Anthropic released Claude Opus 5 on Friday, presenting it as a lower‑cost model that retains much of the capability of its top‑tier Fable 5 while targeting routine enterprise workloads. The company frames the launch as evidence that the AI competition is shifting from raw peak capability to the economics of day‑to‑day usage.
Pricing and availability
Opus 5 is available immediately across Anthropic's platforms. Pricing matches that of Opus 4.8: $5 per million input tokens and $25 per million output tokens. The model becomes the default on Claude Max (Anthropic's premium consumer tier) and is the strongest model available on Claude Pro.
Positioning: efficiency for the middle band of work
Anthropic is not claiming Opus 5 is its smartest model — Fable 5 remains the company’s most capable system. Instead, Anthropic argues that many commercially important tasks sit in a middle difficulty band where near‑frontier performance delivered cheaply and efficiently is more valuable than the absolute frontier. An Anthropic spokesperson described the lineup as: Opus 5 for daily use and complex tasks you hand off and review; Fable 5 for the longest, most autonomous projects; Sonnet 5 for large‑scale work where speed per call matters; and Haiku 4.5 for subagents and instant answers.
Benchmarks and limits
Anthropic reports Opus 5 sets new state‑of‑the‑art results on several coding and knowledge‑work evaluations, including Frontier‑Bench and GDPval‑AA. On Frontier‑Bench v0.1 (an agentic terminal coding benchmark), Opus 5 scored 43.3% — more than double Opus 4.8's 18.7% and ahead of Fable 5's 33.7% — at a lower cost per task, according to the company. On ARC‑AGI 3 (novel problem‑solving), Anthropic says Opus 5 scored three times higher than the next best model. On OSWorld 2.0 (computer‑use benchmark) the company reports outperforming Fable 5 at roughly one‑third the cost.
Anthropic also candidly notes shortcomings: Opus 5 trails Mythos 5 on cybersecurity and biology research, and an OpenAI‑family model leads on one agentic coding benchmark. The spokesperson emphasized that Opus 5 wins on bounded evals with specific outcomes, while Fable 5 is meant for long‑horizon, multi‑step coherence over hours or days. They recommend running both models on representative workloads: one bounded task and one long‑horizon job.
Token efficiency and customer feedback
A central theme of the launch is token efficiency: Opus 5 is designed to perform well on a per‑dollar basis. The model includes an adjustable “effort” setting that lets customers trade intelligence for speed and token savings, and Anthropic highlights performance at a given cost, not only peak scores. Early customers gave concrete examples: Harvey, a legal AI firm, said Opus 5 matched Opus 4.8's max‑reasoning mode while generating about 26% fewer tokens on average, according to Niko Grupen, head of applied research. Richard Pham at Fundamental Research Lab reported that on hard financial‑modeling tasks accuracy rose by nine percentage points while using roughly one‑third fewer turns and tool calls and 60% less time. Wade Foster, CEO of Zapier, said Opus 5 topped his company's AutomationBench leaderboard by running a full churn‑prevention workflow end‑to‑end without spending more tokens than prior Claude models; previous models did not pass, Opus 5 hit 100%. Scott Wu, CEO of Cognition (developer of the Devin coding agent), reported that on FrontierCode 1.1 "Claude Opus 5 approaches Fable‑level performance at half the cost," with particular strength in debugging and root‑cause analysis.
Self‑verification and hidden automation costs
Anthropic sells a behavioral story: Opus 5 verifies and iterates on its work until it succeeds. In tests the company described, the model built its own computer vision pipeline to extract geometry from pixels when it was intentionally given no viewing method, repeatedly solving a Frontier‑Bench task no competitor solved in five attempts. In another case it fixed a real bug in an open‑source package manager, addressing an edge case the community patch missed; a competing model patched only the symptom. An engineer at a trading firm used Opus 5 to build a market data feed in a single session; lacking a live feed for validation, the model constructed its own test harness. Customer reports reflected similar behavior in production. Cristian Rivera, staff software engineer at Stripe, said he assigned the model a "chief‑of‑staff" role over his dev environments for a weekend: it built a monitor, managed boxes, and only involved him for judgment calls. Anthropic argues this self‑verification capability narrows the gap between demo and deployable systems: reducing human review — the largest hidden cost for enterprise AI today — is why customers emphasize fewer turns, fewer passes, and less time rather than just higher raw scores.
