In brief:
- Claude Opus 5 was released on July 24, 2026, and immediately became the default model in Claude Max and the most powerful option in Claude Pro.
- Anthropic positions it not as a "successor to Opus 4.8," but as a model approaching the level of Claude Fable 5 at roughly half the price.
- Key innovations include the "effort dial" mode (low/medium/high/xhigh/max), Fast mode with ~2.5x speed, a default context window of 1 million tokens, and thinking enabled by default.
- Pricing remains unchanged relative to Opus 4.8 — $5 per million input tokens and $25 per million output tokens.
- The model leads on Frontier-Bench, GDPval-AA, and ARC-AGI 3, but trails Claude Mythos 5 in cybersecurity exploit development and has known weaknesses in logical errors, race conditions, and API interaction.
Contents
What is Claude Opus 5
Claude Opus 5 is Anthropic's new flagship model in the Opus line, succeeding Claude Opus 4.8. According to the company's official description, it is a "thoughtful and proactive" model that approaches the level of Anthropic's most powerful model, Claude Fable 5, at roughly half the price (Anthropic, Introducing Claude Opus 5).
It's important to distinguish terms right away: Opus 5 is not the smartest model in Anthropic's portfolio. Above it are Claude Fable 5 (a commercial version with enhanced safeguards) and Claude Mythos 5 (a model for a trusted circle of organizations with advanced capabilities in cybersecurity and biology). Opus 5 fills the niche of a "workhorse" — a model that developers and companies use daily for real-world tasks, not for breaking benchmarks at any cost.
When was Claude Opus 5 released
Anthropic released Claude Opus 5 on July 24, 2026, simultaneously across all platforms: on Claude.ai, Claude Code, Claude Cowork, and via the Claude API (official announcement). This is the fourth model the company has released in less than two months — following Claude Mythos 5, Claude Fable 5, and Claude Sonnet 5 in June 2026 (Fortune). On the same day, Anthropic also published the model's system card — a document detailing its safety tests.
Positioning: Opus 5 is not "4.8 successor," but "almost Fable 5 for half the price"
This is a key point that is confused in most news materials. Formally, Opus 5 replaces Opus 4.8 in the lineup, but in terms of intelligence, Anthropic deliberately compares the model not to its predecessor, but to Claude Fable 5 — its top-tier model. The company explicitly states: Opus 5 approaches the level of Fable 5 but costs half as much (Anthropic).
In practice, this means the following market positioning:
| Model | Role in Portfolio | Estimated Price |
|---|
| Claude Mythos 5 | Most powerful model, available to a limited circle of partners | — |
| Claude Fable 5 | Top commercial model for the most complex and longest tasks | Higher than Opus 5 |
| Claude Opus 5 | Model for daily use, "almost Fable 5" for half the price | $5 / $25 per million tokens |
| Claude Sonnet 5 | Balance of price and quality, released earlier, in June 2026 | Lower than Opus 5 |
Therefore, the question "is it worth upgrading from 4.8" actually consists of two different questions: how much has Opus itself improved, and is this level sufficient, or does the project specifically need Fable 5. More on this in the section about Fable 5 below.
What are the main changes compared to Claude Opus 4.8
According to Claude Platform documentation, Opus 5 represents a qualitative leap over Opus 4.8, not a routine update: the model has seen the greatest gains in deep reasoning, agentic and long-form tasks, as well as in scaling test-time compute (Claude Platform Docs). Technically, this looks like this:
- The context window is now 1 million tokens by default — and also the maximum value; there is no longer a separate "short" model variant.
- The maximum output tokens per request is 128 thousand.
- Thinking mode is enabled by default — this is a change in behavior compared to how and when thinking could be disabled.
And in terms of performance metrics, the gains look like this:
- On Frontier-Bench v0.1, the model more than doubles the result of Opus 4.8, and is cheaper per task.
- On internal benchmarks for organic chemistry, Opus 5 scores 10.2 p.p. higher than Opus 4.8, and on protein mutation impact prediction tasks — 7.7 p.p. higher.
Together, these changes mean that out-of-the-box, Opus 5 is now suitable for tasks that previously required manual control over context size or a separate switch to Fable 5: large codebases, long agent sessions, and tasks where input data had to be artificially truncated to the token limit, now fit within the model's standard operating mode.
Effort dial: low / medium / high modes — how it works
The main practical feature of the release is the ability to control how much computational resources the model expends on a task. Opus 5 offers five levels of "effort": low, medium, high, xhigh, and max (Sim, model parameter reference). This is not just a speed toggle — it's a trade-off between response quality, token usage, and request cost.
According to a legal agent team that tested the model with Anthropic, Opus 5 maintains almost the same quality at lower reasoning levels, using on average 26% fewer tokens compared to Opus 4.8 at maximum reasoning (Anthropic, partner feedback). A financial modeling team reported an average accuracy increase of 9 percentage points with one-third fewer steps and tool calls, and a 60% reduction in execution time across all effort levels.
