Claude Opus 5 vs Fable 5: The API Routing Guide

Anthropic released Claude Opus 5 on July 24, 2026 at $5 per million input tokens and $25 per million output. That’s half of Claude Fable 5’s $10/$50, and by Anthropic’s own launch table it matches or beats Fable 5 on most comparable benchmarks (Anthropic, 2026). Startrise’s routing default two days in: Opus 5 everywhere, Fable 5 reserved for the long-horizon workloads that measurably use its ceiling. And one rule decides it before benchmarks even load: Fable 5 requires 30-day data retention, Opus 5 doesn’t. For zero-data-retention orgs the comparison is over before it starts.

The day-one coverage reformats Anthropic’s benchmark table. This guide answers the question an engineering team actually has: which model ID goes in which pipeline, what it costs, and what breaks when you switch. It’s the framework we use to route our own AI automation pipelines, dated July 26, 2026, because this topic moves daily.

Key Takeaways

• Opus 5 shipped July 24, 2026: claude-opus-5, $5/$25 per MTok, 1M context — half Fable 5’s price
• Anthropic’s launch table is a split decision, not a knockout; the lone independent read (Epoch’s index) is a two-point spread
• ZDR orgs can’t touch Fable 5 (Covered Model, 30-day retention) — except via a documented per-workspace override
• Safety classifiers arrive on the Opus line: refusal handling is now mandatory on both frontier models
• Opus 5 fast mode costs exactly Fable 5 money ($10/$50) and buys up to 2.5x speed instead of ceiling
• In our own 12-build benchmark, Fable 5’s 2x rate card produced only a 1.18x bill ($23.84 vs $20.27) — it wrote 42% fewer output tokens than Opus 5
• Migration from Opus 4.8 is a model-ID swap plus three behavior checks

Claude Opus 5: what shipped on July 24?

Claude Opus 5 (claude-opus-5) is Anthropic’s new default frontier model, released July 24, 2026: $5/$25 per million tokens, unchanged from Opus 4.8 (Anthropic, 2026), with a 1M-token context window and up to 128k output tokens (models overview, 2026). Anthropic positions it as “a thoughtful and proactive model that comes close to the frontier intelligence of Claude Fable 5 at half the price.”

The operational changes matter as much as the scores:

• Thinking is on by default. Requests without a thinking field now run with adaptive thinking, the opposite of Opus 4.8’s behavior (migration guide, 2026).
• The effort ladder runs low to max, the cost-versus-capability toggle Fortune led its launch coverage with. Disabling thinking is allowed only at effort high or below; combined with xhigh or max it returns a 400.
• Safety classifiers, new to the Opus line. Opus 5 can decline a request with stop_reason: "refusal", just as Fable 5 does (refusals docs, 2026). Handling details are in the switching section below.
• Fast mode (research preview, Claude API only): $10/$50 for up to 2.5x output speed.
• Prompt-cache minimum halved to 512 tokens, from 1,024 on Opus 4.8.
• Knowledge cutoff May 2026 — four months newer than Fable 5’s January 2026.

Opus 5 vs Fable 5: the comparison table

Here’s the spec-level picture as of July 26, 2026, with Opus 4.8 and Sonnet 5 for context (Anthropic models overview, retention docs, refusals docs, 2026):

One line on the legacy label: Opus 4.8 moved to Anthropic’s Legacy models list on July 24 but “remains available” on all platforms. More on it below.

Benchmarks: is Opus 5 actually better than Fable 5?

It’s a split decision, not the knockout the headlines suggest. On Anthropic’s launch table, Opus 5 leads most directly comparable evaluations: on Frontier-Bench v0.1 it “more than doubles Opus 4.8’s performance at a lower cost per task” (Anthropic, 2026), and Decrypt’s read of the launch table puts the Frontier-Bench split at 43.3% for Opus 5 against Fable 5’s 33.7%. Anthropic also reports Opus 5 within 0.5% of Fable 5’s peak CursorBench score at max effort, at half the cost per task, and ahead of it on OSWorld 2.0 at just over a third of the cost.

Fable 5 isn’t finished, though. Decrypt notes it still “excels by a tiny margin” in legal and health domains, and on The Decoder’s compiled index scores Fable 5 edges Opus 5 on the Epoch Capability Index, 161 to 159. The Decoder’s overall verdict: “No single model can pull away or claim a clear advantage.” Fable 5 also holds the only independent labor-benchmark record: 15.8% of real freelance projects completed at client-acceptable quality on the CAIS Remote Labor Index. That score was set against Opus 4.8’s 8.3%, not against Opus 5, which has no RLI score yet as of July 26, 2026.

