Claude Opus 5 in 2026: Specs, Pricing, Effort Levels and How to Build With It

Claude Opus 5 is Anthropic's agentic coding and enterprise model, released July 24, 2026, priced at $5 per million input tokens and $25 per million output tokens with a 1M-token context window and 128k max output. It sits below Claude Fable 5 in raw capability and above Claude Sonnet 5 in reasoning depth, and it is the model Anthropic's own documentation tells you to start with for complex agentic coding. This guide covers what Claude Opus 5 actually changes, the three migration traps that break code carried over from Opus 4.8, and how to get from an Opus 5 prompt to a production application you own.

Quick Answer

• Claude Opus 5 launched July 24, 2026 at claude-opus-5, with 1M-token context, 128k max output, and $5 / $25 per million tokens, the same price as the Opus 4.8 it replaces.
• The effort guidance inverted. Opus 4.7 and 4.8 told you to start at xhigh for coding. Opus 5 tells you to start at high and sweep from there. Carrying old effort settings forward is the single most common migration mistake.
• Three things break on migration: thinking: {"type": "disabled"} returns a 400 error at xhigh and max, effort no longer reliably shortens responses, and reused effort sweeps are explicitly called out as invalid.
• Opus 5 has the freshest knowledge of any Claude model: a reliable knowledge cutoff of May 2026, four months newer than Fable 5 and Sonnet 5, which both sit at January 2026.
• A strong model is not a shipped app. To turn Opus 5 output into a running product with auth, payments, database and a custom domain, connect it to Totalum over MCP and let it build the project directly.

What Claude Opus 5 is

Claude Opus 5 is the current Opus-tier model in Anthropic's lineup. Anthropic positions it as "a strong agentic coding model built for long-running, multi-step work" and describes it as coming close to the frontier intelligence of Claude Fable 5 at half the price.

In practical terms, Opus tier means the model that Anthropic's own models overview recommends as the starting point for complex agentic coding and enterprise work. Fable 5 is reserved for workloads that need the highest available capability and are willing to pay double for it. Sonnet 5 is the speed and intelligence balance point at $2 / $10.

The API ID is claude-opus-5, with no date suffix. This matters: since the Claude 4.6 generation, dateless model IDs are still pinned snapshots, not evergreen pointers. Writing claude-opus-5 in your code does not silently move you to a future model.

Claude Opus 5 specs and pricing

Attribute	Claude Opus 5
Claude API ID	claude-opus-5
AWS Bedrock ID	anthropic.claude-opus-5
Google Cloud ID	claude-opus-5
Input price	$5 per million tokens
Output price	$25 per million tokens
Context window	1M tokens
Max output (Messages API)	128k tokens
Max output (Batch API, beta header)	300k tokens
Adaptive thinking	Yes
Extended thinking (thinking.type: "enabled")	No
Effort levels supported	low, medium, high, xhigh, max
API default effort	high
Reliable knowledge cutoff	May 2026
Training data cutoff	May 2026
Released	July 24, 2026

Two details in that table are easy to miss and expensive to miss.

First, the Batch API supports up to 300k output tokens on Opus 5 with the output-300k-2026-03-24 beta header, well beyond the 128k synchronous ceiling. If you are generating large artifacts (full file trees, long migrations, bulk document transforms), batching is not just a discount, it raises the hard limit.

Second, Anthropic also ships a Fast Mode for Opus 5 that runs roughly 2.5x faster at 2x the base cost. That is a genuinely different trade to the effort parameter: effort buys you less thinking for less money, Fast Mode buys you the same thinking sooner for more money.