Practical takeaway: for routine tasks (simple CRUD endpoints, formatting, short refactorings), it makes sense to test low/medium — this significantly reduces cost without a noticeable loss in quality. For complex architecture design, multi-step agent planning, or tasks with a high cost of error, it's best to go directly for high or xhigh.
Fast mode: ~2.5x acceleration and double price
Separate from the effort dial, Opus 5 is available in Fast mode, which operates approximately 2.5 times faster than the standard mode. As with Opus 4.8, Fast mode costs double the base price on the Claude Platform, and in Claude Code, it is deducted via usage credits (Anthropic). This is useful for interactive scenarios — for example, when the model works in real-time with a developer, and response latency directly affects usability.
New Model Capabilities
In addition to purely performance improvements, Anthropic has demonstrated several new qualitative capabilities:
- Significantly stronger visual artifacts — interactive 3D visualizations and simulations that the model can generate from scratch.
- The ability to write its own computer-vision pipeline to extract geometry from an image when direct drawing vision is unavailable — this is how the model recreated a 3D FreeCAD model of a part from a drawing.
- Management of its own memory in long production scenarios: the model can check its own previous conclusions against new data, correct memory entries, and independently complete monitoring tasks.
- Noticeably better computer use — on the OSWorld 2.0 benchmark, the model outperforms all others at any price, and it replicates the best result of Fable 5 for less than a third of the cost.
How Coding Quality Has Changed
For developers, this is the most important section, so let's break it down in more detail. On Frontier-Bench v0.1 — a benchmark of real engineering tasks — Opus 5 outperforms all other models. On CursorBench 3.2 with maximum effort, the model lags behind Fable 5's peak result by only 0.5%, while costing half as much per task (Anthropic).
Partners who tested the model in early access specifically note not so much the "raw" code level as the quality of self-checking. In one of the examples provided by Anthropic, the model received a real bug in a popular open-source package manager and found the root cause that the official community patch missed — a competing model in the same test only fixed the external symptom.
At the same time, independent tests present a more restrained picture. A CodeRabbit review of code review quality shows that the model's weak points remain logical errors, race conditions, and incorrect API usage — precisely the types of bugs that are hardest to catch without a deep understanding of the code's runtime behavior (CodeRabbit, Claude Opus 5 Benchmarks for AI Code Review). This means that Opus 5 should be considered a strong supplementary reviewer, rather than the sole line of defense for code where concurrency or API correctness is critical.
What Has Changed in Reasoning
On ARC-AGI 3 — a benchmark that tests the ability to solve new, previously unseen tasks — Opus 5's result is three times higher than its closest competitor (Anthropic). This indicates not learned behavior, but the model's actual flexibility of thought when faced with an unconventional task formulation.
Partners in the legal and financial analysis sectors separately note an increase in quality on tasks where models previously "gave up" — vague, poorly formulated tasks with high weight for each step. The improvement here is most noticeable precisely because these are not tasks with a clear success metric, as in benchmarks, but tasks where the model's judgment itself is important.
Agentic Tasks: What Improvements Has the Model Received
On Zapier AutomationBench, which measures the ability to complete business tasks without human intervention, Opus 5 shows a success rate approximately 1.5 times higher than its closest competitor at the same task cost — and even at the lowest effort level, it completes more tasks than any other model (Anthropic). The CEO of Zapier reported that the model topped the AutomationBench leaderboard without spending more tokens than previous Claude models, and completed a full customer churn processing workflow end-to-end.
Another illustrative example is an engineer from a trading company who used Opus 5 to build a market data feed for a new exchange in one session; previous models couldn't handle this task even with a detailed plan from the engineer. Not finding a live feed for verification, the model independently built its own test harness to validate the exchange data parsing.
How the New Thinking Mode Works
The most noticeable behavioral change is that thinking is now enabled by default, and the documentation directly calls this a behavioral change regarding when and how thinking can be disabled (Claude Platform Docs). For teams migrating from Opus 4.8, this means integrations should be checked separately — the default behavior has changed, and old code that relied on disabled thinking may require updates.
In practice, this works in conjunction with the effort dial: the effort level determines how deeply the model "thinks" before providing a final answer or tool call. Here's how it looks in a real API request — the parameter is passed via output_config.effort, not through a separate thinking object as in older models:
import anthropic
client = anthropic.Anthropic()
response = client.messages.create(
model="claude-opus-5",
max_tokens=64000,
output_config={"effort": "high"},
messages=[
{"role": "user", "content": "Analyze this SQL query for an N+1 problem and suggest a fix."}
],
)
print(response.content)
In the response, the first element of the content array is now typically a thinking block, not text — this should be taken into account if your code previously rigidly expected content[0] as a ready answer:
{
"content": [
{
"type": "thinking",
"thinking": "First, I will check if the query is executed within a loop..."
