Claude Opus 5 is Anthropic's newest everyday flagship, released July 24, 2026, and it delivers near-frontier intelligence at half the cost of Claude Fable 5 — $5 per million input tokens and $25 per million output tokens, unchanged from Opus 4.8. For most people doing coding, research, or knowledge work with AI, it is now the model to reach for first: it thinks on its own, finishes multi-step tasks without leaving stubs, and checks its own work before handing anything back.
The headline feature is the effort dial — five levels (low, medium, high, xhigh, max) that let you control how hard the model thinks on a given task. Start at high (the default), step up for hard problems, step down for speed. This matters because Opus 5 turns extra effort into better results more reliably than any Opus model before it, so the dial actually does something now.
- Released July 24, 2026; available on Claude Pro, Max, Team, Enterprise, and the API
- $5/$25 per million input/output tokens — same as Opus 4.8, half the input price of Fable 5
- 1M-token context window, 128K max output (300K via Batch API)
- Five effort levels: low, medium, high (default), xhigh, max
- Thinking on by default — no need to prompt the model to reason
- #1 on Artificial Analysis Intelligence Index (60.7) and Coding Index (78.0) at launch
- API model ID:
claude-opus-5; Fast mode: ~2.5x speed at 2x base price
What makes Claude Opus 5 different from Opus 4.8?
Opus 5 is a generational capability jump at the same price. The biggest changes are in agentic coding, self-verification, and the effort system.
On benchmarks, the jump is real. Opus 5 more than doubles Opus 4.8's score on Frontier-Bench v0.1 (an agentic terminal-coding benchmark): 43.3% vs 21.1% (Anthropic, July 2026). On ARC-AGI-3 (novel problem-solving), it scores 30.2% — roughly three times the next publicly listed model, GPT-5.6 Sol at 7.8% (codersera analysis, July 2026). On SWE-bench Pro, it reaches 79.2%, just behind Fable 5's 80.0% (codersera, July 2026).
Independent testing from Artificial Analysis ranks Opus 5 #1 on both its Intelligence Index (60.7) and Coding Index (78.0) at launch, above Claude Fable 5 and GPT-5.6 Sol (CloudPrice, July 2026).
| Benchmark | Opus 5 | Opus 4.8 | Fable 5 | GPT-5.6 Sol |
|---|---|---|---|---|
| Frontier-Bench v0.1 (agentic coding) | 43.3% | 21.1% | 33.7% | 34.4% |
| ARC-AGI-3 (novel reasoning) | 30.2% | 1.52% | — | 7.8% |
| SWE-bench Pro | 79.2% | — | 80.0% | — |
| GDPval-AA v2 (knowledge work, Elo) | 1,861 | — | 1,747 | 1,736 |
Sources: Anthropic launch materials, codersera.com, Artificial Analysis via CloudPrice. Verified July 2026.
But the benchmark numbers only matter if they translate to real work. Here is what actually feels different when you use it:
- It finishes things. Older Opus models loved to leave placeholders, stubs, and half-done functions. Opus 5 completes multi-file features without leaving loose ends (Anthropic, July 2026).
- It checks its own work. Anthropic designed Opus 5 to catch its own mistakes before returning a response — opening pages in a browser at desktop and phone widths, finding hidden buttons, fixing them before handing the work back (Anthropic customer testimonials, July 2026).
- Thinking is on by default. You no longer need to tell the model to reason. It decides how much to think on its own, scaling effort to the task.
- The effort dial works. On older models, cranking effort up did not reliably help. On Opus 5, it does — more effort produces measurably better results.
How does the effort dial actually work?
The effort parameter lets you control how many tokens Claude spends thinking when responding. Think of it like a gear shift: more power when you need it, less when you do not. Five levels are available on Opus 5: low, medium, high (the default), xhigh, and max (Claude Platform docs).
| Level | Best for | Token cost |
|---|---|---|
low |
Quick lookups, formatting, simple summaries | Lowest |
medium |
Routine drafting, well-scoped functions, standard edits | Low |
high (default) |
Most intelligence-sensitive work — analysis, writing, debugging | Medium |
xhigh |
Demanding coding and agentic runs (30+ minutes) | High |
max |
Unconstrained token spending when a wrong answer is very expensive | Highest |
Anthropic's own recommendation: start with high, step up to xhigh for demanding coding and agentic work, use max only when a task justifies unconstrained token spending, and use low and medium liberally wherever your evals show quality holds (Claude Platform docs).