Claude Opus 5 vs Fable 5, Sonnet 5 and Opus 4.8

Claude Fable 5	Claude Opus 5	Claude Sonnet 5	Claude Opus 4.8 (legacy)
Input / output per MTok	$10 / $50	$5 / $25	$2 / $10	$5 / $25
Context window	1M	1M	1M	1M
Max output	128k	128k	128k	128k
Reliable knowledge cutoff	Jan 2026	May 2026	Jan 2026	Jan 2026
Comparative latency	Slower	Moderate	Fast	Moderate
xhigh effort	Yes	Yes	Yes	Yes
Status	Current	Current	Current	Legacy

The row worth staring at is price against status. Opus 5 costs exactly what Opus 4.8 cost, and Opus 4.8 is now filed under Legacy in Anthropic's documentation with an explicit suggestion to migrate. There is no pricing reason to stay on 4.8. If you are still pinned to claude-opus-4-8, you are paying current-model prices for a superseded model.

For a full walk through the rest of the lineup and which model to pick per task, see our companion guide on the best Claude model for coding. That post is the selection guide across all four models; this one is the single-model reference for Opus 5.

The effort guidance changed with Claude Opus 5

This is the most consequential and least covered change in the release, and it is a reversal.

For Claude Opus 4.7 and Claude Opus 4.8, Anthropic's effort documentation is unambiguous: "Start with xhigh for coding and agentic use cases." Teams that tuned agents on 4.7 and 4.8 almost universally hard-coded xhigh into their coding paths.

For Claude Opus 5, the guidance is different: "Start with high, the default, and adjust based on your evals." Step up to xhigh for demanding coding and agentic work, up to max when a task justifies unconstrained spend, and use low and medium liberally as the primary cost and latency control wherever evals show quality holds.

Anthropic then says the quiet part out loud: "If you carried effort settings over from an earlier model, run a fresh effort sweep on your evals rather than reusing them."

The practical consequence is a bill. A team that migrates from claude-opus-4-8 at xhigh to claude-opus-5 at xhigh without re-testing is very likely spending meaningfully more tokens than the model needs on that workload, because the reasoning that required xhigh on 4.8 may land at high on 5. The migration is a one-line model ID change; the effort sweep is the work.

Effort levels on Claude Opus 5

Level	What it is for
max	Genuinely frontier problems where unconstrained token spend is justified
xhigh	Long-running agentic and coding tasks over 30 minutes with million-token budgets
high	The default. Complex reasoning, difficult coding, agentic tasks
medium	Balanced agentic work where speed and cost matter alongside quality
low	Simple, scoped tasks and subagents where latency and cost dominate

One clarification that saves confusion: effort is a behavioral signal, not a token budget. At low, Claude still thinks on genuinely hard problems, it just thinks less than it would at high on the same problem. And effort affects every token in the response, including tool calls, so lower effort means fewer tool calls, not just shorter prose.

Three things that break when you migrate from Opus 4.8

Beyond the effort sweep, three concrete behaviours differ. Each of these is a real code change, not a tuning preference.

1. You cannot disable thinking at high effort levels. On Claude Opus 5, requests that set thinking: {"type": "disabled"} while running at xhigh or max effort return a 400 error. Code that unconditionally disables thinking for latency reasons, then raises effort for a hard task, will now fail outright rather than degrade. Audit for that combination before you flip the model ID in production.

2. Effort no longer controls response length. Anthropic states directly that on Opus 5, effort controls thinking volume, not visible response length, and that "changing effort does not reliably shorten responses." If you were using low effort as a de facto brevity control, that lever is gone. Prompt for length explicitly instead.

3. Extended thinking is not available. Like Fable 5 and Sonnet 5, Opus 5 supports adaptive thinking but not the older thinking.type: "enabled" extended-thinking mode. Code paths written against Opus 4.5 or Sonnet 4.5, which do support it, need rewriting rather than re-pointing.

Here is the minimum viable Opus 5 call with explicit effort:

curl https://api.anthropic.com/v1/messages \
  -H "x-api-key: $ANTHROPIC_API_KEY" \
  -H "anthropic-version: 2023-06-01" \
  -H "content-type: application/json" \
  -d '{
    "model": "claude-opus-5",
    "max_tokens": 65536,
    "messages": [{"role": "user", "content": "Refactor this module for testability"}],
    "output_config": {"effort": "xhigh"}
  }'

Note the max_tokens. Anthropic's guidance for running Opus 5 at xhigh or max is to start at 64k so the model has room to think and act across subagents and tool calls. A small max_tokens at high effort is a common way to get truncated, expensive, useless responses.

