Claude Opus 5
Anthropic's flagship Opus model, released July 24, 2026 — near-frontier intelligence at half the price of Claude Fable 5, with thinking on by default and a full low-to-max effort ladder.
1M tokens · Text / Vision / Code · Prompt cache
Claude Opus 5 is Anthropic's current flagship Opus model, released July 24, 2026 as the successor to Opus 4.8. It comes close to Claude Fable 5's frontier intelligence at half the price, and on several coding and knowledge-work evaluations — Frontier-Bench v0.1 and GDPval-AA v2 among them — it is ahead of Fable 5 outright. Pricing is unchanged from Opus 4.8 at $5 input / $25 output per 1M tokens, so it is a drop-in upgrade for agents already running on the Opus tier.
The behavior changes are worth knowing before you switch. Thinking is on by default, where Opus 4.8 ran without thinking unless you asked for it; effort runs the full low to max ladder; and the raw chain of thought is never returned. On VM0 it carries a ×2 credit multiplier, the same as Opus 4.8, and is part of the built-in model lineup for new organizations.
What is Claude Opus 5?
Released July 24, 2026 · Anthropic's flagship Opus tier — below Fable 5 on absolute ceiling, ahead of it on several agentic and knowledge-work evaluations, at half the token price. The default choice for hard agent work on VM0.
Anthropic released Claude Opus 5 on July 24, 2026 across the Claude API, the Claude apps, Amazon Bedrock, Google Cloud, and Microsoft Foundry; VM0 added it as a built-in model on July 25, 2026. It succeeds Opus 4.8 at identical pricing and roughly doubles its Frontier-Bench v0.1 score, which is the headline claim: substantially more capability for the same cost per token.
It is positioned as the workhorse for agentic coding and enterprise knowledge work — multi-file features, larger refactors, long autonomous runs, and end-to-end deliverables — rather than as the absolute capability ceiling. Fable 5 still leads on the harder SWE-bench Pro set and on factual breadth; Opus 5 wins on cost per completed task, which is the number that shows up on the invoice once agent traffic scales.
Two API behaviors changed relative to Opus 4.8. Thinking is on by default, so a request that configures nothing now reasons before answering, and turning thinking off is permitted only at effort high or below. The minimum cacheable prompt also dropped to 512 tokens, so short system prompts that previously could not be cached now are. Opus 5 also ships with elevated cybersecurity safeguards, and requests in prohibited or high-risk areas can come back as a refusal.
What's notable about Claude Opus 5
Headline architecture and capability features.
Opus 5 has a 1M-token context window — the maximum is also the default — and up to 128K output tokens, with a May 2026 knowledge cutoff. Thinking is on by default and reasoning depth is set through effort levels low, medium, high, xhigh, and max; the raw chain of thought is never exposed, only a readable summary. Assistant-message prefill is not supported, and sampling parameters (temperature, top_p, top_k) are rejected — steer behavior with prompting instead. Inputs span text, high-resolution vision, and code.
Specs at a glance
Claude Opus 5 benchmarks
Vendor-reported scores from Anthropic's launch materials. Third-party trackers report different figures for the same benchmarks depending on dataset, effort setting, and agent scaffold, so treat the headline numbers as configuration-dependent.
Claude Opus 5 pricing
Provider list price, per 1M tokens.
How Claude Opus 5 behaves in practice
Observed behaviour from production agent runs.
Cost per completed task
The clearest reason to pick Opus 5 over Fable 5: comparable outcomes on agentic work at roughly half the token price. On Zapier's AutomationBench its pass rate is around 1.5× the next-best model for the same cost per task, and even at its lowest effort setting it passes more tasks than any other model.
Thinking on by default
Unlike Opus 4.8, a request that omits the thinking parameter reasons before answering. Budget for it: max_tokens caps thinking and answer together, so a limit tuned for a thinking-off workload can truncate. Disabling thinking is allowed only at effort high or below.
Effort is the main cost lever
low and medium are unusually strong on this model and often match what previous flagships produced at their top settings. Start at high — xhigh for the hardest coding and agentic runs — then sweep downward and keep the cheapest level your evaluations still accept.
Verifies its own work
Opus 5 checks its output without being told to, so verification instructions carried over from older models tend to produce redundant passes. It also writes longer responses by default; a short conciseness instruction works better than lowering effort.
Best agent tasks for Claude Opus 5
The multi-file feature that lands in one run
Give it the full specification up front and let it run. Opus 5's gains over Opus 4.8 are largest on hard, multi-step engineering work — larger refactors, end-to-end features — and smallest on single-turn edits, so point it at the top of your difficulty range.
The overnight agent you can afford to leave running
Long autonomous runs are where the price difference compounds. At Fable-adjacent quality and half the token cost, a nightly triage, migration, or research agent costs roughly half as much per completed task.
The business workflow that has to finish, not just respond
Multi-step tasks with tools and a real end state — filing tickets, updating spreadsheets, reconciling records. Anthropic's AutomationBench and OSWorld results point at the same thing: it finishes end-to-end tasks that other models drop partway.
