Claude Opus 5 vs Fable 5: Which Should You Actually Run for Support? (2026)

Anthropic’s own documentation makes the starting point clear. Its models overview tells customers to start with Opus 5, and reach for Fable 5 only when a task needs the highest capability available. A vendor rarely says this about its own top-tier model, and it shows how close the two have actually gotten on real work.

Fable 5 launched first, in June, as Anthropic’s most capable public model. Six weeks later, Opus 5 showed up costing half as much and closing most of the gap. Any team that had already built a support workflow around Fable 5 in those six weeks now has an actual decision to make. For a support team weighing the two, that guidance offers a useful starting point, though it doesn’t fully answer the question once real tickets enter the picture.

This piece works through what each model actually is, where each one wins. It also covers the risks that don’t show up in a launch announcement. And it covers which one belongs on a support queue, and where. This piece speaks directly to a support team making that call, keeping the focus on tickets, cost per resolution, and what happens when the model gets something wrong.

---

What Are Claude Opus 5 and Claude Fable 5?

Both are current-generation Claude models from Anthropic, released six weeks apart, and both show up on the same pricing page and the same model picker. What separates them is what each one is built to do.

Claude Opus 5

• Newest model in Anthropic’s Opus tier, released July 24, 2026
• Built for everyday production use, and now the default model on Claude Max and the strongest model offered on Claude Pro
• Carries real reasoning depth, leading Anthropic’s own agentic and knowledge-work evaluations at launch, and is the model behind the agentic workflow wins detailed further down
• Priced at $5 input and $25 output per million tokens, exactly half of Fable 5
• Available across the Claude API, Claude.ai, Claude Code, Claude Cowork, Amazon Bedrock, Google Cloud, and Microsoft Foundry
• Carries no data retention requirement for general access, a detail that matters for any team running it near production customer data

Key takeaway: Opus 5 is built to deliver most of Fable 5’s intelligence at a lower, more sustainable operating cost.

Claude Fable 5

• Anthropic’s most capable model available to the public, sitting above Opus 5 on Anthropic’s capability ladder
• Built for frontier-level reasoning and long-horizon task execution, planning across stages, delegating to sub-agents, and checking its own work over sessions that can run for days
• The stronger fit for complex planning and research-style work that needs to hold context and stay on track across many steps without supervision at every turn
• The highest capability tier generally available today. Anthropic’s only model above it, Mythos 5, stays limited to vetted partners through Project Glasswing and has not shipped to the public
• The model that spent nineteen days offline in June 2026 under a US export control order, a story covered in full in YourGPT’s dedicated Fable 5 breakdown

Key takeaway: Fable 5 excels when a task needs extended reasoning and autonomous execution sustained over a long stretch of work.

Anthropic’s own guidance for choosing between them is direct for a vendor talking about its own flagship. Its models overview tells customers to start with Opus 5, and step up to Fable 5 only for the workloads that genuinely need the highest available capability.

---

Comparing Opus 5 and Fable 5

Pricing below is verified directly against Anthropic’s live pricing page as of this writing. Re-confirm before publishing anything with a hard number in it, since Anthropic updates this page independently of its launch posts.

The retention and Fast Mode rows are the two worth reading twice. Everything else lines up exactly as expected.

---

The Areas Where Fable 5 Leads 

Fable 5’s lead is real but narrow, and it shows up in a specific kind of work, not across the board.

• Keeps a real, if narrow, lead on a handful of coding and reasoning benchmarks, per Anthropic’s own Opus 5 system card: SWE-bench Pro at 80 versus Opus 5’s 79.2, DeepSWE v1.1 at 69.7 versus 68.8, and FrontierCode 1.1’s main set at 53.5 versus 53.4
• These are fractions of a point, too small on their own to drive a deployment decision
• Carries lower factual knowledge risk. On Artificial Analysis’s AA-Omniscience benchmark, Fable 5 posts the highest accuracy of any evaluated model at 65 percent, ahead of Opus 5
• Sits at rank one on the LMArena text leaderboard as claude-fable-5, with an Elo of 1507 across nearly 15,000 votes, though that snapshot predates Opus 5’s release and measures blind human preference, a different signal than benchmark intelligence
• Owns multi-day autonomous work outright. It plans across stages, hands work to sub-agents, and checks its own output once at the end of a long run
• That shape of work runs inside an agent harness, which matters for where this model actually fits in a support stack, covered further down

---

The Areas Where Opus 5 Leads 

Flip to agentic and computer-use work and the picture reverses, and the gaps stop being decimal dust.

