fable 5 for the rare step that still needs the top rung.
Definitions
Opus 5 - Anthropic’s orchestrator-tier model (claude-opus-5). Replaces Opus 4.8 for long-horizon agentic work at the same $5/$25. 1M-token context, up to 128K output tokens, adaptive thinking; effort defaults to high on the Claude API and Claude Code.
Frontier escalation - calling Fable 5 (claude-fable-5, $10/$50) for the single hardest step in a run. Still valid after Opus 5, but the gap narrowed - escalate only when Opus 5 isn’t enough, and budget for refusal handling.
Cyber classifiers - real-time cybersecurity safeguards on Opus 5 that allow source-code vulnerability finding but block binary-based scanning, penetration testing, and exploit generation. Lighter than Fable’s (~85% fewer interventions). Flagged requests can fall back to Opus 4.8.
What shipped
| Opus 5 | Opus 4.8 | Fable 5 | |
|---|---|---|---|
| API ID | claude-opus-5 | claude-opus-4-8 | claude-fable-5 |
| Input / output ($/M tokens) | $5 / $25 | $5 / $25 | $10 / $50 |
| Context / max output | 1M / 128K | 1M / 128K | 1M / 128K |
| Default on claude.ai | Max (strongest on Pro) | Predecessor | Highest-capability plan routing |
| vs predecessor | More than doubles Opus 4.8 on Frontier-Bench v0.1 at lower cost/task | - | Still highest widely available capability |
| Safety surface | Lighter cyber classifiers; optional fallback to 4.8 | No Opus 5-class classifiers | Full refusal contract (stop_reason: "refusal") |
| Data retention | None required for general access | Same | 30-day required (Covered Model) |
Fast mode: same model at ~2.5× default speed for twice the base price ($10/$50) on the Claude Platform and via usage credits in Claude Code.
Beta platform updates shipping with Opus 5: mid-conversation tool changes without invalidating the prompt cache, and automatic fallbacks on the API - flagged Opus 5 or Fable 5 requests can route to another model instead of blocking.
Sources: Introducing Claude Opus 5, Models overview.
Reading the launch numbers
Anthropic’s launch charts put Opus 5 ahead of Opus 4.8 at the same per-token price, and close to Fable on several agentic evals at half Fable’s rate:
| Eval | Opus 5 claim (Anthropic launch) |
|---|---|
| Frontier-Bench v0.1 | Surpasses all other models; more than doubles Opus 4.8 at lower cost per task |
| CursorBench 3.2 (max effort) | Within 0.5% of Fable 5’s peak score at half the cost per task |
| OSWorld 2.0 | Outperforms every other model at any given cost; beats Fable 5’s best at just over ⅓ cost |
| ARC-AGI 3 | ~3× the next-best model |
| AutomationBench | ~1.5× next-best pass rate at the same cost per task |
Migrate off 4.8 for orchestrators. Do not run every turn at Opus rates - Sonnet 5 and Haiku 4.5 still own the volume economics. See Why most agents default to the wrong Claude tier.
Classifier fallbacks: what to wire
Symptom - a request that ran on Opus 4.8 is now flagged and returns no usable output.
Cause - Opus 5’s cybersecurity classifiers are stricter on binary vulnerability scanning, penetration testing, and exploit generation.
Solution - enable fallback to Opus 4.8 (default in Claude.ai, Claude Code, and Claude Cowork). For the beta automatic-fallbacks path on the API, pass a fallbacks key in the request body so flagged Opus 5 or Fable 5 calls route to another allowed model instead of blocking - see Anthropic’s models overview for the current shape.
Biology-related requests blocked on Fable 5 now route to Opus 5 rather than Opus 4.8 - Opus 5 is Anthropic’s most capable generally available model for scientific research under classifier strictness similar to Opus 4.8. Mythos 5 (Anthropic’s research-tier model, not in general availability) remains stronger for long-running autonomous biology and offensive cyber; Opus 5 stays behind Mythos on exploit development even when it finds vulnerabilities at a similar rate.
Where Opus 5 sits in the stack
The role pattern from pick the right Claude tier holds; only the orchestrator ID moved. Same shape as worker-tier routing: plan on the strong tier, execute on the cheap one.
| Role | Tier | Why (post–Opus 5) |
|---|---|---|
| Frontier escalation | Fable 5 | Highest widely available capability - at 2× Opus token rate, plus refusal handling |
| Orchestrator | Opus 5 | Near-Fable agentic quality at $5/$25; Anthropic’s start recommendation for complex agentic coding |
| Default worker | Sonnet 5 | Near-Opus quality at $2/$10 intro / $3/$15 standard |
| Parallel worker | Haiku 4.5 | Cheapest fan-out tier |
FRONTIER=claude-fable-5 # rare escalation only
ORCHESTRATOR=claude-opus-5 # was claude-opus-4-8
DEFAULT=claude-sonnet-5
WORKER=claude-haiku-4-5Escalate to Fable only when Opus 5 still isn’t enough on the hardest step - not because the leaderboard says Fable tops every chart. Escalate to Opus 5 from Sonnet when the task is long-horizon or autonomy-critical; stay on Sonnet with effort: "medium" for volume work.
Bottom line
Migrate orchestrators to claude-opus-5 at the same $5/$25 you already pay for 4.8. Wire classifier fallbacks (or beta automatic fallbacks), keep Sonnet 5 as the daily driver, and treat Fable 5 as a rarer escalation than it was before Opus 5 closed the gap. Next move: update your launcher config, then re-read the tier guide against the four-role table above.