Verdict: Claude Opus 5 ($5/$25 per million tokens) delivers near-frontier performance at half the price of Claude Fable 5 ($10/$50), and on several coding and reasoning benchmarks it actually beats Fable 5 outright. For 90%+ of production work — agentic coding, knowledge tasks, everyday AI pipelines — Opus 5 is the better choice. Fable 5 remains the pick for multi-day autonomous runs and offensive cybersecurity research where every decimal point of capability matters regardless of cost.
Last verified: 2026-07-30
- Opus 5 launched July 24, 2026 at $5/M input, $25/M output — same as Opus 4.8.
- Fable 5 launched June 9, 2026 at $10/M input, $50/M output — exactly double.
- Opus 5 leads on 7 of 12 shared benchmarks; Fable 5 holds an edge on SWE-bench Pro and CursorBench by razor-thin margins.
- ARC-AGI-3 score: Opus 5 scored 30.2%, nearly 4× the next-best model (GPT-5.6 Sol at 7.8%).
- Pricing/limits change often — last checked 2026-07-30.
What is the price difference between Claude Opus 5 and Fable 5?
Claude Opus 5 costs $5 per million input tokens and $25 per million output tokens — identical to its predecessor Opus 4.8, and exactly half of Fable 5's $10/$50 rate. Both models share a 1M-token context window and 128K max output.
| Spec | Claude Opus 5 | Claude Fable 5 |
|---|---|---|
| Input price | $5.00 / 1M | $10.00 / 1M |
| Output price | $25.00 / 1M | $50.00 / 1M |
| Context window | 1M tokens | 1M tokens |
| Max output | 128K tokens | 128K tokens |
| Fast mode | Yes ($10/$50, 2.5× speed) | Not offered |
| Data retention | None (general access) | 30-day mandatory retention |
| Model ID | claude-opus-5 |
claude-fable-5 |
| Knowledge cutoff | May 2026 | January 2026 |
Sources: Anthropic Opus 5 announcement, Artificial Analysis model page, BenchLM Claude API pricing. Pricing verified against official Anthropic documentation, July 2026.
The data retention difference matters operationally: Fable 5 has a 30-day mandatory retention policy on some surfaces, while Opus 5 has no retention requirements for general access. This is why benchmark platforms like ARC-AGI publish Opus 5 scores but not Fable 5's — ARC-AGI's creators will not risk leaking benchmark details to a model with a data retention policy (ARC Prize Foundation).
How does Opus 5 compare to Fable 5 on benchmarks?
On independent and vendor-reported benchmarks, Opus 5 either matches or exceeds Fable 5 on the majority of evaluations while costing half as much per token.
| Benchmark | What it measures | Opus 5 | Fable 5 | Source |
|---|---|---|---|---|
| Frontier-Bench v0.1 | Agentic coding (% pass) | 43.3% | 33.7% | Anthropic |
| ARC-AGI-3 | Novel reasoning (anti-memorization) | 30.2% | ~20% (ARC-AGI estimate) | ARC Prize Foundation |
| GDPval-AA v2 | Knowledge work (Elo) | 1,861 | 1,747 | Artificial Analysis |
| SWE-bench Pro | Real GitHub issues resolved | 79.2% | 80.0% | Anthropic |
| CursorBench 3.2 (max effort) | In-editor coding | Within 0.3 pts | Peak leader | Anthropic |
| OSWorld 2.0 | Computer use (screen + apps) | Beats Fable 5's peak at ~1/3 cost | Peak leader | Anthropic |
| Zapier AutomationBench | End-to-end automation workflows | 100% (full flow) | Lower throughput | Anthropic |
On the Artificial Analysis Intelligence Index — the independent, third-party tracker — Opus 5 (max effort) scores 61, narrowly ahead of Fable 5 at 60 and GPT-5.6 Sol (max) at 59 (Artificial Analysis).
The standout result is ARC-AGI-3. This benchmark, designed by François Chollet to resist memorization, drops an AI agent into interactive, turn-based environments where it must discover rules through trial and error. Opus 5's 30.2% score is roughly four times GPT-5.6 Sol's 7.8% — the previous best — and the ARC Prize Foundation noted a capability "previously unseen from frontier models": Opus 5 turned the test layouts into algebraic notation to perform advanced logical reasoning (ARC Prize Foundation).
