Verdict: Claude Opus 5 is the model that eliminates the impossible trade-off — you no longer have to choose between Fable 5's raw intelligence and Opus 4.8's speed and safety. At the same $5/M input and $25/M output price as Opus 4.8, Opus 5 more than doubles agentic coding performance on Frontier-Bench, hits new state-of-the-art on multiple benchmarks, and does it without the cybersecurity guardrails that make Fable 5 refuse half your security work. For most developers and small businesses already on Claude, this is a no-brainer upgrade. For teams evaluating Claude vs GPT-5, it shifts the calculus.
- Same price as Opus 4.8: $5/M input, $25/M output
- 2x+ improvement on agentic terminal coding (Frontier-Bench)
- New SOTA on Frontier-Bench, GDPval-AA, ARC-AGI 3, and OSWorld 2.0
- Cybersecurity: behind Mythos 5 but no refusals on defensive audits
- Two new beta API features: mid-conversation tool changes and automatic safety fallbacks
- Fast mode: ~2.5x default speed at 2x base price
- Model ID:
claude-opus-5
Pricing and benchmarks change often — last checked July 30, 2026.
What makes Claude Opus 5 different from Opus 4.8?
Claude Opus 5 is not a frontier Mythos-class model — it sits below Fable 5 and Mythos 5 in raw capability. But it does not need to be the strongest model to be the best model for daily work. Anthropic's positioning is deliberate: take near-frontier intelligence, keep it at Opus-tier pricing, remove the safety guardrails that make Fable 5 frustrating, and ship something developers can actually use every day.
The headline numbers from the official Anthropic announcement:
| Benchmark | Opus 5 vs Opus 4.8 | Notes |
|---|---|---|
| Frontier-Bench v0.1 | >2x performance at lower cost per task | New state-of-the-art |
| CursorBench 3.2 | Within 0.5% of Fable 5 peak at half the cost | Best performance-per-cost |
| ARC-AGI 3 | 3x higher score than next-best model | Reasoning and generalization |
| Zapier AutomationBench | ~1.5x next-best pass rate at same cost | Automation and tool use |
| OSWorld 2.0 (computer use) | Surpasses Fable 5 at ~1/3 the cost | Browser and desktop automation |
| GDPval-AA v2 | State-of-the-art | Knowledge work |
In practical terms, Opus 5 is the new default model on Claude Max and the strongest model on Claude Pro, replacing Opus 4.8 in the daily-driver slot.
How much does Claude Opus 5 cost?
Claude Opus 5 is priced identically to Opus 4.8: $5 per million input tokens and $25 per million output tokens, per the Anthropic announcement. For context, Fable 5 costs double that — $10/M input and $50/M output — making Opus 5 half the price of Anthropic's frontier model while approaching its performance on coding and knowledge work.
| Model | Input (per 1M tokens) | Output (per 1M tokens) | Guardrails? |
|---|---|---|---|
| Opus 5 | $5 | $25 | No |
| Opus 4.8 | $5 | $25 | No |
| Fable 5 | $10 | $50 | Yes (safety classifier) |
| Mythos 5 | $10 | $50 | Trusted access only |
Fast mode is available on Claude Platform and Claude Code, running at approximately 2.5x default speed for 2x the base price. This means a fast-mode request costs $10/M input and $50/M output — the same as standard Fable 5 — but with no guardrails and the option to drop back to default speed instantly.
Cache pricing follows Anthropic's standard pattern: cache reads at 10% of base input price, cache writes at 1.25x the base input price for 5-minute cache and 2x for 1-hour cache. The new mid-conversation tool changes beta (covered below) is designed specifically to preserve these cache savings.
Is Claude Opus 5 good at agentic coding?
On agentic terminal coding — the ability to run multi-step coding tasks autonomously through a terminal — Opus 5 represents the biggest single generational jump in the Opus line. According to Anthropic's Frontier-Bench results, Opus 5 more than doubles Opus 4.8's performance on agentic coding tasks while operating at lower cost per task.
Three capabilities from the announcement stand out for developers:
Computer vision improvisation: On a Frontier-Bench task where the model was given no way to view a drawing, Opus 5 wrote its own computer vision pipeline to extract geometry from raw pixels and reconstruct a 3D FreeCAD model. It succeeded repeatedly where no competing model could solve the task in five attempts.
Root-cause debugging: Given a real bug in a popular open-source package manager, Opus 5 found the root cause and fixed an edge case the community's existing patch had missed. A competing model only fixed the surface symptom.
Autonomous test harness creation: An engineer used Opus 5 to build a market data feed for a new exchange in a single session. Finding no live feed to validate against, Opus 5 built its own test harness to verify its code parsed exchange data correctly.
