Claude Opus 5 Is Here: Near-Fable Intelligence at Half the Price

The Release That Competes on Price, Not Just Power

Claude Opus 5 arrived on July 24, 2026, and Anthropic framed it around a claim about economics rather than raw capability: it comes close to the frontier intelligence of Fable 5 at half the price. Crucially, it costs exactly what Opus 4.8 cost - five dollars per million input tokens and twenty-five per million output - which means the upgrade is free in practical terms. It is available on every platform today as claude-opus-5, and it becomes the new default model on Claude Max as well as the strongest model accessible on Claude Pro. The timing explains the positioning. This is the fourth model Anthropic has shipped in under two months, following Mythos 5, Fable 5, and Sonnet 5, and Fable 5 drew real criticism from businesses over its token burn rate, with users blowing through budgets on tasks. Opus 5 is the direct answer to that: Anthropic describes it as designed to be used every day because it works more efficiently than other models. The competitive frame has shifted from who has the smartest model to who has the smartest model you can actually afford to run all day.

The Benchmarks: Doubling Its Predecessor, Tripling Its Rivals

The numbers are where the claim earns its weight. On Frontier-Bench v0.1, which evaluates valuable software engineering work, Opus 5 surpasses every other model and more than doubles Opus 4.8's performance while costing less per task. On CursorBench 3.2 at max effort, it comes within half a percentage point of Fable 5's peak score at half the cost per task, and it delivers more performance per dollar than any other model at high, xhigh, and max effort. On ARC-AGI 3, a test of solving genuinely novel problems the model has never seen, its score is three times the next-best model. On Zapier's AutomationBench, which measures completing real business tasks start to finish, its pass rate is about one and a half times the next-best model at the same cost per task - and even at its lowest effort setting it passes more tasks than any other model does at any setting. Zapier's CEO noted it took a raw account-health workbook and ran a full churn-prevention sequence end to end, hitting one hundred percent where previous models did not pass at all. On OSWorld 2.0 for computer use, it beats Fable 5's best result at just over a third of the cost. Anthropic states plainly that Opus 5 is the new state of the art on coding and knowledge work evaluations, while being honest that it remains behind Mythos 5 on cybersecurity tasks.

What Changed in Practice: Judgment, Self-Verification, and Fewer Turns

The most interesting reports about Opus 5 are not about scores; they describe a shift in behavior. Anthropic says it is much stronger at verifying its own work and iterating carefully until it succeeds, and three examples make that concrete. On one Frontier-Bench task it was given a drawing of a machine part and asked to rebuild it as a 3D model - but was deliberately given no way to view the drawing. It responded by writing its own computer vision pipeline to pull the geometry from raw pixels, then reconstructed the part, repeatedly, where no competing model succeeded in five attempts. Given a real bug in a popular open-source package manager, it found the root cause and fixed an edge case the community's own patch had missed, while a competing model fixed only the surface symptom and declared it resolved. An engineer at a trading firm had it build a market data feed for a new exchange in a single session; finding no live feed to validate against, it built its own test harness. Customers describe the same pattern from different angles: a frontend testing company reported it opened its pages in a browser at desktop and phone widths and caught a product hidden below the mobile fold before handing work back. A financial modeling team measured nine percentage points higher accuracy with a third fewer turns and tool calls and sixty percent less time. A legal team got comparable results while generating twenty-six percent fewer tokens than Opus 4.8 at max reasoning. And one engineer described it pushing back on a proposed design and not folding when pressed - instead narrowing its objection to a single question and proposing a compromise. Lovable, whose platform serves millions of builders, reported it ahead of every model in its family on internal evals, up twenty-two percent over Opus 4.7 on their hardest agentic coding tasks, with far less variance run to run.

