Claude Opus 5 vs Opus 4.8: What Actually Changed

Anthropic shipped Claude Opus 5 on July 24, 2026, and priced it identically to the model it replaces. That single decision reframes the whole Opus 5 vs Opus 4.8 question. You are not weighing capability against budget, because the per-token cost is the same on every line of the pricing sheet. You are weighing capability against churn: prompts that need re-tuning, two features that quietly disappear, and a token bill that moves even though the rate card does not.

The parts that will actually cost you time sit below the benchmark table.

Opus 5 vs Opus 4.8 at a glance

Specs and pricing as on July 2026, from Anthropic's models overview, pricing page, and migration guide.

The top half of that table is identical. Everything that matters for a migration decision sits in the bottom half.

What Opus 5 does better

Anthropic's own comparison puts Opus 5 ahead of Opus 4.8 on every benchmark it published, and on several of them the margin is not subtle.

Vendor-reported figures from Anthropic's Claude Opus 5 announcement, July 2026. Not independently verified.

ARC-AGI-3 is the outlier worth pausing on. Opus 4.8 scores 1.5% on it. Opus 5 scores 30.2%. A jump of that size usually means the older model was failing the task category almost entirely rather than performing it poorly, so treat it as a capability that arrived rather than a capability that improved.

Now read the same table for where the gap closes. Legal work moves 1.3 points. HealthBench moves 2.4. If your workload lives in those columns, the honest read is that Opus 5 gives you a modest lift, not a step change, and the migration cost may outweigh it. The dramatic numbers all cluster in one place: tasks where the model has to keep a plan alive across many tool calls without losing the thread.

Anthropic's early-access partners reported in the same direction. Box measured an 8% overall improvement over Opus 4.8, rising to 17% on due diligence workflows. A financial-modeling partner reported 9 percentage points more accuracy while using roughly a third fewer turns and tool calls.

What gets worse when you switch

Anthropic's migration guide states it directly: Opus 5 supports the same feature set as Opus 4.8 with two exceptions.

Web fetch is gone. The web fetch tool works on Opus 4.8 and is not available on Opus 5. If any part of your stack has Claude pulling a URL directly, that path breaks, and you will need to route fetching through your own code or a different model. For a research agent, this is not a footnote. It is a rebuild.

Priority Tier is not supported. Organizations holding a Priority Tier commitment for reserved capacity keep it on Opus 4.8 and lose it on Opus 5. That turns the upgrade into a procurement conversation rather than an engineering one.

Three behavioral shifts round out the list. Default responses and written deliverables run longer, and dropping the effort level shrinks the thinking without reliably shortening what the reader sees, so verbosity now needs explicit prompting. Opus 5 also verifies its own work unprompted, which means the "add a final verification step" instruction you wrote for an earlier model is now actively harmful and produces over-verification. In multi-agent setups it spawns subagents more freely than Opus 4.8 did, and Anthropic recommends capping that explicitly.

Early community testing has surfaced one more wrinkle. Several users running multi-agent stacks report that Opus 4.8 still outperforms Opus 5 at planning tasks for other agents to execute, even as Opus 5 wins the execution side. This is community-reported, not verified by Anthropic, but it aligns with the documented increase in subagent delegation: a model that wants to do the work itself may be less disciplined about writing the plan for someone else to follow.

None of these throws an error. Your logs will look clean while your output drifts.

The price is identical, so why is your bill higher

Because thinking runs by default now. On Opus 4.8, a request without a thinking field ran without thinking. Send that same request to Opus 5 and it reasons first, then answers, and you pay for both at the same $25 per million output tokens. Independent testers have measured roughly three times the output tokens on small tasks after the switch, which sounds alarming until you look at what happens per completed task rather than per call.

Anthropic's effort-scaling chart plots score against dollars per attempt, and the two curves never cross. Reading approximate values off the plot, Opus 5 at its lowest effort setting scores around 26% at roughly $5.50 per attempt. Opus 4.8 at its highest effort setting reaches about 19% at roughly $17. The cheapest configuration of the new model beats the most expensive configuration of the old one at about a third of the cost.

Match the spend instead and the gap is starker. Around $11.50 per attempt, Opus 5 sits near 39% while Opus 4.8 sits near 15.5%.

Both findings are true at once, and the apparent contradiction dissolves once you pick the right denominator. More tokens per call, fewer calls per finished job. Budget on cost per accepted task, not tokens per request, and the picture inverts.

One detail on the chart deserves a warning. Opus 5's max effort point sits slightly below its xhigh point while costing more. Anthropic's documentation says the same thing in prose: max can show diminishing returns and tends toward overthinking on simpler tasks. Do not turn effort to max on migration day and assume you have bought the best version of the model.

Two caveats on this data. These are internal Anthropic runs on the mini-SWE-agent harness with a GKE backend, averaged over five attempts per task. And per Anthropic's own footnote, Opus 4.8 served as the fallback on safety-classifier refusals for Opus 5, so a slice of the Opus 5 line was in some runs served by Opus 4.8. The gap is far too wide for that to explain it away, but the methodology is worth knowing before you quote the numbers.