Lovable reported a 22% improvement over Opus 4.7 on hardest agentic coding tasks with significantly lower run-to-run variance — meaning you get reliable results build after build, not偶尔 brilliant output mixed with failures.
What about computer use and agent automation?
Computer use — letting AI agents control browsers and desktops — is the fastest-growing capability tier, and Opus 5 pushes it forward significantly. On OSWorld 2.0, the standard benchmark for computer-use tasks, Opus 5 surpasses Fable 5's best result at roughly one-third the cost, according to Anthropic's data.
Zapier CEO Wade Foster described the practical impact: "It took a raw account-health workbook and ran a full churn-prevention sequence end to end — flagging at-risk accounts, alerting the right owner, and summarizing for retention ops. Previous models didn't pass; Opus 5 hit 100%."
This matters for small businesses. If your workflow involves agents that need to navigate dashboards, fill forms, or interact with web interfaces that lack APIs, Opus 5 closes the gap between "works on simple pages" and "handles real enterprise SaaS."
For a deeper dive on the verification architecture that makes autonomous agent loops reliable, see our doer-judge loop engineering guide.
What are the new beta API features shipped with Opus 5?
Anthropic released two beta features alongside Opus 5 that matter for developers building on the Claude API.
1. Mid-conversation tool changes (beta)
Previously, the tools array in an API request sat at the very start of the hashed prompt cache prefix — earlier even than the system prompt. Editing it invalidated the cache for the entire conversation, forcing a full reprocessing of all prior messages. The mid-conversation tool changes beta fixes this.
Instead of editing the tools array, you append a tool_addition or tool_removal block inside the messages array at the point where the change becomes relevant. Everything before that point stays byte-identical, so the cache still hits and only the new message is processed as fresh input.
The beta header is mid-conversation-tool-changes-2026-07-01. It works on Claude Fable 5, Mythos 5, Opus 4.8, and Opus 5 across the Claude API, Amazon Bedrock, and Google Cloud.
Why this matters: In long agentic sessions — the kind that accumulate 50+ cached turns — swapping tools mid-session previously meant paying to reprocess the entire conversation. With this beta, you can add a debugging tool on turn 40 without paying cache rebuild costs on turns 1-39. For teams managing token budgets, this is a meaningful cost saver.
2. Automatic safety classifier fallbacks
This feature builds on the fallback system Anthropic introduced with Fable 5. When a request is flagged by Anthropic's safety classifier, it can now be automatically routed to another model — typically from Fable 5 to an Opus model — without returning a refusal to the user.
According to the Anthropic cookbook on classifier fallback billing, input tokens on a request the classifier blocks are not billed at all. When the request re-runs on the fallback model, those input tokens bill at the cache-read rate (10% of base input price) rather than the full cache-write rate.
For Opus 5 specifically, this matters in reverse: Opus 5 does not have the heavy safety classifier that Fable 5 does, so fewer requests get rerouted. But the fallback infrastructure is there if you build multi-model pipelines.
How does Opus 5 handle cybersecurity tasks?
This is where Opus 5's positioning becomes clearest. Anthropic is explicit about the trade-off: "Opus 5 does not advance the frontier in risky dual-use capabilities. In rigorous evaluations conducted alongside private sector and government partners, we found it remains behind Mythos 5 in both biology research and offensive cybersecurity."
However — and this is the key practical point — Opus 5 is still notably better at defensive cybersecurity than Opus 4.8. Heise reports that while Opus 5 was not specifically trained on cybersecurity tasks, its general capabilities are good enough to partially match Mythos 5 in finding vulnerabilities. It lags behind Mythos in exploiting those vulnerabilities, which is exactly the safety boundary Anthropic intends.
What this means for developers: Fable 5 and Mythos 5 refuse to work on security tasks — they hit guardrails and reroute. Opus 5 does not. You can run a security audit, ask it to analyze code for vulnerabilities, or use it for defensive cybersecurity work without triggering refusals. That makes it the practical choice for security-conscious teams who do not want their AI assistant to refuse half their work.
If you are working on AI security specifically — including the guardrail asymmetry problem where defender models refuse to help during incidents — our guide to AI guardrail asymmetry covers the framework for handling it.
Claude Opus 5 vs the competition: where does it fit?
| Capability | Opus 5 | Fable 5 | GPT-5.6 All | Sonnet 5 |
|---|---|---|---|---|
| Price (input/output per 1M) | $5 / $25 | $10 / $50 | Varies by provider | $3 / $15 |
| Agentic coding (Frontier-Bench) | SOTA | Within 0.5% | Competitive | Below Opus tier |
| Computer use (OSWorld 2.0) | SOTA at 1/3 Fable cost | Strong | Not benchmarked | Below Opus tier |
| Cybersecurity | Defensive only, no guardrails | Guardrails block offensive | Varies | Not the focus |
| Guardrails | None | Safety classifier reroutes | Standard safety | Standard safety |
| Best for | Daily driver coding, agents, security audits | Frontier intelligence when cost is not the constraint | Teams on OpenAI stack | Cost-sensitive production |
The honest comparison: Opus 5 is not the strongest model Anthropic makes. Fable 5 and Mythos 5 are smarter. But Fable 5 costs double and has guardrails that block entire categories of work. GPT-5.6 from OpenAI remains strong on certain benchmarks, but switching model families means switching tooling, prompt structures, and API patterns.
