Claude Opus 5 Migration Guide: Costs and Checklist

Anthropic released Claude Opus 5 on July 24, 2026. For teams already using Opus 4.8, the headline is unusually simple: Opus 5 keeps the same standard API price while adding stronger long-horizon agent behavior. The migration is close to drop-in, but two changes around adaptive thinking can break existing requests.

This Claude Opus 5 migration guide explains what changed, who should upgrade, how to estimate costs, and what to test before sending production traffic to the new model. Facts and prices were checked against Anthropic’s documentation on August 6, 2026.

Claude Opus 5 at a glance

Anthropic positions Opus 5 for complex agentic coding and enterprise work. According to the official model overview, it has a one-million-token context window, a 128,000-token maximum output, vision support, adaptive thinking, and moderate comparative latency.

The model ID is a fixed snapshot despite having no date suffix. Anthropic says dateless IDs introduced with the 4.6 generation are pinned versions, not pointers that silently move to a newer model.

In its launch announcement, Anthropic reports that Opus 5 outperformed Opus 4.8 on coding, automation, computer-use, knowledge-work, and life-science evaluations. It also says the model performs near the more expensive Fable 5 on some tasks. Treat those numbers as vendor-reported evidence, not a substitute for testing your own workload. The public GDPval-AA v2 leaderboard provides useful outside context, but no general benchmark can predict reliability on your tools, data, or acceptance criteria.

Should you migrate from Opus 4.8?

The strongest reason to migrate is better capability at the same token price. Anthropic charges both Opus 4.8 and Opus 5 at $5 per million input tokens and $25 per million output tokens. A team paying standard rates does not accept a higher list price merely by changing models.

Opus 5 is a strong candidate when your workload includes:

• multi-step research or analysis that must keep its goal over many actions;
• difficult debugging, code review, or root-cause investigation;
• agents that use several tools and must verify the result;
• document, spreadsheet, diagram, or visual-analysis workflows;
• large prompts or histories that benefit from the one-million-token window.

Do not switch all traffic immediately if you depend on Priority Tier, because Anthropic’s migration guide says it is not supported for Opus 5. The first-party web fetch tool is also unavailable on this model. Existing external search or retrieval tools may still work, but that integration needs its own test.

If you are new to agent systems, build a narrow evaluation workflow before adopting the most capable model. NextPJ’s five-day AI agents learning plan provides a practical sequence from task definition to testing.

The two breaking changes to check first

Anthropic describes Opus 5 as a drop-in upgrade from Opus 4.8, with two important exceptions.

1. Adaptive thinking is now on by default

On Opus 4.8, a request without an explicit thinking setting runs without thinking. On Opus 5, the same request uses adaptive thinking. This can improve difficult work, but it changes token use, latency, and the amount of the output allowance consumed before the visible answer appears.

The maximum-output setting remains a hard ceiling across thinking tokens and visible response tokens. A limit tuned for non-thinking Opus 4.8 traffic may therefore be too small after migration. Measure completion rates and truncated responses rather than assuming the old ceiling is safe.

2. Thinking cannot be disabled at the two highest effort levels

Opus 5 lets developers disable thinking only at low, medium, or high effort. Combining disabled thinking with xhigh or max effort returns HTTP 400. Opus 4.8 accepted that combination, so search request builders, presets, and per-route overrides for it before rollout.

Anthropic recommends keeping thinking enabled and lowering effort for most cost-sensitive tasks. Its prompting documentation warns that disabling thinking can occasionally expose tool-call text or internal XML-like tags in visible output, especially in tool-heavy workflows.

A safe Claude Opus 5 migration checklist

1. Change only the model first

Replace the Opus 4.8 model ID with claude-opus-5, but keep prompts, tools, and business logic stable for the first comparison. Changing several variables at once makes regressions hard to diagnose.

2. Run an effort sweep on real tasks

Test low, medium, high, xhigh, and max effort against the same evaluation set. Anthropic recommends starting with high, then using low or medium where quality holds. Reserve xhigh or max for tasks where a measurable accuracy gain justifies more tokens and latency.