Here is the cost curve. Independent measurements from Artificial Analysis reported Intelligence Index scores of 56, 59, 60, and 61 from medium through max effort, with evaluation spend rising from $1,114.96 to $3,835.51 — meaning max effort cost roughly 94% more than medium while adding only a few Intelligence Index points (TokenMix Research Lab, July 2026). That makes effort a routing decision: don't run everything at max. Route 80% of traffic through low or medium, escalate the hard 20% to high or xhigh, and reserve max for tasks where a wrong answer costs more than the extra tokens.
One important detail: thinking tokens bill as output tokens. Even at unchanged per-token rates, your spend per request can rise because the model is thinking more. Compare total cost per task, not just per-token price.
How much does Claude Opus 5 cost?
Claude Opus 5 keeps the same headline pricing as Opus 4.8: $5 per million input tokens and $25 per million output tokens (Anthropic, July 2026). That is roughly half the input price of Claude Fable 5 ($10/$50).
| Model | Input / 1M tokens | Output / 1M tokens | Context window | Max output |
|---|---|---|---|---|
| Claude Opus 5 | $5 | $25 | 1M | 128K |
| Claude Opus 5 Fast | $10 | $50 | 1M | 128K |
| Claude Fable 5 | $10 | $50 | 1M | 128K |
| Claude Sonnet 5 | $2 (intro) / $3 | $10 (intro) / $15 | 1M | 128K |
| Claude Haiku 4.5 | $1 | $5 | 200K | 64K |
Source: Anthropic pricing, CloudPrice. Verified July 2026.
Two cost levers make Opus 5 cheaper in practice than the sticker price suggests:
- Prompt caching saves up to 90% on repeated context — the minimum cacheable prompt dropped from 1,024 to 512 tokens with Opus 5 (Anthropic, July 2026).
- Batch processing saves 50% on non-urgent work.
Fast mode runs at roughly 2.5x the default speed but doubles the price to $10/$50 per million tokens. Use it when wall-clock time matters more than token cost.
For subscription users: Opus 5 is the default model on Claude Max and the strongest model available on Claude Pro (Anthropic, July 2026). It is not available on the Claude Free plan — free users get Claude Sonnet 5.
Is Claude Opus 5 better than Fable 5?
It depends on the task. Opus 5 comes close to Fable 5's frontier intelligence at half the price, but Fable 5 still leads on the hardest workloads.
Opus 5 wins outright on:
- Frontier-Bench agentic coding: 43.3% vs Fable 5's 33.7%
- ARC-AGI-3 novel reasoning: 30.2% vs no published Fable 5 score
- GDPval-AA v2 knowledge work: 1,861 Elo vs 1,747
Fable 5 still leads on:
- SWE-bench Pro: 80.0% vs 79.2%
- The hardest, multi-day autonomous agent work (Anthropic recommendation, July 2026)
The practical guidance: use Opus 5 as your daily driver and keep Fable 5 in reserve for the heaviest autonomous jobs. If you are on Claude Max, you already have both — the effort dial on Opus 5 handles most of what Fable 5 used to be needed for.
Should you switch from Opus 4.8 to Opus 5?
Yes — the upgrade is low-risk and free (same price, same API contract). Anthropic designed Opus 5 as a drop-in replacement with improved performance across coding, agents, and professional work. The API model ID changes from claude-opus-4-8 to claude-opus-5, but everything else — context window, pricing, max output — stays the same.
One migration note: Opus 4.8 is now listed as "available" (legacy), and Opus 4.1 retires on August 5, 2026 (Oflight, July 2026). If you are still on Opus 4.1, migrate now.
How to set up Claude Opus 5 for daily work
The model is available through several paths. Here is what to use for each scenario:
For the Claude API
# API model ID
model = "claude-opus-5"
# Effort parameter (defaults to "high")
# Set explicitly to override the default
anthropic.messages.create(
model="claude-opus-5",
max_tokens=16000,
effort="high", # low | medium | high | xhigh | max
messages=[{"role": "user", "content": "Your prompt here"}]
)
For long-horizon agentic runs at xhigh or max effort, set a large max_tokens — starting at 64K is a reasonable default (Claude Platform docs).
Breaking change: You cannot disable thinking at xhigh or max effort. Passing thinking: {"type": "disabled"} with those levels returns an HTTP 400 (Oflight, July 2026).