Opus 5 win is the model you reach for without thinking — fast enough, smart enough, cheap enough, and no guardrail friction. If you want a comparison of the broader model landscape including Chinese alternatives, see our Chinese AI models decision guide.
How to get started with Claude Opus 5
On Claude.ai: Opus 5 is the new default model on Claude Max and available on Claude Pro. If you have either plan, you are already on it.
On the API: Set
model: "claude-opus-5"in your requests. The model ID is live on the Claude API, Amazon Bedrock, and Google Cloud Vertex AI.In Claude Code: Update to the latest version and select Opus 5 from the model picker. Fast mode is available at 2x base cost if you need speed.
For budget-conscious setups: If you are running Claude Code but want to manage costs, our free Claude Code OpenRouter setup guide shows how to route to free models for testing and use paid Opus 5 only for production work.
For mid-conversation tool changes: Add the
anthropic-beta: mid-conversation-tool-changes-2026-07-01header to your API request. Declare all tools upfront withdefer_loading: trueon tools you want to hold back, then usetool_additionblocks to surface them mid-conversation without invalidating cache.
What this means for you
If you are a developer already on Claude: Switch to Opus 5 today. Same price, better code, no downside. The upgrade from Opus 4.8 to Opus 5 is free if you are on the API, and automatic if you are on Claude Max.
If you are a small business using AI for automation: Opus 5's OSWorld 2.0 and Zapier AutomationBench results mean your browser-based agent workflows will work more reliably. The computer-use improvement is the biggest practical win for non-coding automation.
If you are a security team: Opus 5 is the model to use for defensive security audits. No guardrails means no refusals, and its vulnerability-finding capability is good enough for code review and penetration testing prep. For offensive work, you still need Mythos 5 through trusted access — and our guide to self-improving AI agents vs static copilots covers which agent architecture handles that workflow best.
If you are evaluating Claude vs GPT-5: Opus 5's value proposition is clear — near-frontier intelligence at half the frontier price, with API features (mid-conversation tool changes, fallback routing) that OpenAI does not yet match. The ecosystem tooling (Claude Code, MCP) is also more mature for agentic work.
FAQ
Q: Is Claude Opus 5 free on Claude Pro? A: Claude Opus 5 is available on Claude Pro as the strongest model on that plan, and it is the new default model on Claude Max. On the API, it costs $5 per million input tokens and $25 per million output tokens — the same as Opus 4.8.
Q: Does Claude Opus 5 have safety guardrails like Fable 5? A: No. Opus 5 does not have the safety classifier that makes Fable 5 reroute cybersecurity and biology queries. This means you can use it for defensive security audits and code vulnerability analysis without triggering refusals. However, it also means Opus 5 lags behind Mythos 5 on offensive cybersecurity capabilities — by design.
Q: How fast is Claude Opus 5 compared to Opus 4.8? A: Opus 5 offers a fast mode that runs at approximately 2.5x the default speed for 2x the base price. A trading firm reported Opus 5 used roughly one-eighth the reasoning tokens and under half the latency of Opus 4.8 on their benchmark, suggesting it is more token-efficient even at default speed.
Q: Can I use Claude Opus 5 on Amazon Bedrock and Google Cloud?
A: Yes. Claude Opus 5 is available on the Claude API, Amazon Bedrock, and Google Cloud Vertex AI. The mid-conversation tool changes beta also works across all three platforms with the mid-conversation-tool-changes-2026-07-01 header.
Q: Should I upgrade from Opus 4.8 to Opus 5? A: Yes — the pricing is identical and the performance gains are across the board. There is no scenario where Opus 4.8 is the better choice after Opus 5's release, unless you have a specific compatibility constraint. Opus 5 outperforms 4.8 on coding, knowledge work, computer use, and defensive cybersecurity.
Q: What is the difference between Opus 5 and Fable 5? A: Fable 5 is a Mythos-class frontier model — stronger on raw intelligence but costs double ($10/M input, $50/M output) and has safety guardrails that reroute cybersecurity and biology queries. Opus 5 approaches Fable 5 on coding and knowledge work at half the cost, has no guardrails, and is the better daily-driver model for most developers.
Discussion
0 comments