},
{
"type": "text",
"text": "This query has a classic N+1 problem: ..."
}
],
"stop_reason": "end_turn"
}
An important nuance for migration: the old manual syntax thinking: {"type": "enabled", "budget_tokens": N} no longer works on Opus 5 and returns a 400 error. Additionally, thinking cannot be completely disabled with effort levels xhigh or max — the API will return the same error. And since thinking tokens are now included in the same limit as max_tokens, for high-effort tasks, you should immediately allocate at least 64,000 tokens, otherwise the response risks being cut off.
Changes in Claude API: Mid-conversation Tool Changes and Automatic Fallbacks
Concurrently with Opus 5, Anthropic released two beta features on the Claude Platform:
- Mid-conversation tool changes. Developers can now change the set of tools available to Claude mid-conversation without invalidating the prompt cache — previously, changing tools often meant re-paying for the entire context anew.
- Automatic fallbacks. Requests flagged by safety classifiers on Opus 5 (or Fable 5) can now be automatically redirected to another model instead of a complete denial of response. With automatic fallbacks enabled, a request always returns a result from the best available model rather than being blocked.
Both features are particularly relevant for production agents that previously had either a rigidly fixed set of tools for the entire dialogue or the risk of receiving a complete denial instead of a response to a borderline request.
Claude Opus 5 Pricing
The price of Opus 5 has not changed compared to Opus 4.8: $5 per million input tokens and $25 per million output tokens (Anthropic). According to the Sim model tracker, cached input tokens cost $0.5 per million (Sim). Fast mode, as mentioned above, costs twice the base price.
| Parameter | Value |
|---|
| Input tokens | $5 / million |
| Output tokens | $25 / million |
| Cached input tokens | $0.5 / million |
| Fast mode | 2x base price |
| Context window | 1 million tokens (default and maximum) |
| Maximum output tokens | 128k per request |
On which benchmarks does the model perform best
According to official Anthropic data, Opus 5 ranks first or is the most effective model on: Frontier-Bench v0.1, CursorBench 3.2 (at high effort levels), ARC-AGI 3, GDPval-AA v2, OSWorld 2.0, AutomationBench, and DeepSearchQA (Anthropic). The independent Artificial Analysis ranking on the release day placed Opus 5 first in both Intelligence Index and Agentic Index (BOVO Digital, system card analysis).
Where Claude Opus 5 falls short of competitors
Here, it's worth distinguishing between two types of weaknesses – competitive and product-specific.
Compared to other Anthropic models: Opus 5 significantly lags behind Claude Mythos 5 in cybersecurity exploit development (although results for vulnerability discovery itself are close) and in tasks related to biological research – Anthropic directly acknowledges this in the official announcement (Anthropic).
Compared to competitors from other companies: OpenAI's flagship GPT-5.6 Sol outperforms Claude Fable 5 on the Agents' Last Exam benchmark by 13.1 points, while it itself trails on SWE-Bench Pro by 15.4 points (GCN) – meaning the competitive landscape depends on the specific class of tasks being measured.
The problem with hallucinations – and why 50% doesn't mean "the model lies in half of its answers". Independent system card analysis records a hallucination rate of about 50% on test sets specifically focused on borderline, uncertain queries – tasks where the correct answer is often "I don't know" or "insufficient data" (BOVO Digital). The reason is not poor factual accuracy as such, but a behavioral shift: the model has become more willing to give a concrete answer instead of admitting uncertainty. For ordinary tasks with clear answers, this is almost imperceptible – but for scenarios where the cost of false confidence is high (medicine, legal analysis, financial calculations), I would definitely add guardrails and not rely on the model's "raw" answers without verification.
Tendency towards over-engineering. I want to specifically address the conclusion from Zvi Mowshowitz's independent system card review: Opus 5's weaker performance in some scenarios is explained by unproductive self-correction and incorrect assessment of task scale (Zvi Mowshowitz, Claude Opus 5: The System Card). In practice, I see it like this: if you ask the model to "add an endpoint to update a user's email," Opus 5 at high effort might, instead of a simple PATCH handler with validation, propose a full event-driven mechanism with an abstraction layer for future profile update types, an event queue for potential future integration with an email confirmation service, and a separate change audit service – none of which were in the task, but the model "anticipates" possible future requirements and builds infrastructure for them immediately. Formally, the code will be functional and even well-structured, but for a simple task, it's unnecessary complexity that someone will have to maintain later. At lower effort levels (low/medium), this tendency appears much less frequently – another practical argument for not setting max effort "by default just in case."
Who is Claude Opus 5 suitable for
- Teams that previously used Opus 4.8 for daily development and want a significant quality increase without an increase in API bills.
