Claude Opus 5 Review: API Pricing, Specs, and Opus 4.8 Comparison

Anthropic released Claude Opus 5 on July 24, 2026. The interesting part is not another collection of first-place benchmark claims. Anthropic is trying to make the Opus tier more practical: keep the Opus 4.8 API rates, but move coding agents, tool use, and professional work closer to the much more expensive Fable 5.

At first glance, that sounds like a free upgrade. At the task level, it is not so simple. Opus 5 decides when to think, tends to write longer answers, and is not especially fast in independent testing. The rate card may be unchanged while the cost and wait for a completed task move considerably.

We checked Anthropic’s announcement and pricing material against Artificial Analysis, a hands-on ComputingForGeeks comparison, and the first two days of Hacker News discussion. Vendor benchmarks, customer quotes, and independent tests are kept separate because they answer different questions. Figures in this article were checked on July 26, 2026.

Our take: Opus 5 belongs on the shortlist when the work is difficult and a failed attempt is costly. For high-volume classification, routine rewriting, or small code completions, it is usually too slow and too expensive. The 1M context and 128K output limit look impressive, but cost and time per successful task are what should decide an upgrade.

Key Takeaways

• claude-opus-5 supports a 1M-token input and up to 128K output. Adaptive thinking is the default, with effort controls from low to max.
• Standard API rates remain $5 per million input tokens and $25 per million output tokens. Fast mode is roughly 2.5 times faster at twice the price.
• The max-effort configuration currently tops the Artificial Analysis Intelligence Index, but produces only about 52.6 tokens per second and uses far more output than the median model.
• In one small, three-task DevOps comparison, Opus 5 cost roughly twice as much as Opus 4.8. The sample is tiny, but it makes the practical point: identical token rates do not guarantee an identical bill.

What Is Claude Opus 5?

Think of Opus 5 as Anthropic’s model for difficult everyday work: close to Fable ambitions, but priced in the Opus tier. Anthropic calls it “thoughtful and proactive.” In practice, that means it is less eager to hand back the first plausible answer and more likely to inspect the environment, verify a result, and try again after a failure.

That difference is easier to see in work such as:

• finding a problem in a large repository, editing multiple files, running tests, and continuing until the change passes;
• using terminals, computer interfaces, or business tools to complete a multi-step workflow;
• analyzing contracts, financial material, enterprise documents, and complex tables;
• performing scientific or analytical work that benefits from repeated verification;
• understanding large collections of code, logs, specifications, images, and documents within a 1M-token context.

Anthropic still has not said how many total or active parameters the model has, whether it uses a particular mixture-of-experts design, or how large its training set is. The specifications we can verify are the context, output limit, modalities, and reasoning controls. Weight-count estimates circulating online remain guesses.

Claude Opus 5 Specifications and Features

The two headline limits are easy to overread. A 1M context means the request can hold that much material; it does not promise perfect recall from a million-token prompt. Long-document systems still need retrieval, chunking, and citation checks. Dumping an entire knowledge base into one call merely hides the retrieval problem inside the model.

Likewise, 128K is a hard ceiling, not a sensible target. Thinking and visible text draw from the same response budget, so production requests still need a realistic max_tokens. Otherwise, a wandering agent can produce an expensive report before anyone notices it has gone off course.

There is also a migration trap. Opus 5 uses adaptive thinking and does not accept the old fixed-budget thinking: {"type": "enabled"} shape. Disabling thinking has limits when combined with the highest effort settings, so replacing the model ID alone is not a safe migration plan.

How Much Does Claude Opus 5 Cost?

Anthropic did not use the release to raise the base rate. Standard API pricing is:

Consider a standard request with 100,000 input tokens and 10,000 output tokens, excluding tools and retries:

0.10 × $5 + 0.01 × $25 = $0.75

If all 100,000 input tokens are cache hits, the same output would cost about:

0.10 × $0.50 + 0.01 × $25 = $0.30

The same request would be about $0.375 as an eligible Batch job and about $1.50 in Fast mode. Cache behavior, whether the work can run asynchronously, and whether low latency is genuinely valuable often matter more than the headline fact that the model rate did not change.

Price the Task, Not Just the Tokens

Output tokens cost five times as much as input. One more reasoning pass, tool call, or verification loop can make an “unchanged-price” request more expensive. The opposite can also be true: if Opus 5 succeeds on the first attempt and avoids an engineer’s rework, the extra inference cost may be money well spent.

