Anthropic released Claude Opus 5 on July 24, 2026 — priced identically to its predecessor Opus 4.8 at $5 per million input tokens and $25 per million output tokens, while benchmarking close to the flagship Fable 5 at half its $10/$50 cost. It is available on the Claude API, Amazon Bedrock, Google Cloud, and Microsoft Foundry. For operators running Anthropic models in production, the substitution economics are now compelling — but two API-level changes mean upgrading is not a straight model-ID swap.

Key takeaways:

• Opus 5 is priced at $5/$25 per million tokens — same as Opus 4.8, half of Fable 5 ($10/$50)
• All benchmark results are vendor-reported; no independent replication as of this writing
• Breaking change: passing thinking: disabled at effort xhigh or max now returns a 400 error
• If your team is paying Fable 5 prices for everyday complex work, a substitution test is now justified

What did Anthropic actually ship with Opus 5?

On Anthropic’s own benchmarks — all vendor-reported — the performance gap over Opus 4.8 is large. On Frontier-Bench v0.1, Opus 5 more than doubles Opus 4.8’s score at the same cost per task. Zapier CEO Wade Foster reported that Opus 5 passed 100% of AutomationBench tasks — real business workflows run end to end — at 1.5× the next-best model’s pass rate for the same token spend. On CursorBench 3.2, it lands within 0.5% of Fable 5’s peak score at half the cost per task.

Anthropic now frames its lineup explicitly: Opus 5 for complex daily work you hand off and review; Fable 5 for long-horizon autonomous projects; Sonnet 5 for high-volume scale; Haiku 4.5 for subagents. Opus 5 is the new default on Claude Max and the strongest model on Claude Pro. One disclosed limit: Opus 5 trails Mythos 5 on cybersecurity tasks.

What this means for operators running Anthropic in production

The cost case for Fable 5 just narrowed. If your team uses Fable 5 for everyday coding, analysis, or document review, Opus 5 benchmarks close to parity at half the price. That is not a reason to switch without testing — but it is a reason to run both on your own workloads before your next contract renewal.

The Opus 4.8 → Opus 5 upgrade has two documented API changes. First, per Anthropic’s adaptive thinking documentation, thinking is now on by default in Opus 5 — it was off by default in Opus 4.8. Second, Anthropic’s troubleshooting guide confirms: passing thinking: {"type": "disabled"} at effort xhigh or max returns a 400 error at request time. Pipelines at effort high or below with thinking disabled are not affected. Audit your effort settings and thinking configuration before swapping model IDs.

Token economics also shift. Per Anthropic’s prompt caching documentation, Opus 5’s minimum cacheable prompt length is 512 tokens — down from 1,024 tokens on Opus 4.8. Prompts in the 512–1,023 token range that were previously too short to cache now qualify, including many system prompts. Operators running Claude Code multi-agent systems or other Anthropic token-cost workflows should revalidate cost projections after migrating.

The near-term move is not a full fleet swap — it is a one-workload test. Watch next: independent replication of Anthropic’s benchmark claims over the next two to four weeks. That is the signal that justifies broader adoption.

FAQ

Is Claude Opus 5 better than Fable 5?

On Anthropic’s benchmarks — vendor-reported, not yet independently replicated — Opus 5 lands within striking distance of Fable 5 on coding and automation at half the cost. Fable 5 retains an advantage on long-horizon autonomous work. Run your own workload comparison before switching.

Will upgrading from Opus 4.8 to Opus 5 break our existing API integrations?

Only if your pipelines pass thinking: {"type": "disabled"} at effort xhigh or max — Anthropic confirms this returns a 400 error on Opus 5. Pipelines at effort high or below are not affected. Audit your effort and thinking settings before upgrading the model ID.