Narrative change, July 24, 2026
This page was rewritten after Opus 5 shipped: the recommendation changed from Fable 5-first to Opus 5-first. Anthropic's own guidance is now to start with Opus 5 and use Fable 5 only for workloads that need the highest available capability.
Why the recommendation flippedUpdated August 4, 2026
Claude Fable 5 vs Claude Opus 5
The July 24, 2026 launch flipped the default: Anthropic now recommends starting with Opus 5 at half the price, and reserving Fable 5 for the four job types it still dominates. Here is the full, sourced comparison.
The short answer
Default to Opus 5 — it matches Fable 5's capability band on the published benchmarks at half the price ($5/$25 vs $10/$50), it is the default model on Claude Max, and it carries no data retention requirement. Upgrade to Fable 5 only when you hit one of four cases: long-horizon agent runs, security-critical coding at Mythos level, full-repo 1M-context analysis, or a compliance policy that requires the 30-day retention only Fable 5 provides (where ZDR is not available). Full comparison below.
Specs and pricing side by side
API list prices from src/lib/model-pricing.ts, last verified August 4, 2026 against Anthropic's pricing documentation. Re-check before budgeting — both models are less than two months old.
| Claude Fable 5 | Claude Opus 5 | |
|---|---|---|
| Released | 2026-06-09 | 2026-07-24 |
| API model ID | claude-fable-5 | claude-opus-5 |
| Input / MTok | $10.00 | $5.00 |
| Output / MTok | $50.00 | $25.00 |
| Cache read / MTok | $1.00 | $0.50 |
| Cache write (5 min) / MTok | $12.50 | $6.25 |
| Batch (input / output) | $5.00 / $25.00 | $2.50 / $12.50 |
| Context window | 1M | 1M |
| Max output | 128K | 128K |
| Data retention | 30 days, mandatory — ZDR agreements do not apply | None for general access |
| Subscription access | Max / Team Premium, 50% of weekly pool | Max default · strongest on Pro |
| Fast mode (Claude apps) | Not offered for Fable 5 | ~2.5x speed at 2x price ($10/$50) |
Benchmarks: 12 comparisons, with sources
Every row is labeled: Anthropic = published in the vendor's own launch materials; AA = run by Artificial Analysis (independent). Anthropic published no Opus 5 figures for several classic benchmarks — those rows say not confirmed rather than guess.
| Claude Fable 5 | Claude Opus 5 | Source | |
|---|---|---|---|
| SWE-Bench Pro | 80.3% | Not confirmed (as of 2026-08-04) | |
| FrontierCode Diamond | 29.3% | Not confirmed (as of 2026-08-04) | |
| Frontier-Bench v0.1 | 33.7% | 43.3% | |
| CursorBench 3.2 | Holds the top published score | Within 0.5% of Fable 5's peak, ~half the cost per task | |
| ARC-AGI 3 | Not tested | ~3× the next-best model | |
| Zapier AutomationBench | Not published | ~1.5× the next-best pass rate | |
| OSWorld 2.0 / OSWorld-Verified | 85.0% | Beats Fable 5's best at ~1/3 the cost | |
| GDPval-AA | 1932 | 1861 | |
| AA-Briefcase | Not published | 1720 | |
| Humanity's Last Exam | 53% | 53% | |
| Terminal-Bench v2.1 | Not published | 89% | |
| AA Intelligence Index | 60 | 61 |
Anthropic: Opus 5 launch announcement (July 24, 2026) and Fable 5 launch table (June 9, 2026). AA: Artificial Analysis article on Opus 5 (July 24, 2026). Fable 5's SWE-Bench Pro 80.3% and FrontierCode 29.3% remain the highest Anthropic has published for a widely available model — an Opus 5 score for either has not been published as of August 4, 2026. The GDPval-AA figures come from different harness versions (Anthropic's June 9 table for Fable 5 vs AA's v2 for Opus 5) and are not directly comparable.
Decision matrix: if you're… use…
The four upgrade-to-Fable-5 cases, the default Opus 5 case, and the budget case. Everything else starts at Opus 5.
You run multi-hour autonomous agents
Long-horizon agent loops (30-90+ minutes, many tool calls, file memory) are what Fable 5 was built for — it is Anthropic's 'long-running agents' model. Fable 5.
You do security-critical coding
Mythos-level security work (exploit-resistant code, defensive tooling) is Fable 5's domain — Opus 5 itself 'remains behind Mythos 5' here. Fable 5 (or a Mythos variant).
You analyze a full repository in one context
Whole-repo reads that saturate the 1M window — both models have 1M, but Fable 5's peak on the hardest coding tasks (SWE-Bench Pro 80.3%) is the published ceiling. Fable 5.
Your compliance needs retention, or ZDR is unavailable
Your data policy requires retention logs, or ZDR is not available. Fable 5's mandatory 30-day retention is exactly the log-retention some policies require — and where zero-data-retention cannot be agreed, Fable 5 leaves no retention gap to negotiate. Opus 5, by contrast, has no retention at all, which fails a retention-required policy. Fable 5.
You're starting a new project
Complex agentic coding and enterprise work — Anthropic's own recommendation is to start with Opus 5, and it is already the Max default. Opus 5 @ high.
You're optimizing budget at scale
High-volume coding and tool use at 1/5 the price of Fable 5's list — Sonnet 5 ($2/$10 intro through August 31, 2026) covers most routine work. Sonnet 5.