That is why a useful evaluation tracks total cost per successful task: input, cache hits, thinking, visible output, tool calls, retries, wall-clock time, and human correction. A procurement decision based only on two numbers from a pricing table misses most of the bill.

The Upgrade Is About Finishing Work, Not Better Chat

Coding Agents: Does the Model Check Its Own Work?

“Verify” and “iterate” appear repeatedly in Anthropic’s description of Opus 5. For a coding agent, that matters more than producing an elegant first draft. When a test fails, can the model understand the failure, revisit its assumption, and finish the job?

Migrations, cross-file refactors, and multi-tool workflows benefit from that habit. They also consume more steps and tokens. Production agents still need hard limits on steps, spend, and writable scope, with human approval for deletion, deployment, payments, or other irreversible actions.

Professional Work: Better Output Still Needs Review

GDPval-AA, AutomationBench, and DeepSearchQA receive prominent space in the launch material. Anthropic is plainly selling beyond developers to finance, legal, and enterprise-research teams.

The likely gain is faster research, drafting, and cross-checking—not the removal of final review. Contract terms, financial figures, medical claims, and security decisions still need traceable sources and an accountable person signing off.

Science and Visual Artifacts

Anthropic also reports gains over Opus 4.8 in structural biology, organic chemistry, and bioinformatics. The launch page shows an interactive wind tunnel and cell visualization, but “visual output” needs clarification here: the model wrote code for interactive pages. The standard API still returns text; it does not directly generate images.

Choosing Effort and Fast Mode

Higher effort is not automatically better. Formatting, routine rewriting, and small maintenance changes are cheaper at low effort; difficult debugging, architecture, and long-running agents are where higher settings earn their keep. Fast mode is a simpler trade: pay twice the rate for about 2.5 times the inference speed. It makes sense only when waiting has a measurable cost.

The Official Numbers Look Great—But They Are Still Vendor Numbers

Anthropic assembled an impressive launch scorecard for Opus 5. These figures show what the company wants the model to be good at, but they are not third-party certification: the benchmark mix, effort, tools, and budget were selected or cited by the vendor.

The launch post also includes a rare losing result. Opus 5 comes close to the restricted Mythos 5 at finding vulnerabilities, but remains substantially worse at developing exploits. That tells buyers more about the model’s boundary than another first-place chart would.

Customer Stories Are Worth Reading, but They Are Not Independent Tests

The launch page packs in 24 customer and partner statements. Box provides the clearest named result: CTO Ben Kus says its internal evaluation put Opus 5 8% ahead of Opus 4.8 overall, including an 11% gain in data analysis and 17% in due diligence.

Other cases selected for the launch page report that:

• a financial-modeling team averaged 9 percentage points higher accuracy across effort levels, with one-third fewer turns and tool calls and 60% less time;
• a legal-agent team maintained similar quality at lower reasoning levels while generating 26% fewer tokens on average than Opus 4.8 at max reasoning;
• the Kiro team found that Opus 5 maintained longer, multi-step coding work and clarified feature and bug-fix requirements more effectively.

That is more useful than “we love the model,” but Anthropic still selected these success stories from its early customers. Full test sets, failures, significance, and a common cost method are missing. Treat them as reasons to run a trial, not as reproduced findings.

Independent Tests Show the Other Side of the Tradeoff

Artificial Analysis: First in Intelligence, 121st in Speed

On July 26, Artificial Analysis scored Claude Opus 5 (Adaptive Reasoning, Max Effort) at 61 on its Intelligence Index, first among the 190 models shown on the page. Scroll further, though, and the cost of that capability becomes obvious. Output measured about 52.6 tokens per second, only 121st for speed, and the evaluation consumed roughly 100 million output tokens against a class median near 63 million.

Artificial Analysis spent approximately $3,835.51 to complete the Intelligence Index. Opus 5 at max effort is clearly near the capability frontier; it just does not get there by being fast or frugal. The first-place label is also a July 26 snapshot of one suite and one configuration, not a permanent title.

Three DevOps Tasks: More Complete Code and a Larger Bill

On July 25, ComputingForGeeks gave Opus 5 and Opus 4.8 the same three tasks at default settings: write an OpenTofu VPC module, repair a Kubernetes Deployment seeded with six defects, and produce a Bash log-rotation script. The outputs went through tofu validate, tflint, kubeconform, and shellcheck rather than a subjective review.