The knowledge cutoff advantage nobody is discussing

Almost every Opus 5 review published in the last month leads with benchmarks. Almost none of them mention the row that most changes day-to-day coding output.

Claude Opus 5 has a reliable knowledge cutoff of May 2026. Claude Fable 5, the more capable and more expensive model, sits at January 2026. So does Claude Sonnet 5. Claude Haiku 4.5 is back at February 2025.

That is a four-month freshness advantage for Opus 5 over the flagship, and it is not a rounding error in a field where framework majors ship quarterly. If you are generating code against a fast-moving stack, the model with newer knowledge of that stack's current API surface can outperform a nominally smarter model with staler knowledge, because the smarter model confidently writes deprecated calls.

The honest caveat: knowledge cutoff is not capability, and for reasoning-dominant work Fable 5 remains ahead. But for the specific job of writing current, idiomatic application code against 2026 framework versions, Opus 5 being both cheaper and fresher than the flagship is a real and underrated combination. It is also a reason to prefer retrieval or documentation tooling over raw model recall regardless of which model you pick.

What the Claude Opus 5 benchmarks actually claim

Anthropic's launch announcement makes specific claims. They are worth reading precisely, because the precision is where the useful signal lives.

Benchmark	Anthropic's stated claim
Frontier-Bench v0.1	Surpasses all other models; more than doubles Opus 4.8's score
ARC-AGI 3	Roughly three times the next-best model's score
CursorBench 3.2	Within 0.5% of Fable 5's peak score, at half the cost
OSWorld 2.0	Exceeds Fable 5's result at one-third the cost
Zapier AutomationBench	About 1.5x the pass rate of the next-best model at equal cost

Read those as vendor claims with the vendor's own framing, because that is what they are. The pattern across them is consistent and plausible: Opus 5's gains concentrate in agentic, long-horizon, tool-using work rather than in single-shot question answering. That matches the positioning and it matches what the model is priced for.

Two caveats before you plan a migration around these numbers. First, several are self-reported on internal or partner evaluations rather than independent leaderboards. Second, benchmark deltas on agentic tasks are notoriously sensitive to harness design, so a 2x on a vendor harness rarely reproduces as 2x in your pipeline. For an independent view of how AI builders and agents compare on delivered output rather than model scores, see ai-agents-benchmark.com.

Your own evals remain the only benchmark that decides your migration.

Where you can use Claude Opus 5

Claude Opus 5 is available across every surface Anthropic ships:

• Claude API as claude-opus-5
• Claude Code, where it is the default agentic coding model
• claude.ai, Claude Pro and Claude Max
• Claude Cowork
• Amazon Bedrock and Claude Platform on AWS
• Google Cloud Vertex AI
• Microsoft Foundry
• GitHub Copilot, added on launch day

If you are driving Opus 5 through Claude Code, the same effort and max_tokens guidance applies, and it composes with Claude Code subagents, where running subagents at low effort while the orchestrator runs at high is a clean cost structure.

From a Claude Opus 5 prompt to a running application

Here is the gap that no amount of model capability closes on its own.

Claude Opus 5 is excellent at reasoning about an application and writing the code for it. It does not, by itself, give you a live product. Between a strong model response and a real application sit authentication, a database, file storage, payments, environment configuration, a deployment target and a custom domain. That is the work that turns a promising session into something a customer can pay for, and it is the work most people underestimate.

Totalum is an AI app builder that produces real Next.js applications with those pieces already wired: auth, payments, database, file storage, AI integrations, deployment and custom domains, with code you own and can export.

The part that matters for Opus 5 users specifically is that Totalum is drivable by an agent. Rather than treating the model and the builder as separate steps, you connect Claude Code running Opus 5 to the Totalum MCP server and let the model create and maintain the project directly through Totalum. Opus 5 does the reasoning and orchestration it is best at; Totalum materializes and maintains a production application on the other end of the tool call.