How does Opus 5 compare to GPT-5.6 Sol?
GPT-5.6 Sol (OpenAI) costs $5 per million input tokens and $30 per million output tokens, making it comparable to Opus 5 on input but 20% more expensive on output. On the Artificial Analysis indices, Opus 5 edges GPT-5.6 Sol on intelligence (61 vs 59) and ties on coding (78 each at different effort levels). GPT-5.6 Sol remains the better value for price-to-performance on routine tasks — costing roughly half of Opus 5 per completed task per Artificial Analysis cost-per-task charts (Artificial Analysis, OpenAI pricing).
For a deeper look at how Opus 5 benchmarks and pricing stack up across the full landscape, see our Claude Opus 5 benchmarks and developer pricing guide.
How do Opus 5 and Fable 5 perform on real coding tasks?
In hands-on coding tests — building an F1 racing game in a single HTML file with Three.js, then building a full-stack personal finance dashboard with a frontend and backend — both Opus 5 and Fable 5 produced working applications from a single prompt. The results reveal nuances that benchmark scores alone don't capture.
F1 game test: Both models generated playable 3D racing games with collision detection, lap timing, and positioning. Opus 5's version featured more polished 3D car models and better track aesthetics. Fable 5's game had a camera wobble effect on turns that some might prefer. Both were functional. GPT-5.6 Sol produced a game where the track broke halfway through, and Kimi K3 (the 2.8T open-weight model from Moonshot AI, launched July 16, 2026) produced a solid result despite being open-weight.
Dashboard test: Both Opus 5 and Fable 5 built working full-stack apps. Opus 5 chose React + React Router + Express + Node SQLite — a conventional, production-ready stack. Fable 5 also used React and Express with Node SQLite but built its own custom routing library instead of using React Router. GPT-5.6 Sol chose Next.js + Drizzle + Cloudflare/Vinxd — a valid but unusual stack for a local finance app.
Cost analysis: On these real-world tasks, token consumption varied significantly. The AI Artificial Analysis cost-per-task measurement shows Opus 5 at roughly $2.03 per Intelligence Index task versus Fable 5 at $2.75 — a 26% savings (Artificial Analysis). However, per-task cost depends heavily on effort level and token consumption, which can vary by 5× between models on identical tasks.
For practical tips on reducing LLM costs across any provider, see our model-agnostic LLM cost optimization guide.
When should you use Opus 5 instead of Fable 5?
Opus 5 is the right default for most teams doing agentic coding, knowledge work, and production AI pipelines. Fable 5 is the right choice only for specific high-endurance workloads.
Choose Opus 5 when:
- You need agentic coding (PR review, multi-file refactors, terminal agents)
- You do knowledge work at scale (research, financial modeling, legal review)
- Cost-per-task matters ( startups, small teams, subscription-constrained work)
- You want no data retention restrictions
- You use Claude Code or Claude Cowork daily
Choose Fable 5 when:
- You run multi-day autonomous agents (long-horizon tasks that need maximum capability)
- You need peak SWE-bench performance for the hardest real-world code fixes
- You do offensive cybersecurity research (Fable 5 has broader permission for exploitation work)
- You need Mythos-class capability for biology or advanced security tasks
Anthropic's own guidance echoes this: start with Opus 5 for complex agentic coding and enterprise work; reach for Fable 5 when you need the highest available capability, especially multi-day autonomy (Anthropic).
For a look at how Opus 5 handles single-prompt app builds including 3D games, see our Opus 5 one-shot 3D app build test.
How does Opus 5 handle cybersecurity differently from Fable 5?
Opus 5's cyber classifiers are proportionally less restrictive than Fable 5's, allowing source-code vulnerability discovery while blocking binary-based vulnerability scanning, penetration testing, and exploit generation. Anthropic reports these classifiers intervene approximately 85% less often than Fable 5's (Anthropic).
On the OSS-Fuzz cybersecurity evaluation, Opus 5 approaches Mythos 5's capability at identifying software vulnerabilities but is considerably less successful at developing exploits — an intentional design choice. When Opus 5's classifiers flag a request in Claude.ai, Claude Code, or Claude Cowork, the system falls back to Opus 4.8 by default, configurable at the API level.
