Overview
Claude Opus 5 is Anthropic's July 24, 2026 Opus-tier release — an upgrade over Opus 4.8 that Anthropic positions as "a thoughtful and proactive model that comes close to the frontier intelligence of Claude Fable 5 at half the price." Rather than chasing a single benchmark ceiling, Opus 5 is built to be the default option for everyday coding, research, and business work — the model Anthropic expects most teams to reach for day to day, with Fable 5 reserved for the hardest long-horizon jobs. Claude Sonnet 5 launched on June 30, 2026; Opus 5 launched on July 24, 2026 and became the new default on Claude Max and the strongest model available on Claude Pro.
The model is available today on claude.ai, the Claude API as claude-opus-5, Claude Code, Claude Cowork, Amazon Bedrock, Claude Platform on AWS, Google Cloud Vertex AI, and Microsoft Foundry. It carries a 1M-token context window and up to 128k output tokens (300k on the Message Batches API with the output-300k-2026-03-24 beta header) — unchanged from Opus 4.8 — but its reliable knowledge cutoff moves forward to May 2026, five months later than Opus 4.8's January 2026 cutoff.
What's New
Near-Fable 5 Intelligence at Half the Price
The headline result is cost efficiency, not a new capability ceiling. On CursorBench 3.2 at max effort, Opus 5 lands within 0.5% of Fable 5's score while costing half as much per token. On OSWorld 2.0, Opus 5 actually surpasses Fable 5's best result, at just over a third of the cost. Anthropic frames Opus 5 as approaching — not replacing — Fable 5: for the very hardest long-horizon agentic work, Fable 5 remains the recommended model.
Stronger Verification and Iteration
Opus 5 is "much stronger at verifying its work and iterating carefully until it succeeds," per Anthropic's release benchmarks. On Frontier-Bench v0.1 it surpasses every model Anthropic tested, roughly doubling Opus 4.8's performance, and it scores about 3× higher than the next-best model on ARC-AGI 3 — a benchmark built around novel reasoning rather than memorized patterns.
Life Sciences and Scientific Research Gains
Opus 5 outperforms Opus 4.8 on every life-sciences evaluation Anthropic reported, including a 10.2-percentage-point jump over Opus 4.8 in organic chemistry and a 7.7-point gain on predicting how protein-sequence variations affect function, according to Anthropic's internal benchmark.
Developer-Facing API Changes
Two beta API features ship alongside the model: mid-conversation tool changes, enabled with the mid-conversation-tool-changes-2026-07-01 beta header, which let developers modify the set of available tools mid-session without invalidating the prompt cache, and automatic fallbacks (beta), enabled with the server-side-fallback-2026-07-01 header, which reroute classifier refusals to Anthropic's recommended fallback model rather than treating them as request failures.
Anthropic's Most Aligned Model to Date
Anthropic's automated behavioral audit scores Opus 5 at 2.3, and the company states it adheres to Claude's Constitution more closely than Opus 4.8, Sonnet 5, or Fable 5, with the lowest observed rate of deceptive behavior across the current model lineup. Opus 5 remains behind Claude Mythos 5 in both biology research and offensive cybersecurity; it is close on vulnerability identification but substantially behind on exploit development.
Performance Benchmarks
Headline numbers from Anthropic's release:
| Benchmark / Task | Claude Opus 5 | Comparison |
|---|---|---|
| Frontier-Bench v0.1 | Surpasses all tested models | ~2× Opus 4.8's performance |
| CursorBench 3.2 (max effort) | Within 0.5% of Fable 5 | At half Fable 5's per-token cost |
| OSWorld 2.0 | Surpasses Fable 5's best result | At ~1/3 of Fable 5's cost |
| ARC-AGI 3 | ~3× the next-best model | — |
| Zapier AutomationBench | ~1.5× the next-best model | At the same cost |
| Organic chemistry (life sciences) | +10.2 points | vs. Opus 4.8 |
| Protein sequence-variation effect prediction | +7.7 points | vs. Opus 4.8 on Anthropic's internal benchmark |
| Automated behavioral audit | 2.3 overall misaligned-behavior score | Lowest among Anthropic's recent models; separately, Opus 5 showed the lowest deceptive-behavior rate |
The CursorBench and OSWorld results matter most for teams running AI agent and AI code checker pipelines that were previously choosing between Opus 4.8's price and Fable 5's ceiling — Opus 5 narrows that gap from the price side rather than the capability side.