What changes if you use Claude Pro or Max

Not everyone using these models is editing JSON request bodies. Plenty of people meet them through a subscription, and the changes there are real but different.

As of July 2026, Opus 5 is the default model on Claude Max and the strongest model available on Claude Pro. Opus 4.8 stays selectable, which is your rollback path and requires no code at all.

There is also a context-window upgrade that only shows up in conversations on Claude.ai. On paid plans, Opus 4.8 runs at 500K tokens in chat. Opus 5 runs at 1M. If you work in long conversations with large file uploads, that is a meaningful jump, and it is easy to miss because the API and Claude Code context window has been 1M for both models all along.

The fallback behavior is easy to get wrong. Opus 5 ships with cybersecurity classifiers, and in Claude.ai, Claude Code, and Claude Cowork, any flagged request falls back to Opus 4.8 by default. On the API, that same fallback is opt-in through the fallbacks parameter. The distinction matters: default-on in the apps, opt-in on the API.

What happens after that fallback fires matters too. In the Claude apps, the model picker stays on the less capable model for the rest of that conversation, and switching back can trigger the same fallback again if the original request is still in the thread. For long sessions, the per-response model label is more reliable than the picker state at the top. Anthropic documents this in its help center article on automatic model switching.

Anthropic also expects those classifiers to fire around 85% less often on Opus 5 than on Fable 5, and biology requests blocked on Fable 5 now route to Opus 5 rather than Opus 4.8.

What changes if you build on the API

The swap itself is one line, from claude-opus-4-8 to claude-opus-5. There is exactly one hard break to audit.

Sending thinking: {"type": "disabled"} alongside effort xhigh or max returns a 400 error. Opus 4.8 accepted that combination. Validation runs per request, so raising effort mid-conversation gets rejected even when earlier turns went through. You have two fixes: drop the thinking field and keep the effort level, or keep thinking disabled and lower effort to high or below.

If you keep thinking disabled, Anthropic warns that the model can occasionally write a tool call into its visible text instead of emitting a tool_use block, or leak internal XML tags. In an agentic loop that text stays in the history and contaminates later turns, which is the expensive kind of bug. Prefer lower effort with thinking on.

Three smaller changes are worth a grep. Because max_tokens caps thinking plus response together, limits tuned on a thinking-free Opus 4.8 workload can now truncate answers, and Anthropic suggests starting at 64k for xhigh or max. Prompt caching now kicks in at 512 tokens rather than 1,024, so short system prefixes that never cached before will start creating entries. And mid-conversation tool changes are available in beta behind the mid-conversation-tool-changes-2026-07-01 header, letting you add or remove tools between turns without invalidating the cache.

Both models use the same tokenizer, so token counts for the same text are roughly unchanged. Any increase you see is behavioral, not tokenization.

When staying on Opus 4.8 is the right call

Four situations, and none of them is nostalgia.

You use web fetch. Until you have replaced it, Opus 5 cannot run that workload.

You hold Priority Tier. Reserved capacity does not follow you to Opus 5.

Your integration must run thinking-free above high effort. Opus 4.8 is currently the only model that serves both conditions.

Your prompts are heavily tuned and you cannot re-baseline this cycle. A frozen Opus 4.8 lane beats a half-migrated Opus 5 lane running instructions written for a different model.

Retirement pressure is not on that list. Anthropic's deprecation page lists both as Active, with tentative earliest retirement of May 28, 2027 for Opus 4.8 and July 24, 2027 for Opus 5, plus a commitment to at least 60 days' notice.

The bottom line on Opus 5 vs Opus 4.8

Migrate, in most cases, and do it deliberately rather than globally. Opus 5 leads Opus 4.8 across Anthropic's entire published benchmark set at identical pricing, and the effort curve suggests it delivers more per dollar at every setting on the ladder. For agentic coding, long-horizon tool use, and computer-use workloads, the gap is large enough that staying put has a cost.

Audit three things first. Search your codebase for web fetch, check whether you hold Priority Tier, and grep for thinking: disabled near any xhigh or max effort setting. Then strip the verification instructions you wrote for older models, because leaving them in is where the token increase turns into a bill. Keep Opus 4.8 as the rollback lane while you measure, since nothing is forcing you off it before mid-2027.

All of that assumes you are the person wiring a model into an application: auditing request bodies, re-baselining effort sweeps, rewriting prompts tuned for a model that no longer behaves the same way.

If your goal is shipping working software rather than managing that migration, Emergent handles the model layer for you. Opus 5 is already available on the platform alongside the rest of the supported Claude, OpenAI GPT, and Google Gemini models. You describe the product; Emergent handles the model selection, the backend, and the deployment. What comes out is a full-stack application with a real backend, real integrations, and real code you own. Emergent also supports MCP connectors for wiring those applications into the tools a business already runs on.

Start Building on Emergent.