GuidesNews
Claude Opus 5: What Changed, Pricing, and Who It’s For
Anthropic’s Claude Opus 5 (24 July 2026): what’s new vs Opus 4.8, API pricing, Max/Pro availability, Fast mode, safeguards, and when to use it in Claude Code.
Brand marks are the property of their respective owners
Claude Opus 5 is Anthropic’s July 2026 upgrade to the Opus line: a model aimed at long-running agentic coding and professional knowledge work, priced at the same API rates as Opus 4.8, and positioned as the everyday high-capability option on Claude Pro / Max.
This is a launch analysis, not a press rewrite. Official claims come from Anthropic’s announcement and platform docs. Vendor benchmarks are their measurements — we translate them into decisions: when Opus 5 is worth the usage budget, when Sonnet 5 or Fable 5 still fit, and what to change in Claude Code workflows.
Verified against Anthropic’s public launch materials as of 2026-08-05. Model defaults and plan packaging can move; re-check the vendor page before you budget.
The short answer
- What launched: Claude Opus 5 (
claude-opus-5) on 24 July 2026, available on Claude apps and the Claude API (and cloud partners Anthropic lists). - What’s new vs Opus 4.8: Large claimed gains on agentic coding, computer use, verification/iteration, and several knowledge-work evals at the same $5 / $25 per 1M tokens API list rates.
- Where it sits in the family: Strongest model on Claude Pro; new default on Claude Max. Anthropic still routes some high-risk cyber/biology work through safeguards and fallbacks.
- PromptHive take: Treat it as the default heavy agent model if you already live in Claude’s stack — then measure your diffs and refusal rate, not only vendor charts.
Related evergreen pages: Claude Code pricing, how to use Claude Code, Claude Code review, Claude review, OpenAI Codex vs Claude Code, complete AI coding guide.
What was officially announced
From Anthropic’s product announcement (24 July 2026):
| Claim | Practical meaning |
|---|---|
| Available today | Consumer Claude surfaces + API + listed cloud hosts |
| Near Fable 5 intelligence at half the price (on some cost-per-task charts) | Opus tier becomes the “use daily” high model, not only a rare upgrade |
| New default on Claude Max; strongest on Claude Pro | Paid subscribers should expect Opus 5 in the picker / routing |
| Same API price as Opus 4.8: $5 in / $25 out per 1M tokens | No sticker shock on list rates; usage still scales with effort/tools |
| Fast mode ~2.5× speed at 2× base price | Optional latency trade for Platform / Code credit paths |
| Effort setting | Tune intelligence vs token burn |
| Mid-conversation tool changes (beta) | Swap tools without nuking prompt cache |
| Automatic API fallbacks (beta) | Safety-flagged requests can route to another model instead of hard-blocking |
Anthropic also published a system card and notes Opus 5 does not lead Mythos 5 on cybersecurity exploitation tasks, despite stronger general capability.
What’s new compared with Opus 4.8
| Dimension | Opus 4.8 era | Opus 5 (announced) |
|---|---|---|
| Role | Prior Opus-class workhorse | Step-change Opus for agents + coding |
| API list price | $5 / $25 per 1M tokens | Unchanged list rates |
| Agentic coding | Strong | Anthropic reports large gains (e.g. Frontier-Bench cost/performance charts vs 4.8) |
| Computer use | Competitive | Claims SOTA-style cost curves on OSWorld 2.0-class evals |
| Verification behaviour | Good | Emphasised: iterate until checks pass, not just first draft |
| Cyber posture | Classifiers + CVP | Still blocks many exploit/pentest paths; find-vuln easier than exploit; fallbacks to older Opus on flags |
| Daily product default | Varied by plan | Max default; Pro’s strongest model |
What did not change: you still need supervision for anything that edits files or runs tools. A better model writes more confident wrong code when unconstrained — see our agents guide.
Features explained (without the brochure)
Effort levels
Opus 5 exposes an effort control so you can trade depth for cost/latency. That matters more than marketing “IQ” numbers: a Max-effort coding agent run can burn a Pro day even when list rates look familiar.
Fast mode
Anthropic describes Fast mode as roughly 2.5× default speed at twice base price on Platform / Code usage-credit paths. Use it when wall-clock time matters more than token efficiency — not as a free lunch.
Tooling platform betas
- Mid-conversation tool changes keep prompt cache valid when you enable/disable tools mid-session — useful for multi-stage agents.
- Automatic fallbacks reduce hard refusals by routing flagged traffic to another model (often older Opus) when classifiers fire.
Safeguards you will feel
Cyber classifiers are less restrictive than Fable 5 for source-code vulnerability finding, but still block many exploit-generation and binary-scanning styles of work. On Claude.ai / Claude Code / Cowork, flagged requests can fall back to Opus 4.8. Enterprises in Anthropic’s Cyber Verification Program get a less restricted path.
Biology-related blocks that hit Fable can route to Opus 5 rather than 4.8 — important if you use Claude for scientific research tooling.