The three Opus 5 runs cost about $0.14 in total, versus roughly $0.07 for Opus 4.8. The newer model wrote more complete infrastructure code, but it was slower and longer. On the Kubernetes task it also ignored the explicit request to return raw YAML without Markdown fences.

Three prompts are nowhere near enough to declare an overall winner. They do expose something the pricing page cannot: at the same token rate, extra thinking and longer answers can double the cost of a task.

What Developers Like—and What They Are Complaining About

At verification time, the Hacker News launch thread had about 1,747 points and 1,266 comments available through the Algolia API. It is not a satisfaction survey, but it does reveal what early users noticed first.

The positive stories are about genuinely hard problems. One commenter said Opus 5 triaged and fixed a Linux kernel issue in under 30 minutes after Sonnet 5 and Opus 4.8 had failed for hours. Another, after a few hours of use, thought its coding was slightly better than Fable 5. These are anecdotes, not reproducible tests, but they explain why some developers will tolerate the price.

The complaint is remarkably consistent: it is slow. Several users describe it as slower than the already unhurried Opus 4.8; one said even routine refactoring felt delayed after using the model for most of a day. That lines up with the Artificial Analysis throughput result.

There is no community consensus on value. A practical pattern is emerging instead: reserve Opus for the hard planning and implementation, then hand routine execution to Sonnet or a cheaper model. That preserves access to the model’s ceiling without forcing every request to absorb Opus latency and cost.

What Work Should Go to Opus 5?

Good Places to Start

• Long-running coding agents: cross-file refactors, difficult debugging, migrations, and test repair;
• High-value professional work: contracts, finance, enterprise content, and research material;
• Multi-tool automation: workflows that must plan, call tools, inspect results, and iterate;
• Large repositories and document sets: workloads that genuinely need 1M context and can add retrieval and validation;
• Quality-first offline tasks: jobs that can tolerate latency or use Batch to reduce cost;
• High-value interactive work: cases where quality and latency both matter enough to test Fast mode at twice the rate.

Workloads That Should Probably Stay Put

• high-volume, low-value classification, extraction, rewriting, or support responses;
• applications with strict latency requirements but no need for frontier reasoning;
• agents that produce long output without an explicit acceptance test;
• generated YAML, shell, SQL, or infrastructure configuration executed without validation;
• legal, medical, financial, or production changes without human review;
• requests that use the 1M limit as a reason to submit an entire unfiltered knowledge base.

What Changes When Migrating from Opus 4.8?

A migration is not difficult, but it is more than changing a string. Check at least these eight items before rollout:

1. Change the model ID to claude-opus-5.
2. Remove the old fixed-budget thinking: {"type": "enabled"} request shape and use adaptive thinking as supported by the current interface.
3. Revisit max_tokens, because thinking and visible output share the response budget.
4. Evaluate low, medium, high, xhigh, and max as separate configurations.
5. Record usage.output_tokens_details.thinking_tokens to measure reasoning overhead.
6. Keep JSON Schema, syntax checks, tests, Markdown-fence cleanup, and approval for dangerous actions.
7. Compare cost per successful task rather than input and output rates alone.
8. If default latency is unacceptable, test Fast mode on real traffic instead of treating “2.5× speed” as a guaranteed end-to-end improvement.

Anthropic also shipped two betas: tools can change mid-conversation without invalidating the prompt cache, and safety-classifier triggers can automatically fall back to another model. Both are useful for long-running agents. Before relying on them, confirm that the account has access and make sure logs reveal which fallback model ran and how it was billed.

Calling Claude Opus 5 through UnifyLLM

If the model is enabled for your UnifyLLM account, you can begin with the unified OpenAI-compatible interface:

curl https://api.unifyllm.top/v1/chat/completions \
  -H "Authorization: Bearer <YOUR_UNIFYLLM_API_KEY>" \
  -H "Content-Type: application/json" \
  -d '{
    "model": "claude-opus-5",
    "messages": [
      {
        "role": "user",
        "content": "Review this change plan, identify high-risk dependencies, and provide verifiable rollback steps."
      }
    ],
    "max_tokens": 4096
  }'

The example does not assume a thinking field because compatibility protocols, SDKs, and upstream channels may map effort, caching, and tools differently. Check the UnifyLLM API reference alongside current Anthropic documentation before using provider-specific controls. Account availability, group multipliers, and actual billing appear on the model pricing page and in the console.