Effort Levels, Science Gains, and the Safeguards You Will Notice

Three practical layers round out the release. First, effort control: Opus 5 exposes settings from low through medium, high, xhigh, and max, letting you optimize for intelligence or conserve tokens for faster and cheaper results. What matters is that it holds quality at lower effort better than its predecessor did - which is precisely why teams report comparable output at meaningfully lower token counts. The practical habit is to start at medium or high and raise the effort only when you actually hit a wall. Second, science: Opus 5 beats Opus 4.8 on every life sciences evaluation Anthropic ran, with the largest gains in organic chemistry, scoring ten point two percentage points higher on inferring molecular structures from spectroscopy data, and seven point seven points higher on predicting how protein sequence variations affect function. That makes it Anthropic's most capable generally available model for scientific research. Third, safeguards, which is where daily experience differs sharply from Fable 5. Opus 5's cyber classifiers are proportionally less restrictive and are expected to intervene around eighty-five percent less often than Fable 5's. It is allowed to find vulnerabilities in source code, while binary-based scanning, penetration testing, and exploit generation remain blocked; flagged requests fall back to Opus 4.8 by default across Claude.ai, Claude Code, and Cowork. Biology requests that get blocked on Fable 5 now route to Opus 5 rather than 4.8. On safety, Anthropic reports that Opus 5 does not advance the frontier of dual-use capability: it comes close to Mythos 5 at finding vulnerabilities but stays substantially behind at turning them into working exploits. And on alignment, its automated behavioral audit makes it Anthropic's most aligned model to date, scoring lowest on misaligned behavior with the lowest rates of deception and the least susceptibility to being tricked into misuse. Two beta updates shipped alongside it: mid-conversation tool changes that no longer invalidate the prompt cache, and automatic fallbacks on the API so flagged requests route to another model instead of failing outright. Unlike Fable 5, Opus 5 carries no data retention requirement for general access.

Prompt

# CLAUDE OPUS 5 - LAUNCHED JULY 24, 2026

# ─── THE ESSENTIALS ───
# API model string: claude-opus-5
# Price: $5 / M input · $25 / M output  (SAME as Opus 4.8)
# Positioning: comes close to Fable 5 intelligence at HALF the price
# New DEFAULT model on Claude Max · strongest model on Claude Pro
# Available today on all platforms
# Fast mode: ~2.5x default speed, at 2x base price (or usage credits in Claude Code)
# No data retention requirement for general access (unlike Fable 5)

# ─── HEADLINE BENCHMARKS ───
# Frontier-Bench v0.1 → beats ALL other models; MORE THAN DOUBLES Opus 4.8
#                        at a lower cost per task
# CursorBench 3.2     → at max effort, within 0.5% of Fable 5 peak, HALF the cost
# ARC-AGI 3           → 3x the score of the next-best model (novel problem solving)
# Zapier AutomationBench → ~1.5x next-best pass rate at same cost;
#                          even LOWEST effort beats every other model
# OSWorld 2.0 (computer use) → beats Fable 5 best result at ~1/3 the cost
# GDPval-AA v2, HLE, DeepSearchQA → best + most cost-efficient
# Still BEHIND Mythos 5 on cybersecurity

# ─── EFFORT LEVELS (the cost/intelligence dial) ───
# low · medium · high · xhigh · max
# Lower effort = fewer tokens, faster, cheaper
# Higher effort = deeper reasoning, better on hard tasks
# Key finding: Opus 5 holds quality at LOWER effort better than 4.8
#   (one legal team: similar performance with 26% fewer tokens vs 4.8 at max)

# ─── SCIENCE GAINS OVER OPUS 4.8 ───
# Better on EVERY life-sciences eval Anthropic ran
# Organic chemistry (molecular structure from spectroscopy): +10.2 points
# Protein sequence-variation effects: +7.7 points
# Now Anthropic's most capable GENERALLY AVAILABLE model for scientific research

# ─── SAFEGUARDS & FALLBACKS (what you will actually notice) ───
# Cyber classifiers are LESS restrictive than Fable 5 - expected to intervene
#   about 85% LESS often
# Allowed: finding vulnerabilities in source code
# Blocked: binary-based vuln scanning, penetration testing, exploit generation
# Flagged requests fall back to Opus 4.8 by default in Claude.ai / Code / Cowork
# Biology requests blocked on Fable 5 now route to Opus 5 (not 4.8)

# ─── TWO BETA UPDATES SHIPPED ALONGSIDE ───
# Mid-conversation tool changes → swap Claude's tools mid-conversation WITHOUT
#   invalidating the prompt cache (big token saver for agents)
# Automatic fallbacks on API → flagged requests auto-route to another model
#   instead of being blocked

# ─── WHICH MODEL TO USE NOW ───
# Everyday work, coding, agents, analysis → OPUS 5 (best value at the frontier)
# Multi-day autonomous projects / absolute hardest tasks → FABLE 5
# High-volume simple work → SONNET 5 / HAIKU
# Rule: start at Opus 5 medium/high effort; raise effort only when you hit a wall.