Practically, that looks like:

1. Connect the Totalum MCP server to Claude Code, Cursor, or any MCP-capable agent. Our guide to the best MCP servers covers the setup pattern.
2. Give Opus 5 the product brief rather than file-level instructions. Long-horizon planning is where this model earns its price.
3. Run the orchestrator at high or xhigh effort with a generous max_tokens, and any subagents at low.
4. Iterate against the running application, not against a local scratch directory.

The same path is available through the Totalum REST API, which is how software agencies and SaaS companies embed the builder in their own products and client workflows.

Common mistakes with Claude Opus 5

• Reusing Opus 4.8 effort settings. Anthropic explicitly says to re-sweep. Most teams do not, and quietly overpay.
• Setting a small max_tokens at xhigh. Start at 64k and tune down. Anything less truncates expensive reasoning.
• Disabling thinking at xhigh or max. That combination is a 400 error on Opus 5, not a silent fallback.
• Varying effort inside a cached conversation. Effort shapes the rendered prompt, so changing it invalidates prompt caching. Pick a level at the start of a long session and hold it.
• Reaching for Fable 5 by reflex. It costs double and has a four-month older knowledge cutoff. For most application code, Opus 5 is the better trade.
• Treating the model as the product. The model writes; something still has to run, authenticate, bill and deploy.

FAQ

What is Claude Opus 5?

Claude Opus 5 is Anthropic's Opus-tier model, released July 24, 2026, built for complex agentic coding and enterprise work. It uses the API ID claude-opus-5, has a 1M-token context window and 128k max output, and costs $5 per million input tokens and $25 per million output tokens.

How much does Claude Opus 5 cost?

$5 per million input tokens and $25 per million output tokens on the Claude API. That is identical to Claude Opus 4.8, which is now a legacy model, and half the price of Claude Fable 5 at $10 / $50. Batch API pricing and prompt caching reduce this further. Fast Mode runs roughly 2.5x faster at 2x the base cost.

Is Claude Opus 5 better than Claude Fable 5?

Not in raw capability. Fable 5 is Anthropic's most capable widely released model and is recommended for workloads needing the highest available capability. Opus 5 costs half as much, claims scores within 0.5% of Fable 5 on CursorBench 3.2, exceeds it on OSWorld 2.0, and has a newer reliable knowledge cutoff of May 2026 versus Fable 5's January 2026. For most application coding, Opus 5 is the better value.

What effort level should I use with Claude Opus 5?

Start at high, which is the API and Claude Code default. Step up to xhigh for demanding long-running coding and agentic work, and to max only when a task justifies unconstrained token spend. Use low and medium where your evals show quality holds. Do not carry effort settings over from Opus 4.7 or 4.8; Anthropic's guidance for those models was to start at xhigh, and it changed for Opus 5.

Does Claude Opus 5 support extended thinking?

No. Claude Opus 5 supports adaptive thinking but not the older extended-thinking mode set with thinking.type: "enabled". Additionally, thinking cannot be disabled at xhigh or max effort: requests that set thinking: {"type": "disabled"} at those levels return a 400 error.

Should I migrate from Claude Opus 4.8 to Claude Opus 5?

In most cases yes, because the price is identical and Opus 4.8 is now listed as legacy. The migration is a model ID change plus a fresh effort sweep on your evals, and an audit for any code that disables thinking while running at high effort levels.

How do I turn Claude Opus 5 output into a deployed app?

Connect Opus 5 to a builder that produces complete applications. With the Totalum MCP server attached to Claude Code or another MCP-capable agent, Opus 5 can create and maintain a real Next.js project with auth, payments, database, storage and a custom domain already wired, and you keep and can export the code.

Ready to build with Totalum?

Claude Opus 5 gives you the reasoning. Totalum gives you the application: real Next.js code with authentication, payments, database, file storage, deployment and custom domains, owned and exportable from day one, and drivable by your agent over MCP or REST.

Start building free at totalum.app and connect your Opus 5 agent to it in minutes.