For the comparison, use work your team already does rather than inventing a handful of impressive-looking prompts. First-pass success, serious errors, format compliance, time to first token, total time, thinking tokens, tool calls, and final cost all affect the production experience. The Getting Started guide and Integrations cover the initial connection.

Should You Upgrade from Opus 4.8?

If one failed attempt burns half a day of engineering time, or the model has to inspect a complex environment and carry the work to completion, the upgrade is easy to justify testing. In that setting, saved rework matters more than the per-million-token rate.

If the work is clearly specified, easy to verify, and high volume, there is little reason to send all of it to Opus. Let Sonnet or a smaller model handle ordinary calls, then escalate failures, low-confidence cases, and valuable tasks. Use Batch for asynchronous work and prompt caching for repeated large contexts. Fast mode belongs only where an extra minute of waiting has a real business cost.

We would not replace Opus 4.8 everywhere on day one. Run an A/B test on real work, then route by difficulty. That is far more reliable than making a fleet-wide decision from launch-day charts.

Summary

Opus 5 is a substantial upgrade: the token rates stay put, the ceiling on difficult work rises, and the agent is more willing to finish what it started. It also thinks, verifies, waits, and writes—and each of those can add time and money.

This is a model for tasks where getting the answer wrong is expensive, not an argument that every request needs the strongest available model. Measure success rate, wait time, and the complete bill on your own workload before deciding where it belongs. For Opus 5, those numbers matter more than any launch leaderboard.

Frequently Asked Questions

Has Claude Opus 5 Been Officially Released?

Yes. Anthropic’s official publication record shows a July 24, 2026 release, with claude-opus-5 as the API model ID. Availability may still differ across cloud platforms, regions, and accounts.

What Is Claude Opus 5 API Pricing?

Standard pricing is $5 per million input tokens and $25 per million output tokens, the same as Opus 4.8. Prompt-cache hits can reduce input cost by 90%, Batch reduces token rates by 50%, Fast mode costs twice the standard rate, and US-only inference costs 1.1 times the standard rate.

What Are the Context and Output Limits?

The input context is 1,000,000 tokens and the maximum single response is 128,000 tokens. A 1M capacity does not guarantee perfect recall across one million tokens, so production systems should still use retrieval, chunking, citations, and validation.

How Many Parameters Does Claude Opus 5 Have?

Anthropic has not disclosed the total or active parameter count or the model’s MoE structure. Publicly verifiable information covers context, output, modalities, effort, thinking, and pricing. Community estimates should not be presented as official specifications.

Does Adaptive Thinking Cost Extra?

Thinking tokens are billed at the output-token rate and consume the response budget. They may improve success on difficult tasks but can also increase cost and latency. Evaluate thinking tokens together with final pass rate and total time.

Is Claude Opus 5 Cheaper than Opus 4.8?

Their standard token rates are equal, but task costs can differ. In one independent three-task sample, Opus 5 cost about twice as much because of thinking and longer output. On work where it prevents failures and rework, its cost per successful task may still be lower.

Is Claude Opus 5 Better than Fable 5?

Sometimes; not always. Anthropic positions Opus 5 as approaching Fable 5 capability at roughly half the price. Opus wins some coding, automation, and cost-performance comparisons, while gaps remain in areas such as cybersecurity. Compare them with the same effort, tools, budget, and tasks or the result is meaningless.

References

• Anthropic: Introducing Claude Opus 5
• Anthropic: Claude Opus product page
• Anthropic Docs: Models overview
• Anthropic Docs: What’s new in Claude Opus 5
• Anthropic Docs: Pricing
• Artificial Analysis: Claude Opus 5 Performance & Price Analysis
• ComputingForGeeks: Claude Opus 5 Released—Benchmarks, Pricing and API Changes
• Hacker News: Claude Opus 5 launch discussion
• Hacker News Algolia API: Discussion data

Facts were checked on July 26, 2026. The Anthropic documentation site redirected this verification environment because of regional restrictions. The release date, positioning, standard pricing, Fast mode, caching, Batch, and US-only rates were rechecked on directly accessible official Anthropic pages. The 1M/128K limits, adaptive thinking, and effort capabilities were cross-checked through a published live Models API response and independent specification pages.