• Anything with a tight agent loop goes the other way, and the margins are not close, per Anthropic’s launch post
• FrontierBench v0.1: Opus 5 43.3 versus Fable 5 33.8, a gap of +9.5
• OSWorld 2.0 (computer use): Opus 5 70.6 versus Fable 5 66.1, a gap of +4.5
• Zapier AutomationBench: Opus 5 26.0 versus Fable 5 17.4, a gap of +8.6
• GDPval-AA v2 (Elo): Opus 5 1861 versus Fable 5 1747, a gap of +114
• On Zapier’s own benchmark, Opus 5 hit 100 percent on a churn-prevention workflow that every prior model had failed outright, without spending more tokens than earlier Claude models
• On OSWorld 2.0, Anthropic states Opus 5 surpasses Fable 5’s best result at just over a third of the cost
• Artificial Analysis ran an independent agentic knowledge-work benchmark, AA-Briefcase, that tells the same story from a different angle
• At high effort, Opus 5 scored 1606 Elo at $10.41 per task against Fable 5’s 1574 Elo at $22.30 per task
• A lower-cost setting on the cheaper model beat the flagship on both quality and price at once, which does not happen often in this category

---

Data Retention and Hallucination Risk: Read This Before Choosing

This is the section that matters most for anything customer-facing, more than any benchmark table above it.

• Fable 5’s retention requirement is mandatory. Per Anthropic’s support documentation, Fable 5 retains prompts and outputs for 30 days, with no zero-retention option. Support conversations carry PII and payment details, so a zero-retention compliance baseline has to route around Fable 5. Opus 5 carries no retention requirement for general access.
  
• Fable 5’s safety classifiers intervene far more often than Opus 5’s. On the FrontierBench run, Anthropic’s system card recorded Fable 5’s classifiers flagging 42 percent of calls across 26 percent of trials, against 5 and 4 percent for Opus 5, about 85 percent less often. Cybersecurity requests fall back to Opus 4.8. Biology requests now route to Opus 5. Most support traffic avoids these categories, but B2B verticals like security or healthcare can trip them, and a silent mid-conversation swap is worse than an expected one.
  
• Opus 5 answers more often when it’s uncertain, and that has a cost. On AA-Omniscience, Opus 5’s accuracy rose 7 points, but hallucinations climbed 14 points to 50 percent, with lower factual knowledge than Fable 5. A wrong guess just fails a test in code. On a refund question, it’s the whole failure. What the model is grounded in, and when it escalates to a human, matters more than which model runs underneath. More on hallucination mechanics.

---

The Right Model for Different Teams 

Ticket type is one axis. Team size and stage is the other, and it changes how much either flagship’s price tag actually matters.

Startup

Recommendation: Opus 5

• Better economics at low ticket volume and a tight budget, since Opus 5 costs half of Fable 5 on every metered line
• Quality strong enough for most early-stage support work once grounded in a real knowledge base, which matters more at this stage than squeezing out the last benchmark point

Mid-Market SaaS

Recommendation: Mostly Opus 5, with Fable 5 used selectively

• Opus 5 covers the working queue: order status, account questions, standard troubleshooting, at a cost that scales with ticket volume without a flagship-sized bill
• Fable 5 earns its price on the smaller slice of complex, multi-system tickets that benefit from its long-horizon planning, the same category covered in the routing table below

Enterprise Support

Recommendation: A hybrid approach across both models

• Opus 5 handles the large majority of standard resolution work at high effort, with Fable 5 reserved for the escalation tier, multi-system billing disputes, account reconciliation, and cases that need sustained reasoning across many steps
• The exact split depends on each team’s own ticket distribution, set by testing against real tickets

This mirrors how mature AI support systems already route requests by complexity, the mechanics of which are covered next.

---

The Hybrid Routing Strategy

Picking one model for an entire queue is the easy decision. Picking the right model for each ticket is the one that actually controls the bill and the quality bar at the same time.

A basic version of this pattern looks like:

1. Customer message arrives
2. Ticket gets classified by intent and complexity
3. Standard tickets route to Opus 5
4. Complex, multi-system tickets route to Fable 5
5. A human takes over when confidence drops below a set threshold

Every comparison in this piece up to this point has been framed as Opus 5 against Fable 5. That is how the decision usually gets presented. In production, it rarely stays a binary choice. A single queue holds order-status questions that need almost no reasoning next to multi-system billing disputes that need all of it, often within the same hour. Committing the whole queue to Fable 5 means paying flagship prices for tickets that never needed flagship reasoning. Committing the whole queue to Opus 5 means the small slice of genuinely hard tickets gets a model built for daily use. The days-long autonomous planning those tickets sometimes need is Fable 5’s job.

Routing solves both problems at once. It sends routine tickets to whichever model handles them at the lowest defensible cost. It reserves the expensive reasoning for the tickets that actually need it. A single best model is a leaderboard answer. A routed queue is a production answer.

---

Choosing the Right Model for Your Support Queue

The right pick depends on what a wrong answer costs on that specific ticket type.

The short version: default to Opus 5 at high effort for the hard tickets that used to justify Fable 5’s price tag, keep Fable 5 for the narrow slice of genuinely multi-day autonomous work, and keep both flagships off the routine 70 to 80 percent of a queue entirely.