The honest read: for most workloads, the 2x premium no longer buys a visible gap — but “most” is doing real work in that sentence. Two days post-launch, nearly every head-to-head number traces back to Anthropic’s own table; the lone independent read is Epoch’s capability index, where two points separate the models. Treat the launch numbers as the vendor’s, and trust evals on your own tasks over any launch table.

What does each model actually cost to run?

Sticker price is $5/$25 versus $10/$50 — a clean 2x. The real math has more moving parts.

We ran the invoice, and 2x sticker came out 1.18x real. In the Startrise LLM Benchmark’s cost analysis (run 2026-07-26-0159, 12 single-shot frontend builds per model), Fable 5’s suite bill was $23.84 against Opus 5’s $20.27 — about 18% more, on a rate card that’s exactly double. The reason: Fable wrote 42% fewer output tokens, 469,422 against Opus’s 803,391. Double the rate, 0.58x the volume, 1.18x the bill. Note the direction — Fable is expensive because of its rate, not verbosity; it’s one of the terser frontier models in our data. The same run produced two inversions worth staring at: Haiku 4.5 is 17% cheaper per output token than GPT-5.6 Luna yet billed 57% more ($0.74 vs $0.47), and Sonnet 5 is 33% cheaper per output token than Kimi K3 yet billed 29% more ($9.26 vs $7.17) — each wrote about 1.9x the tokens. So budget in output tokens per task, not dollars per million: verbosity does as much work as price, and it’s on no pricing page. Caveat honestly stated — one frontend suite, run once.

The old framing died on July 24. June-era comparisons leaned on “Fable bills thinking, Opus doesn’t.” Both models now run adaptive thinking by default. The residual difference: Opus 5 can still disable thinking at effort high or below; Fable 5 returns a 400 at any effort level (migration guide, 2026). Fable has a hard price floor per request. Opus 5 keeps an escape hatch for cheap turns.

effort is the real spend lever on both models. The ladder runs to max, and thinking tokens bill as output.

The equivalence nobody prices out: Opus 5 fast mode costs $10/$50, Fable 5’s standard rate to the dollar. The same dollars buy up to 2.5x output speed on Opus, or the capability ceiling on Fable. That’s a genuine fork: latency or intelligence, at identical spend.

Refusal billing — now on both frontier models: refusals stopped before any output aren’t billed, mid-stream refusals bill the input plus the output already streamed, and the fallbacks credit refunds prompt-cache cost on retry (refusals docs, 2026).

Caching: Opus 5’s minimum cacheable prompt is 512 tokens, half of Opus 4.8’s 1,024. Short system prompts that never cached now do.

The compliance fork: ZDR decides first

For a zero-data-retention organization, the model choice is made before anyone opens a benchmark table. Claude Fable 5 and Claude Mythos 5 are the only designated Covered Models: they require 30-day data retention, ZDR “is therefore not available for either model,” and an API request to Fable 5 from a ZDR org returns a 400 invalid_request_error telling you retention must be enabled (Anthropic retention docs, 2026). Opus 5 appears nowhere on that list. The announcement states it “does not have data retention requirements for general access” (Anthropic, 2026). Opus 5 is the ZDR-compatible frontier option.

Now the detail no ranking page mentions: the ban is scopeable. Anthropic documents that a ZDR organization can enable 30-day retention for a single workspace (Claude Console > Settings > Workspaces > Privacy controls) and run Fable 5 only there, while every other workspace keeps zero data retention (retention docs, 2026). “Fable is banned for us” becomes an architecture decision: one fenced workspace with retention, everything else untouched. (Two footnotes, used sparingly: HIPAA readiness is an org-level alternative to ZDR, and content flagged by trust-and-safety systems may be retained up to two years under any arrangement.)

Deciding which workloads may touch a Covered Model is an inventory-and-controls question, the same one EU AI Act questionnaires ask, and it’s what an AI compliance readiness engagement maps.

What breaks (and what doesn’t) when you switch?