Your evaluation should score task completion, factual accuracy, tool success, latency, input and output tokens, refusal rate, and human correction time. For agent observability patterns, see NextPJ’s guide to tracing AI agents with OpenTelemetry.

3. Revisit output limits and timeouts

Increase the output ceiling for demanding agent jobs, then cap it using measured data. Anthropic suggests starting at 64,000 tokens for xhigh or max effort and tuning from there. Also review application, proxy, worker, and tool-call timeouts; a model request can remain healthy while an upstream timeout terminates the workflow.

4. Remove redundant self-check instructions

Opus 5 verifies its work more readily than earlier models. Prompts that repeatedly demand a final check, a second verification pass, and a separate verification agent can cause over-verification and unnecessary spending. Remove that inherited scaffolding, then confirm on your evaluation set that quality does not fall.

Be explicit about scope and response length. Lowering effort controls thinking, not reliably the length of the visible response. If concise output matters, state the desired length or format directly.

5. Control subagent fan-out

Anthropic says Opus 5 delegates more readily than earlier models. Set a deterministic maximum number of subagents and permit delegation only for genuinely independent, substantial work. This prevents a small request from multiplying into several model runs.

6. Test caching and safety fallbacks

The minimum cacheable prompt is 512 tokens, down from 1,024 on Opus 4.8. That can make more system prompts eligible for caching. Also handle refusal responses explicitly. Anthropic offers an automatic fallback feature in beta for certain safety-classifier refusals, but production teams should monitor fallback frequency rather than hiding it.

For a broader control set around tools, isolation, approvals, and secrets, use the AI agent sandbox security checklist.

7. Canary before full rollout

Route a small, reversible share of eligible traffic to Opus 5. Compare it with Opus 4.8 by workload, not only in aggregate. Expand traffic only after quality, cost per successful task, latency, and refusal rates meet predefined thresholds. Keep the previous model available as a rollback target until the canary covers peak traffic and uncommon failure paths.

What Claude Opus 5 actually costs

The Claude API pricing page lists these standard rates as of August 6, 2026:

• Standard: $5 per million input tokens and $25 per million output tokens.
• Fast mode: $10 per million input tokens and $50 per million output tokens.
• Batch API: $2.50 per million input tokens and $12.50 per million output tokens.
• Prompt-cache hit: $0.50 per million input tokens, plus normal output charges.
• US-only first-party inference: a 1.1-times multiplier across token categories.

Here are illustrative per-request calculations before taxes or negotiated discounts:

Fast mode is roughly 2.5 times the default speed according to Anthropic, but costs twice the standard token rate and cannot be combined with Batch. Batch is the better fit for asynchronous evaluations, document processing, and other work that does not need an immediate response.

Prompt caching matters when a long, stable prefix repeats. A cache hit costs 10% of the normal input rate, while writes cost more than ordinary input: 1.25 times for a five-minute cache and two times for a one-hour cache. Measure hit rate and prefix stability before assuming caching will save money.

Limitations that benchmarks do not remove

A one-million-token window is capacity, not a reason to send every available document. At standard pricing, one million uncached input tokens alone cost $5 before any output. Retrieval, document selection, and compaction can still improve speed, cost, and focus.

Opus 5 also has product-specific constraints. Priority Tier and Anthropic’s web fetch tool are unavailable. Cybersecurity classifiers may refuse some requests, while approved participants in Anthropic’s Cyber Verification Program can access fewer restrictions. The Claude Opus 5 System Card documents safety evaluations and safeguards; teams in regulated or dual-use domains should review it before approval.

Finally, stronger self-correction does not make an autonomous action safe. Keep approval gates for payments, deletion, account changes, production deployments, and other hard-to-reverse operations.

Conclusion

For existing Opus 4.8 users, Claude Opus 5 is a compelling upgrade because the standard token price is unchanged and the model is designed for harder, longer-running work. The safest path is not a blind model-name swap: audit thinking settings, re-tune output limits and effort, remove redundant verification prompts, cap subagent fan-out, and run a measured canary.

If Opus 5 lowers the cost per successful task on your own evaluation set, expand it. If it only increases tokens or latency without improving accepted outcomes, keep lower effort or route that workload to a cheaper model.