- Products with agentic scenarios (process automation, data processing, long workflows) where stability between runs is important, rather than a one-time peak result.
- Teams that require a large context window (1M tokens) by default without additional configuration.
- Projects sensitive to cost, where the effort dial allows for flexible balancing between quality and request price within a single model.
When to choose Fable 5 over Opus 5
Anthropic directly recommends Fable 5 for more advanced projects, including those that the model can perform autonomously for several days (Fortune). This is precisely the scenario where the difference between "almost Fable 5" and "Fable 5" becomes noticeable: the longer and more autonomous the chain of actions without human supervision, the more small deviations in quality accumulate.
In practice, I would focus on three specific signals that a task has outgrown the Opus 5 level.
First – the autonomy horizon. If the task fits within a single work session with a human in the loop (you set the task, the model performs several steps, you check the result and correct) – Opus 5 will handle it well, and the effort dial will also allow cost control. But as soon as it comes to a scenario of "set the task and come back for the result in a day or two" – this is a different category of risk. I described above that Opus 5 tends towards unproductive self-correction and over-engineering even in short tasks; over a long autonomous horizon, this tendency does not disappear but accumulates – the model simply has time to make more small architectural decisions that no one checks in time.
Second – the cost of rollback. If an error in the middle of an autonomous chain is costly – a broken production migration, an incorrect conclusion in a financial model that will be further included in a report – I would not rely on a model that is "almost" Fable 5, precisely because "almost" in long chains of actions means not a constant small gap in quality, but an increasing probability that somewhere at step 40 out of 60, the accumulated error will push the entire result beyond acceptable limits. Anthropic itself formulates this as a trade-off of "intelligence close to Fable 5 for half the price" – and half the price is justified exactly as long as the cost of an error is lower than the savings.
Third – the subject area. For biological research and tasks where risks are associated with long autonomous research sessions, Anthropic openly calls Mythos 5 a stronger model – and this is not a marketing disclaimer, but a direct conclusion from their own safety tests. If your task is close to this category (e.g., a research agent that independently formulates hypotheses and plans subsequent experiments without intermediate human approval), I would immediately look towards Fable 5 or even Mythos 5, rather than trying to save on Opus 5 and compensate for the difference with additional supervision – the very purpose of such a task is that it should not require supervision.
If none of these three signals apply – the task is short, the cost of error is moderate, and the subject area is not critical – Opus 5 remains the smarter choice simply because it costs half as much for comparable quality.
Is it worth upgrading from the previous version
In short: for the same price, the team gets a noticeably better model in almost all scenarios – from code to agent tasks. But the upgrade should not be done "blindly":
- Check integrations that relied on disabled-by-default reasoning – behavior has changed.
- If you used custom "skills" for Opus 4.8, retest them for compatibility.
- For tasks sensitive to logical errors, race conditions, or correct API usage, do not rely solely on Opus 5 as the sole code reviewer – maintain an additional check.
- Use the effort dial for calibration: start with medium and increase the level only where justified by the task complexity.
Frequently Asked Questions
When was Claude Opus 5 released?
On July 24, 2026, immediately on all Anthropic platforms – in Claude.ai, Claude Code, Claude Cowork, and via Claude API.
How much does Claude Opus 5 cost?
$5 per million input tokens and $25 per million output tokens – the price has not changed compared to Opus 4.8.
How does Claude Opus 5 differ from Claude Fable 5?
Fable 5 is Anthropic's most powerful commercial model, focused on the most complex and longest autonomous tasks. Opus 5 approaches its intelligence level at approximately half the price, but Anthropic still recommends Fable 5 for tasks performed autonomously over several days.
What is the effort dial in Claude Opus 5?
It is a parameter with five levels (low, medium, high, xhigh, max) that allows you to control how much computation the model spends on a task, balancing between response quality and request cost.
What is the context window in Claude Opus 5?
1 million tokens is both the default and maximum value; there is no separate "short" version of the model.
Conclusions
Having analyzed official Anthropic data and independent tests, I see Claude Opus 5 not as a cosmetic version number update, but as a model that significantly closes the gap with the top-tier Fable 5 at an unchanged price. The main practical innovations for developers – the effort dial, Fast mode, the default million-token context window, and beta API features – provide real tools for controlling cost and model behavior in production, and these are the tools I would recommend testing first before a full transition.
At the same time, I would not rush to consider Opus 5 a universal replacement for all tasks. The model is not without its weaknesses: logical errors, race conditions, incorrect API usage, and lagging behind Mythos 5 in exploit development – these are things to consider when choosing a model for a specific task, especially where the cost of error is high. My advice: start with medium effort on non-critical tasks, and make the decision about a full transition based on your own acceptance tests, not just vendor benchmarks.
