Claude Opus 5: Cost, Benchmarks, and When to Use It

Claude Opus 5 is the premium Claude model to test when Sonnet 5 misses something consequential. It brings near-Fable capability to the existing Opus $5 input / $25 output per million tokens price tier, but it is not categorically cheaper than Sonnet 5 per completed task. Effort, latency, tool-loop length, safeguards, and factual reliability decide whether the premium produces a better result or merely a larger bill.

The direct routing decision:

1. Start routine volume with Claude Sonnet 5.
2. Escalate difficult agentic coding, review, research, and knowledge work to Opus 5.
3. Test Fable only when measured task gains justify its $10/$50 price, retention terms, and guardrails.
4. Treat Mythos as approved-access infrastructure, not an Enterprise-seat entitlement.

Quick facts

Availability

An Enterprise seat alone does not establish Mythos access. Anthropic’s current documentation keeps Mythos 5 within Project Glasswing and approved organizations.

API economics: measure the completed task

Opus 5 keeps Opus 4.8’s standard $5/$25 token price. Standard prompt-cache writes cost $6.25 per million tokens and cache reads cost $0.50. Fast mode is a Claude API research preview at $10/$50 and is not currently available through Bedrock, Google Cloud, or Microsoft Foundry.

Those rates do not determine the cheapest successful route. A complete-task comparison must include thinking tokens, tool results, retries, cache behavior, the number of agent turns, human review, and latency.

Artificial Analysis reported:

So the precise conclusion is:

• Opus 5 max was not cheaper than Sonnet 5 max on the Intelligence Index task set: $2.03 versus $1.53 per task.
• Artificial Analysis reports that Opus 5 at high and xhigh can outperform Sonnet 5 and Opus 4.8 at lower cost per task on specific evaluations.
• On AA-Briefcase, Opus 5 max cost 20% less than Fable while scoring higher. High effort still narrowly exceeded Fable and cost less than half as much.

These are benchmark-specific results, not a universal promise about your repository or research workflow.

Independent evidence and its caveats

Artificial Analysis scored Opus 5 max at 61 on its Intelligence Index, effectively tied with Fable 5 at 60. It also reported a leading 1861 Elo on GDPval-AA v2 and 1720 Elo on AA-Briefcase.

The cost comes partly in time. On AA-Briefcase, high, xhigh, and max averaged 25.7, 34.3, and 36.2 minutes per task, versus 24.1 minutes for Opus 4.8 max. Higher effort also generated far more turns. Do not route interactive work to max effort merely because it leads a chart.

There is also a factual-reliability caveat. Artificial Analysis reported better AA-Omniscience accuracy than Opus 4.8 but a 50% hallucination rate on that evaluation because Opus 5 answered more often when uncertain. That result does not mean half of ordinary answers are false; it means retrieval and citation checks still matter even when agentic scores improve.

Fable and Mythos context

Fable 5 remains generally available. On subscription plans, the temporary included allowance ended July 7 and current use is through usage credits; API and cloud routes have their own entitlements. At $10/$50, it is an escalation only when a measured improvement over Opus 5 changes the outcome.

Mythos 5 remains limited to approved Project Glasswing organizations. Neither a Claude Enterprise seat nor access to Fable proves Mythos entitlement.

Migration from Opus 4.8

Changing the model ID is the easy part:

1

claude-opus-4-8  ->  claude-opus-5

Review these behavior changes before production:

• Thinking is on by default. thinking: {"type": "adaptive"} remains equivalent to the default.
• Five effort levels matter. Start at the default high, then move down when quality holds or up only for measured hard tasks.
• max_tokens includes thinking and visible output. Increase it where Opus 4.8 previously ran without thinking, especially at xhigh or max.
• Thinking cannot be disabled at xhigh or max. That combination returns HTTP 400; use high or below, or leave thinking enabled.
• Thinking-disabled tool output needs testing. Anthropic warns the model can occasionally emit a tool call as text or expose internal XML rather than a tool_use block.
• Remove redundant verification prompts. Opus 5 self-verifies more often; inherited “verify everything again” instructions can create costly over-verification.

Anthropic’s system card and launch post also describe stronger narrow cyber guardrails than Opus 4.8. On first-party surfaces, flagged requests can fall back to Opus 4.8; API fallbacks must be configured and observed rather than assumed.

Practical routing ladder

The best setting is the lowest effort that clears the acceptance bar. Record cost per accepted task, wall-clock time, tool failures, factual corrections, and human review—not tokens alone.

Related links

• /models/claude-sonnet-5/
• /models/claude-opus-4-8/
• /compare/models/premium/
• /value/smart-spend/
• /posts/claude-fable-5-mythos-5-cost-guardrails/
• /posts/how-to-read-ai-benchmarks/

Sources

Vendor evidence

• Anthropic: Introducing Claude Opus 5 (Archive)
• Anthropic Platform: What’s new in Claude Opus 5 (Archive)
• Claude: Plans and pricing (Archive)
• Anthropic Platform: Models overview (Archive)
• Anthropic: Claude Opus 5 system card (Archive)
• Anthropic: Redeploying Fable 5 (Archive)

Independent evidence

• Artificial Analysis: Opus 5 Intelligence Index analysis (Archive)
• Artificial Analysis: Claude Opus 5 on AA-Briefcase (Archive)

AIHackers has not reproduced the vendor or Artificial Analysis results.

Last verified: July 26, 2026. Pricing, plan entitlements, benchmark dashboards, safeguards, and cloud availability can change independently.