Migration Guide
Opus 5 is a same-price upgrade from Opus 4.8 with two API breaking changes:
- Thinking is now on by default. Requests without a
thinkingfield now use adaptive thinking, so revisitmax_tokens, cost, and latency. - Disabling thinking requires effort
highor below. Usingthinking: {"type": "disabled"}withxhighormaxreturns a 400 error. - Effort no longer defaults to
highon claude.ai. On Opus 4.8, theeffortparameter defaulted tohighon every surface, including claude.ai. On Opus 5, it defaults tohighonly on the Claude API and Claude Code — set it explicitly if your claude.ai workflow assumed the old default. - Knowledge cutoff moved to May 2026. Five months newer than Opus 4.8's January 2026 cutoff — re-verify any prompts that relied on Opus 4.8's specific training-data boundary.
- Mid-conversation tool changes preserve prompt cache. Because this feature is beta, include the
mid-conversation-tool-changes-2026-07-01header before swapping tools mid-session without invalidating the prompt cache. - Same tokenizer and context ceiling as Opus 4.8. Token counts, the 1M-token context window, and the 128k standard output limit carry over, but workloads that previously ran without thinking should revisit
max_tokens, cost, and latency. Web fetch and Priority Tier are also unavailable on Opus 5.
Pricing & Plans
| Plan | Price |
|---|---|
| API – Input (regular) | $5 per million tokens |
| API – Output (regular) | $25 per million tokens |
| API – Fast mode research preview (~2.5× speed; Claude API only) | $10 input / $50 output per million tokens |
| Claude.ai Pro / Max | Included within plan limits — Opus 5 is the new Claude Max default |
| Team | Standard seats: $25 per member/month on monthly billing or $20 per member/month on annual billing; Premium seats: $125 or $100 respectively |
| Enterprise | $20 per seat/month, billed annually, plus usage at standard API rates; tailored quotes are available |
| Cloud availability | Amazon Bedrock, Claude Platform on AWS, Google Cloud Vertex AI, Microsoft Foundry |
Fast mode note: Opus 5's fast mode is a research preview on the first-party Claude API and through usage credits in Claude Code, running roughly 2.5× faster at $10/$50 per million tokens on the API. Standard requests stay at $5/$25 — identical to Opus 4.8.
Best For
- Teams choosing between Opus 4.8's price and Fable 5's ceiling who want most of the capability gap closed without the $10/$50 price tag
- Claude Code users and Claude Max subscribers, where Opus 5 is now the default model
- Life-sciences and scientific research teams working on organic chemistry or protein-sequence tasks
- Developers building tool-calling agents that can adopt the beta headers for mid-conversation tool changes and server-side fallback
- Anyone currently on Opus 4.8 who wants a same-price upgrade and can review thinking-on-by-default behavior, the
xhigh/maxrestriction, and feature availability - Teams that need AI agent reliability but don't require Fable 5's absolute reasoning ceiling
FAQ
How is Claude Opus 5 different from Opus 4.8?
Opus 5 holds the same $5/$25 pricing and 1M-token context window as Opus 4.8 but more than doubles Opus 4.8 on Frontier-Bench, scores about 3× the next-best model on ARC-AGI 3, and improves every reported life-sciences evaluation (+10.2 points in organic chemistry). Its reliable knowledge cutoff also moves forward to May 2026. The two API breaking changes are thinking on by default and a 400 error when thinking is disabled at xhigh or max; the claude.ai effort default also changed.
Is Claude Opus 5 as capable as Fable 5?
Close, but not equal. On CursorBench 3.2 at max effort, Opus 5 lands within 0.5% of Fable 5's score at half the price, and it even surpasses Fable 5 on OSWorld 2.0. For the hardest long-horizon reasoning and agentic tasks, Anthropic still recommends Fable 5.
How much does Claude Opus 5 cost?
Standard pricing is $5 per million input tokens and $25 per million output tokens — unchanged from Opus 4.8. An optional fast mode runs about 2.5× faster at double the price ($10/$50 per million tokens). On claude.ai, Opus 5 is included within Pro, Max, Team, and Enterprise plan limits, and is now the default model on Claude Max.
Where can I use Claude Opus 5?
Opus 5 is available on claude.ai, the Claude API as claude-opus-5, Claude Code, Claude Cowork, Amazon Bedrock, Claude Platform on AWS, Google Cloud Vertex AI, and Microsoft Foundry.
Should I upgrade from Opus 4.8 to Opus 5?
For most teams, yes — it's the same price with measurable gains across reasoning, coding, and life-sciences benchmarks, plus a newer knowledge cutoff. Before migrating, check thinking-on-by-default behavior, the xhigh/max restriction when thinking is disabled, the claude.ai effort default, and unsupported web fetch and Priority Tier.