When to skip Claude Opus 5
Skip Opus 5 for high-volume, latency-sensitive, or straightforward work — Sonnet 5 covers routine agent loops at half the multiplier, and GPT 5.6 Luna is cheaper still. Reach for Fable 5 instead when a task sits at the ceiling of your difficulty range or leans on SWE-bench-Pro-style hard coding and broad factual recall. Security research and other high-risk domains can trip Opus 5's safeguards and return a refusal.
Claude Opus 5 vs other models
Claude Opus 5 vs Claude Fable 5
Fable 5 is the higher ceiling; Opus 5 is the better deal. Opus 5 leads on Frontier-Bench (43.3% vs 33.7%), GDPval-AA v2, and SWE-bench Verified, while Fable 5 keeps SWE-bench Pro and factual breadth. Opus 5 costs half as much ($5/$25 vs $10/$50) and carries a ×2 VM0 multiplier against Fable 5's ×4 — route the hardest steps to Fable 5 and everything else here.
Claude Opus 5 vs Claude Opus 4.8
Same price, same tier, materially more capable: Opus 5 roughly doubles Opus 4.8's Frontier-Bench score and adds seven points on SWE-bench Verified. Migrating is a model-ID swap plus prompt re-tuning — thinking is now on by default, so re-check max_tokens, and drop any "verify your work" instructions, which now cause redundant passes.
Claude Opus 5 vs Claude Sonnet 5
Sonnet 5 (×1) is the cost-efficient default for routine agent loops; Opus 5 (×2) is what you promote a task to when it stops finishing reliably. The common pattern is Opus 5 orchestrating and Sonnet 5 handling the fan-out work.
Claude Opus 5 vs GPT 5.6 Sol
Both are top-tier agentic models. Opus 5 leads Sol on Frontier-Bench (43.3% vs 34.4%); Sol leads on DeepSWE v1.1 (72.7% vs 68.8%). On VM0 that is a routing decision rather than a procurement one — run the same agent on both and compare.
Bottom line: should you use Claude Opus 5?
The default flagship on VM0. Opus 5 gets close enough to Fable 5's ceiling that price becomes the deciding factor on most agentic work, and it is a drop-in upgrade for anything already running on Opus 4.8. Keep Sonnet 5 for routine loops and reserve Fable 5 for the few tasks that genuinely need the top of the range.
Frequently asked questions
When was Claude Opus 5 released?
Anthropic released Opus 5 on July 24, 2026, across the Claude API and the Claude apps as well as Amazon Bedrock, Google Cloud, and Microsoft Foundry. VM0 added it as a built-in model on July 25, 2026 (model id claude-opus-5).
How much does Claude Opus 5 cost?
$5 per 1M input tokens and $25 per 1M output — unchanged from Opus 4.8, and exactly half Claude Fable 5's $10/$50. Cache reads are $0.50 per 1M. On VM0 it carries a ×2 credit multiplier.
Is Claude Opus 5 better than Claude Fable 5?
On several evaluations, yes: Opus 5 leads Fable 5 on Frontier-Bench v0.1, GDPval-AA v2, and SWE-bench Verified. Fable 5 still leads on the harder SWE-bench Pro set and on factual breadth, and it remains the higher ceiling for the hardest reasoning. For most agentic work Opus 5 wins on cost per completed task.
Is thinking on by default on Claude Opus 5?
Yes. A request that omits the thinking parameter reasons before answering — a change from Opus 4.8, where thinking was off unless requested. Depth is controlled with effort (low through max), and thinking can be disabled only at effort high or below. The raw chain of thought is never returned, only a readable summary.
What changes when migrating from Opus 4.8?
The model id, and then prompt tuning. Thinking on by default means max_tokens now covers reasoning as well as the answer; verification instructions should be removed rather than rewritten, since Opus 5 self-verifies; and responses run longer by default, so add an explicit conciseness instruction if length matters.
Alternatives
Using Claude Opus 5 on VM0
Two ways to access Claude Opus 5 on VM0
VM0 supports Claude Opus 5 as a Built-in model billed in VM0 credits, and through bring-your-own with a Anthropic API key. The Built-in path uses VM0 Managed routing and the credit multiplier explained below; the bring-your-own path bills you directly with the upstream vendor and skips the VM0 credit conversion entirely.
VM0's recommendation
VM0 positions Claude Opus 5 as a core agent model, recommended alongside Claude Opus 4.7, Claude Opus 4.6, and Claude Sonnet 4.6 for the steps that drive the actual outcome of an agent run. These are the models we'd pick for the orchestrator role, for code-touching agents, and for any step where a wrong answer is expensive.
Credits and the ×2 multiplier
Every Built-in model on VM0 is priced as a multiple of Claude Sonnet 4.6, which sits at the ×1 credit baseline. Claude Opus 5 bills at ×2 credits. The multiplier is what shows up on your VM0 invoice; the vendor list price in the pricing table above is what the upstream provider charges before VM0 converts it into credits.
Claude Opus 5 bills at ×2, which means a step here costs 2× the credits of an equivalent step on Sonnet 4.6 (the ×1 baseline). It's a premium tier on VM0, so the cost-effective pattern is to default to a cheaper model and route only the steps that genuinely need the extra reasoning depth to Claude Opus 5.
Available on VM0 since July 25, 2026.