Decision guide as of August 4, 2026, based on Anthropic's published guidance, published benchmarks and Anthropic's plan documentation. Your own evals beat any default.
Coding decision: which model solves what at what cost
The coding numbers, plus the arithmetic that decides: cost per solved task, not pass rate.
Agentic coding benchmarks
SWE-Bench Pro: Fable 5 at 80.3% is the highest published figure — Opus 5 has no published score (not confirmed). FrontierCode Diamond: Fable 5 at 29.3%; Opus 5 not published (not confirmed). Frontier-Bench v0.1, Anthropic's newest agentic-coding test: Opus 5 at 43.3% beats Fable 5's 33.7% — this is the benchmark Anthropic led with for Opus 5.
What real usage shows
Artificial Analysis measured Opus 5 at 61 vs Fable 5's 60 on its Intelligence Index at max effort, and Terminal-Bench v2.1 at 89% for Opus 5 (no published Fable 5 figure). Lovable reported a 22% internal improvement over Opus 4.7 on its hardest agentic-coding evals. Fable 5's 80.3% SWE-Bench Pro figure uses Anthropic's own scaffolding — score only within the same harness.
Cost per solved task
At max effort, Artificial Analysis measured roughly $2.03 per task for Opus 5 against $2.75 for Fable 5 — a 26% saving. Worked example: 100 coding tasks at high effort, ~10K input tokens each (50% cached) and ~8K output. Fable 5 ≈ $45.50; Opus 5 ≈ $22.75. If Fable 5 solves 2% more of those tasks, the premium buys two extra solves; at 10,000 tasks a month, the gap funds a whole team.
Safeguard fallback: on Claude.ai, Claude Code and Claude Cowork, requests flagged by Fable 5's classifiers fall back to Opus 4.8 — billed at Opus 4.8 rates. Cyber classifiers on Opus 5 intervene roughly 85% less often than on Fable 5.
What a month of real workloads costs
Three representative workloads priced at list rates, before batching and caching — see the cost calculator for your real numbers.
Agentic coding (1,000 tasks/mo)
Fable 5 ≈ $1,720 · Opus 5 ≈ $860
1M-context analysis (50 jobs/mo)
Fable 5 ≈ $470 · Opus 5 ≈ $235
High-volume support (200K requests/mo)
Fable 5 ≈ $12,750 · Opus 5 ≈ $6,375 · Sonnet 5 ≈ $2,550 (intro)
Priced at API list rates: 50% cache hit on input, +30% output tokens for agentic runs. The calculator has the exact math for your volumes. Because Opus 5 can spend more tokens per answer than Fable 5, the realised gap is often smaller than the 2x list price — never budget from list price alone.
The difference that decides enterprise deals
Fable 5 requires a mandatory 30-day retention window for safety monitoring, and zero-data-retention agreements do not apply to it. Opus 5 has no data retention requirement for general access — Anthropic's spokesperson raised this unprompted at launch, specifically for customers with a hard ZDR requirement. If a compliance review blocked Fable 5 at your company, that blocker does not exist on Opus 5.
Fable 5's unique angle: safeguard routing you can't see in a benchmark
Three things separate the models that no benchmark table shows.
Classifier-driven routing
Fable 5's safety classifiers divert flagged requests to Opus 4.8 (on Claude.ai, Claude Code, Claude Cowork) or to Opus 5 (biology requests, since July 24). Opus 5's cyber classifiers intervene roughly 85% less often.
30-day retention is mandatory
Fable 5 requires 30-day retention for safety monitoring and is excluded from ZDR agreements. Opus 5 has no retention requirement for general access.
Mythos 5 is the actual ceiling
The top published cybersecurity and biology scores belong to Mythos 5 (invitation-only via Project Glasswing). Fable 5 is the widely available tier; Opus 5 'remains behind Mythos 5' on those domains.
When Fable 5 is still the right pick
Four cases where the premium still pays — the upgrade cases from the decision matrix.
Long-horizon autonomy is the job
Multi-hour agent runs with file memory are what Fable 5 was built for, and it has the longer public track record on them. Anthropic describes it as its 'long-running agents' model.
You need Mythos-level security work
On cybersecurity and biology, Opus 5 itself remains behind Mythos 5 — and Fable 5 is the widely available tier of that family. If your workload is security-critical coding, this is not a price question.
You're on Max and inside the 50% weekly pool
Max and Team Premium include Fable 5 up to 50% of the weekly allowance. Inside that pool the marginal cost is zero — the arithmetic changes entirely.
Your harness is tuned to it and your bill is fine
Migration is a one-line change, but it is not free — prompt caches, effort levels and evals all carry over. If your setup is stable and the bill acceptable, there is no urgency.
How to switch
Both models take the same request shape, so in most codebases this is a one-line change.
API — change the model ID
Replace claude-fable-5 with claude-opus-5. The effort parameter (defaults to high on both), the absent temperature/top_p/top_k, and adaptive thinking all carry over unchanged. Opus 5 also added automatic fallback routing on the API.
Claude Code — /model opus
Run /model opus in the session, or set ANTHROPIC_MODEL=claude-opus-5. On Max, Opus 5 is already the default — you may have been using it since July 24 without noticing.
Claude apps — nothing to do on Max
Opus 5 became the default on Claude Max and the strongest model on Claude Pro on launch day, with no change to subscription pricing.
Price your real workload before you switch
Run your actual monthly volume through the calculator at both price points — with caching and batching stacked, the gap is often larger than the 2x list price suggests.