---

Turning the Right Model Into a Working Support Bot

This is where the decision above turns into an actual, running support bot.

• Picking a model answers one part of the question. The other part is whether a support stack can act on the answer without a rebuild the next time a new model ships, which is exactly what happened here six weeks after Fable 5 launched
  
• A platform where the model is one configuration setting turns “Opus 5 just closed most of Fable 5’s gap” from a re-architecture into a settings change
  
• YourGPT already lets teams choose between models from OpenAI, Anthropic, Google, and xAI in a single setting, and pairs that choice with the parts of a deployment that decide answer quality regardless of which model is underneath
  
• That includes retrieval grounded in a business’s own knowledge base, confidence-gated human handoff with full conversation context, and a Studio layer for building the routing logic described above without touching the underlying prompts
  
• Before routing production ticket volume to either flagship, three checks apply regardless of platform
  
• Confirm the account’s data retention settings against whichever model is selected, since the two behave differently by design
  
• Ground answers in retrieval sources the team controls, since that grounding is what determines whether a 50 percent hallucination rate on a closed-book benchmark ever reaches a customer
  
• Test the actual ticket distribution against both models before committing a queue to one, since the benchmark gaps above were measured on general agentic and coding work, and a live support queue behaves differently

---

Frequently Asked Questions

---

What is the main difference between Claude Opus 5 and Claude Fable 5? ▼

Both have a 1M-token context window and a 128,000-token maximum output. Fable 5 costs $10 per million input tokens and $50 per million output tokens, compared with Opus 5 at $5 and $25. Fable 5 retains advantages on some evaluations, while Opus 5 leads on several agentic and computer-use benchmarks. On AA-Briefcase, Opus 5 at high effort scored above Fable 5 while costing less than half as much per task.

How much does Claude Opus 5 cost compared to Fable 5? ▼

Opus 5 costs $5 per million input tokens and $25 per million output tokens. Fable 5 costs exactly double at $10 and $50. Both use the same Batch API and prompt-caching discount structure, so the 2:1 price difference also holds across standard, batch, and prompt-cache pricing.

Which model should run a support queue, Opus 5 or Fable 5? ▼

Neither should be the default for high-volume, routine tickets such as order status or FAQs. For harder tickets involving complex, multi-system resolution, Opus 5 at high effort offers a strong balance of capability and cost. Fable 5 is particularly well suited to genuinely multi-day, highly autonomous workflows where maximum capability matters more than cost.

Does Claude Opus 5 hallucinate more than Fable 5? ▼

On Artificial Analysis’s AA-Omniscience benchmark, yes. Opus 5 recorded a 50% hallucination rate and lower factual-knowledge performance than Fable 5. That does not mean Opus 5 will hallucinate more in every application. For customer-facing bots, grounding answers in a controlled knowledge base and routing uncertain responses to a human remains important regardless of the model used.

Is Claude Fable 5 safe to use with customer support data? ▼

It can be, but teams should check their retention requirements first. Fable 5 requires 30-day data retention for safety monitoring and is not available under zero data retention. Opus 5 does not have Fable 5’s model-specific 30-day retention requirement and can be used with eligible ZDR configurations.

What effort level should Claude Opus 5 run at for support work? ▼

High effort is a practical starting point for demanding support workflows. On AA-Briefcase, Opus 5 at high effort showed a strong balance between performance and cost. Max effort can improve results further, but it uses more output tokens and costs more per task, so it is better reserved for the smaller share of tickets that genuinely need additional reasoning.

Can I use Claude Opus 5 or Fable 5 inside YourGPT? ▼

YourGPT lets teams choose between models from OpenAI, Anthropic, Google, and xAI without locking a support bot to a single model provider. The specific Claude models available can change as providers release new models and they are added to the platform, so check the model settings inside Studio for the current list.

What happens to a support bot if Claude Fable 5 goes offline again? ▼

Anthropic suspended Fable 5 access on June 12, 2026, following a US export-control directive, and restored access on July 1 after the controls were lifted. A support stack tied to Fable 5 without a fallback model could lose service during a provider-level suspension. On a platform such as YourGPT, where the model can be changed through configuration, workloads can shift to another available model without rebuilding the bot.

---

Conclusion

Opus 5 changes the calculus behind six weeks of Fable 5 deployments. It closes most of the agentic-benchmark gap at half the price, and on Artificial Analysis’s own benchmark, it beats Fable 5 outright at a fraction of the cost. Fable 5 still leads a short list of coding benchmarks and remains the right tool for genuinely multi-day, autonomous work.

Neither belongs on the routine tickets that make up most of a real queue. That job goes to a cheaper, faster tier. What decides how either flagship performs on its suited tickets hasn’t changed. It’s what the model is grounded in. It’s also what it can do without a human checking first. And it’s how fast a wrong answer gets caught before it reaches a customer.