Safety approach: capability constraints, classifiers, and fallbacks
The launch highlights Anthropic's multi‑layered safety approach, which includes intentionally avoiding teaching certain skills. An automated behavioral audit found Opus 5 to be the company's most aligned generally available model to date, scoring 2.3 on overall misaligned behavior — lower than Opus 4.8, Sonnet 5, or Fable 5 — with the lowest deceptive behavior rates and least susceptibility to misuse. Anthropic says it deliberately avoided training Opus 5 on cyber tasks, as with Opus 4.8, yet general capability gains improved its discovery abilities. On Anthropic's OSS‑Fuzz evaluation Opus 5 identified vulnerabilities at a 79.4% rate (close to Mythos 5's ~80%) but succeeded at developing exploits in only four challenges versus Mythos 5's 13. That asymmetry — strong at discovery, weak at exploitation — appears intended, and the product's safeguards reflect that logic. Anthropic expects Opus 5's cyber classifiers to intervene about 85% less often than Fable 5's. When a classifier triggers in Claude.ai, Claude Code, or Claude Cowork, requests fall back to Opus 4.8 by default. Anthropic's explanation is that the fallback model's lower capability reduces the risk of harmful use; users are shown a message when such a fallback occurs. The company acknowledges the tradeoffs implicit in relying on classifiers to decide which model answers a request. On biology, Opus 5 is Anthropic's most capable generally available model for scientific research, scoring 10.2 percentage points higher than Opus 4.8 on the company's internal chemistry benchmark, though Mythos 5 remains stronger for long‑horizon, open‑ended tasks like autonomous drug design campaigns.
Commercial stakes and regulatory backdrop
The release comes amid strong commercial momentum and heavy infrastructure commitments. Reuters reported in February that Anthropic's valuation was roughly $380 billion in its latest funding round. Contrary Research estimated Claude held about 40% of enterprise large language model usage by late 2025, and Claude Code had reached roughly $1 billion in annualized revenue. Reported internal targets for 2026 ranged from $20 to $26 billion in revenue, supported by large compute deals including a reported $30 billion Azure compute agreement alongside partnerships with Google Cloud and Nvidia. These commitments make token efficiency a core product imperative: offering significantly more capability at the same listed price effectively reduces cost per unit of capability and expands the set of workloads that are economical to automate, creating recurring token revenue.
Additional features and legal notes
Also released on Friday were a Fast mode (about 2.5× default speed at twice the base price), automatic API fallback routing, and mid‑conversation tool changes that no longer invalidate the prompt cache — a small but practical feature for agent developers. Consistent with prior Opus models, Opus 5 carries no general data retention requirement for access, which Anthropic highlighted for customers with strict zero‑data‑retention needs. Developers can access the model as claude‑opus‑5 on the Claude API starting the day of the announcement. A U.S. judge this week gave final approval to Anthropic's $1.5 billion copyright settlement with book authors, resolving litigation over early training data. Reports in June also indicated U.S. government moves to restrict foreign access to Anthropic's most advanced models, underscoring how export policy now affects which customers can buy which systems.
What will decide Opus 5's success?
Two questions will determine the launch's payoff: whether Opus 5's efficiency claims hold up under real, large‑scale production workloads, and whether enterprises accept a reality where safety classifiers sometimes determine which model responds. The broader implication of the release is clear: the industry's center of gravity has shifted from competing on what the best model can do on its best day to what a very good model can do every day at much lower cost. Anthropic is betting that the greater commercial opportunity lies just behind the frontier — and Opus 5 is their wager on that position.