Anthropic highlighted a real-world example: Opus 5 identified the root cause of a bug in a popular open-source package manager and patched an edge case the community's own fix had missed, while a competing model addressed only the surface symptom and incorrectly reported the issue as resolved (Anthropic).
For more on building AI-powered defense systems, see our autonomous AI agent cyberattack defender playbook.
What is the Kimi K3 wildcard in this comparison?
Kimi K3, the 2.8-trillion-parameter open-weight model from Moonshot AI (released July 16, 2026), is the wild card. At 57 on the Artificial Analysis Intelligence Index, it sits just behind Opus 5 (61), Fable 5 (60), and GPT-5.6 Sol (59) — but you can run it on your own infrastructure at marginal compute cost (Artificial Analysis, Better Stack).
In hands-on coding tests, Kimi K3 produced working applications from single prompts, though its UI output lagged behind the frontier models and its backend code choices (using a JSON file as a database instead of a real database) were less production-ready. For teams willing to trade some capability for full control and zero per-token cost, Kimi K3 is worth evaluating — especially alongside a framework like Hermes Agent for self-hosted intelligence.
What this means for you
If you're already paying for Opus 4.8 or Fable 5 through any Anthropic plan, switching to Opus 5 is a no-brainer. You get materially better performance at the same (or half the) price, no data retention restrictions, and it is the new default on Claude Max. The real decision is whether you also need Fable 5 in your routing table — and for most teams, the answer is no. Keep Fable 5 as an overflow tier for the hardest long-horizon tasks, but make Opus 5 your everyday driver.
For those on competing models: GPT-5.6 Sol remains the price-to-performance leader on routine tasks. Run your own benchmarks on your own workloads — the per-task cost delta can swing based on how many tokens each model needs to solve your problems.
FAQ
Q: Is Claude Opus 5 actually better than Fable 5? A: On 7 of 12 shared benchmarks, yes. Opus 5 leads on agentic coding (Frontier-Bench), novel reasoning (ARC-AGI-3), and knowledge work (GDPval-AA v2). Fable 5 holds a razor-thin edge on SWE-bench Pro (80.0% vs 79.2%) and remains Anthropic's recommendation for multi-day autonomous agent work. For 90%+ of production tasks, Opus 5 is both cheaper and at least as capable.
Q: How much does Claude Opus 5 cost? A: $5 per million input tokens and $25 per million output tokens — the same as Opus 4.8 and exactly half of Fable 5's $10/$50 rate. An optional fast mode doubles the price to $10/$50 for approximately 2.5× speed.
Q: Does Opus 5 have data retention restrictions? A: No. Opus 5 has no mandatory data retention for general access. Fable 5 has a 30-day mandatory retention policy on some surfaces, which is why benchmark platforms like ARC-AGI publish Opus 5 scores but not Fable 5's.
Q: Can Opus 5 find security vulnerabilities in code? A: Yes. Opus 5 is permitted to find vulnerabilities in source code, but it blocks binary-based vulnerability scanning, penetration testing, and exploit generation. Anthropic reports its classifiers intervene approximately 85% less often than Fable 5's.
Q: Should I switch from Opus 4.8 to Opus 5? A: Yes — there is no reason not to. The price is identical ($5/$25) and Opus 5 materially outperforms Opus 4.8 across all benchmarks (e.g., 43.3% vs 18.7% on Frontier-Bench, a 30.2% vs 1.5% on ARC-AGI-3). Opus 5 is the new default model on Claude Max.
Q: Is GPT-5.6 Sol cheaper than Opus 5? A: On raw token price they are comparable ($5 input for both; GPT-5.6 is $30 output vs Opus 5's $25). On cost-per-task, GPT-5.6 Sol is cheaper for routine work because it tends to use fewer tokens. Test on your own workloads to determine the cost winner for your specific tasks.
Q: Can Kimi K3 compete with Opus 5? A: Kimi K3 (score 57) is competitive but trails Opus 5 (61) and Fable 5 (60) on the Artificial Analysis Intelligence Index. It's a 2.8T open-weight model from Moonshot AI that you can self-host at marginal compute cost, trading some capability for zero per-token pricing.
Discussion
0 comments