For Claude Code
/model claude-opus-5
For Claude.ai (web and desktop)
If you are on Claude Pro or Max, Opus 5 is already available. On Max, it is the default model. No setup needed — just start a new conversation.
Getting started with the effort dial: 4 pro tips
- Don't touch the fancy settings on day one. Leave effort on the default (high). Just talk to it normally and see what you get. Once you are comfortable, start moving the dial — down for quick tasks, up for hard ones.
- Stop adding "please double-check your work" to your prompts. On older models, that helped. On Opus 5, it already checks its own work, and Anthropic says those old verification instructions can cause it to overcheck. Drop them and let the model run the way it was built to.
- Give it room. If you are asking for something big — especially at higher effort levels — don't cut the output short. Let it think and finish. Set generous
max_tokens. - Set up a memory system early. The real power of Opus 5 is not one clever prompt — it is a system that remembers what you built and gets better every week. When you wire Opus 5 into an agent runtime with persistent memory, every task's results get saved, and the next task starts from where the last one left off. You can build this with a free agent OS using Hermes Agent and Obsidian or a personal AI agent operating system.
What this means for you
If you are a developer or small team using AI daily, three things changed on July 24, 2026:
- You no longer need to choose between "smart" and "affordable." Opus 5 gives you near-Fable 5 intelligence at Opus 4.8 prices. The effort dial lets one model cover both cheap-and-fast and top-tier work.
- You can trust the output more. Opus 5 catches its own mistakes and finishes multi-file features without leaving stubs. You spend less time reviewing and re-prompting because the model already caught the thing you were about to point out.
- You should pair it with a memory layer. An AI model that thinks hard and finishes clean is powerful. An AI system that also remembers what it did across every project is a compounding advantage. If you have not set up persistent memory for your agents yet, now is the time — Opus 5's agentic coding quality makes the memory investment pay off faster. See our guide on self-improving agents vs static copilots for why this matters.
If you are running Claude Code and want a free or low-cost path, Opus 5 is also available through OpenRouter with a free Claude Code setup, so you can try it before committing to a subscription.
FAQ
Q: Is Claude Opus 5 out yet?
A: Yes. Anthropic released Claude Opus 5 on July 24, 2026. It is available through the Claude API (model ID claude-opus-5), Claude.ai, Claude Code, Amazon Bedrock, Google Cloud Vertex AI, and Microsoft Azure AI Foundry (Anthropic, July 2026).
Q: How much does Claude Opus 5 cost?
A: Claude Opus 5 costs $5 per million input tokens and $25 per million output tokens — the same as Opus 4.8 and half the input price of Fable 5 ($10/$50). An optional Fast mode runs at roughly 2.5x speed for twice the base price ($10/$50). Prompt caching saves up to 90% and batch processing saves 50% (Anthropic, July 2026).
Q: What is the Claude Opus 5 effort setting?
A: The effort setting lets you choose how hard Opus 5 works on a task, from low to high (the default), plus xhigh and max for the hardest jobs. Lower effort is faster and cheaper for simple work; higher effort spends more tokens to reach better reasoning. Anthropic recommends starting at high and adjusting based on your evals (Claude Platform docs).
Q: Is Claude Opus 5 better than Fable 5?
A: Not quite — but it is close, at half the price. Opus 5 leads on Frontier-Bench agentic coding (43.3% vs 33.7%) and ARC-AGI-3 (30.2%) and edges Fable 5 on knowledge work. Fable 5 still leads on SWE-bench Pro (80.0% vs 79.2%) and remains Anthropic's recommendation for the most advanced multi-day autonomous work (Anthropic, July 2026; codersera, July 2026).
Q: What is the Claude Opus 5 context window?
A: Claude Opus 5 has a 1M-token context window (approximately 555,000 words) with a 128K-token maximum output via the synchronous API, or 300K via the Batch API with a beta header. There is no input surcharge for long contexts — unlike GPT-5.6, which charges 2x above 272K tokens (Coursiv, July 2026).
Q: Can I disable thinking on Claude Opus 5?
A: You can set effort to low to minimize thinking, but you cannot fully disable thinking at xhigh or max effort levels — the API returns an HTTP 400 if you pass thinking: {"type": "disabled"} at those levels. At lower effort levels, thinking is adaptive and scales with the task (Oflight, July 2026).
Discussion
0 comments