Pricing and availability
| Channel | What Anthropic states |
|---|---|
| API | $5 / $25 per 1M input/output tokens; ID claude-opus-5 |
| Claude Pro | Strongest model available on Pro |
| Claude Max | New default model |
| Team / Enterprise | Available on Claude for business users per product pages |
| Cloud | Offered via AWS, Google Cloud, Microsoft Foundry (per Anthropic materials) |
| Fast mode | ~2× price for ~2.5× speed (Platform / credits) |
Subscription reality: Claude Code and chat share plan usage. A “better” model that tools more aggressively can still exhaust Pro faster. Read Claude Code pricing for the seat/API shapes we already track.
PromptHive hands-on analysis
We did not re-run every public leaderboard. What matters for PromptHive readers:
- Same list price, higher capability is the rare upgrade that is easy to justify if you already buy Opus-class tokens.
- Default-on Max means Max users get Opus 5 without a picker hunt — budget for heavier sessions.
- Claude Code users should treat Opus 5 as the new baseline for multi-file agents, then re-baseline review time. Better root-cause stories from early-access quotes (Cursor, Devin, Lovable, JetBrains partners on Anthropic’s page) match our existing advice: read the diff.
- Not a free-tier story. Free Claude still is not a free Claude Code tier — see best free AI coding tools.
- Competition frame: OpenAI’s GPT-5.6 Sol/Terra/Luna family and Google’s Gemini Flash line compete on cost tiers; Opus 5’s pitch is Opus-class quality at prior Opus dollars, not “cheapest token.” For ChatGPT-bundled agents, compare Codex vs Claude Code.
Pros
- API list rates held flat vs Opus 4.8 while Anthropic claims large agentic gains
- Clear product placement: Pro strongest / Max default
- Effort + Fast mode give operational knobs
- Strong third-party partner quotes on coding agents (Cursor, Devin, Lovable) in the official post
- Alignment story: Anthropic reports best automated behavioural-audit score among recent Claudes
Cons
- Still behind Mythos on cyber exploitation and dual-use research paths
- Safeguard fallbacks can surprise automation (silent model swap)
- Shared subscription usage can make “same price” feel more expensive in practice
- Vendor evals are not your production suite — measure yourself
- Model landscape is crowded (Fable 5, Sonnet 5, GPT-5.6 family) — defaults matter more than names
Best use cases
| Use case | Fit |
|---|---|
| Multi-file refactors and hard debugging in Claude Code | Primary |
| Long agent loops with tests as the success signal | Primary |
| Computer-use / multi-step business automation evals | Strong on Anthropic’s AutomationBench narrative |
| Everyday chat on free tier | No — paid Claude |
| Offensive security / unrestricted pentest | No without CVP / special access |
| Cheapest high-volume classification | Prefer cheaper tiers (Sonnet / Haiku / other vendors) |
How to trial it in Claude Code (30 minutes)
- Confirm you are on Pro or Max (or API with
claude-opus-5). - Open a feature branch with a clean git status.
- Paste one real ticket into Claude Code with tests as the success signal.
- Score: correctness, review minutes, remaining plan usage, how often you intervened.
- Re-run the same ticket on a cheaper tier (or Opus 4.8 if still available) if you need a control.
If Opus 5 finishes the ticket with less babysitting, keep it as default. If it thrashing-retries and empties the pool, lower effort or shrink the task — a stronger model is not a substitute for a clearer ticket. Workflow detail: how to use Claude Code.
When to stay on Sonnet 5 or another vendor
| Situation | Prefer |
|---|---|
| High-volume, low-stakes rewrites | Sonnet / Haiku class or non-Anthropic cheap tiers |
| ChatGPT already paid and coding is secondary | Codex inside that ecosystem |
| Editor-first flat bill | Cursor |
| Unlimited autocomplete only | GitHub Copilot Pro |
| Free evaluation of coding AI | Not Opus 5 — see best free AI coding tools |
What this means for PromptHive clusters
- Coding: Opus 5 strengthens the case for Claude Code as the paid agent default we already review — update mental models, not every historical comparison table overnight.
- Agents: Better verification behaviour is only valuable if you keep human approval gates (agents guide, MCP advanced).
- App builders: Partner quotes on Anthropic’s post (Lovable, Cursor, Devin) are marketing-adjacent; still pick tools by job (Lovable, Replit Agent).
Final PromptHive verdict
Claude Opus 5 is a real generational Opus upgrade on paper, without a list-price hike. If you already pay for Claude Pro, Max, or Opus API traffic, make it the default for serious coding agents, then spend one afternoon re-running your own harness. If you only need light chat or autocomplete, this launch changes nothing — stay on free tiers and assistants, not Max.
CTA: Builders on Anthropic’s stack — update your Claude Code model setting, pin a CLAUDE.md, and compare one real ticket against Opus 4.8 before changing team defaults. Start from how to use Claude Code and the coding hub.