On both frontier models — and the day-one comparisons get every one of these wrong:

• Refusal handling is mandatory, on Opus 5 too. “Claude Fable 5 and Claude Opus 5 include safety classifiers that can decline a request” (refusals docs, 2026). A refusal arrives as stop_reason: "refusal" inside an HTTP 200, with a stop_details.category of "cyber", "bio", "frontier_llm", "reasoning_extraction", "general_harms", or null. Naive code reads a refusal as success. Wire the fallbacks beta parameter (which now supports a "default" mode) or retry client-side. Comparisons that frame refusals as the Fable tax are pre-July-24 thinking.
• Raw chain of thought is never returned. thinking.display defaults to "omitted" on both models, "summarized" is the ceiling, and “No display setting returns the raw chain of thought” (thinking docs, 2026). Design for summaries or nothing.
• Assistant prefill returns a 400. The migration guide flags this as unchanged from Opus 4.8. It’s a shared constraint, though several ranking comparisons misreport it as Fable-specific.

Choosing Fable 5, the extra work on top of that:

• thinking: {"type": "disabled"} returns a 400 at any effort level (migration guide, 2026). Opus 5 still allows it at effort high or below.
• No web fetch, and the knowledge cutoff is January 2026 (Anthropic docs, 2026).
• 30-day retention is non-negotiable. That’s the Covered-Model fork above.

Choosing Opus 5 from Opus 4.8 is lighter: a model-ID swap plus three behavior checks. Thinking is now on by default, so revisit max_tokens budgets; disabling thinking at effort xhigh or max returns a 400; and classifier refusals now exist on the Opus line, so handle stop_reason: "refusal" even if you never touch Fable. Token counts are roughly unchanged, since Opus 5 uses the same tokenizer generation as 4.8 (migration guide, 2026). Anthropic adds one behavioral note: Opus 5 self-verifies, so strip carried-over “double-check your work” instructions. Over-verification is now a real token cost.

This category of work (refusal handling, fallback paths, escalation when a model declines) is the guardrail engineering of agent development, not something you bolt on later.

Where does Opus 4.8 fit now?

Opus 4.8 moved to Anthropic’s Legacy list on July 24 but is still available on all platforms, at the same $5/$25 as Opus 5, and it’s still ZDR-eligible (migration guide, 2026). The June question (Fable 5 vs Opus 4.8) has a measured answer: the CAIS RLI gap, 15.8% vs 8.3%. But at identical pricing, the live question since July 24 is 4.8 versus Opus 5, and Anthropic’s migration path is the ID swap plus the three checks above.

Legitimate reasons to hold on 4.8: regression-tested pinned workloads you can’t re-evaluate this quarter, and Opus 5’s behavior deltas (longer deliverables, more subagent delegation) that you haven’t measured yet. If you’re deciding whether the Fable tier belongs in your business at all, that case (and the June suspension timeline) lives in our Claude Fable 5 business guide.

The routing decision

The rules we apply, in order:

1. ZDR-bound? Opus 5, or the per-workspace retention carve-out if one workload genuinely needs Fable 5. The Covered-Model rule never softens.
2. Everyday agentic coding and enterprise work → Opus 5. It’s Anthropic’s own default recommendation, at half the frontier price.
3. Long-horizon, multi-hour agentic runs (deep migrations, the legal/health domains where Fable’s narrow launch-table edges and its RLI record live) → Fable 5, with fallbacks wired, if your evals confirm the gap on your tasks.
4. Latency-critical at premium budget → Opus 5 fast mode: Fable dollars, up to 2.5x speed.
5. Routine volume → Sonnet 5, intro-priced $2/$10 through August 31, 2026.
6. Needs post-January-2026 knowledge → Opus 5 (May 2026 cutoff).

These are the rules on our own bill. Startrise routes both tiers daily in production: our five-agent editorial department runs Fable 5 on long-horizon research and strategy legs, Opus-class models on volume legs, with a human operator owning routing and every review gate. This comparison was itself researched by Fable 5 running inside our tooling, Opus-class fallback wired, directed and verified by the named human on the byline, per our no-ghostwriters rule.

Route it once, or have it routed for you

Model routing is not a config value; it’s fallback-and-escalation infrastructure. If you want it built, our AI Agents That Act engagement ships agents on Mastra with guardrails, approval gates, and exactly the fallback and escalation paths this post describes, from $3,500, in 2–4 weeks. If you’d rather have routing operated than owned, a Human-Agent Team embeds senior operators with a configured agent fleet in your infra from $6,000/mo. And if the Covered-Model wall is where your evaluation stopped, start with AI Compliance Readiness, from $7,500, and turn the retention question into an architecture answer.

Two days in, the routing table is clear enough to act on. Just date-stamp your assumptions (this one is July 26, 2026) and re-run your evals when the independents publish.