Anthropic has released Claude Opus 5, a new flagship model priced at $5/$25 per million tokens and positioned as a cost-efficient daily driver for bounded enterprise tasks rather than a peak-capability competitor. The model distinguishes itself through rapid benchmark gains and an agency feature that lets it write its own validation tools to verify outputs before handing them to human reviewers.
Anthropic Unveils Claude Opus 5 as the Daily Driver for Enterprise AI
The new flagship lands at unchanged pricing, carving out a role as a cost-efficient workhorse rather than a peak-capability competitor.
Anthropic released Claude Opus 5 this week, and it's not trying to be the smartest model in the stack. The $5 input, $25 output model is positioned squarely as a daily driver for bounded, production workloads. You hand it complex tasks. You come back to review when it's done. The really ambitious, multi-day projects still go to Fable 5.
There are four Opus variants available through the API, Claude.ai, Claude Code, and Claude Cowork. Pricing stays the same as Opus 4.8. No changes there.
"Opus 5 as your daily driver, the model you hand complex work to and review when it's done." — Anthropic spokesperson
The framing makes sense. Bounded tasks versus long-horizon autonomy is becoming the real axis of model differentiation this year. Anthropic isn't pretending otherwise.
Performance and the Self-Verification Angle
Opus 5 scores 43.3% on Frontier-Bench v0.1. That's more than double Opus 4.8's 18.7% and sits ahead of Fable 5 at 33.7%. It also hits within 0.5% of Fable 5 on CursorBench 3.2 at roughly half the cost per task. If you're running agentic coding pipelines, the numbers are hard to ignore.
The actual differentiator Anthropic is leaning into is self-verification. The model doesn't just produce an answer. It writes its own tooling to validate the output.
In one demo, Opus 5 was given a pixel drawing of a machine part with no way to view it directly. It built a computer vision pipeline from scratch, extracted the geometry, and reconstructed the CAD model. No competing model solved it in five attempts. Another run found and fixed an edge case that a community patch missed. A third handled a trading firm's market data feed, then built its own test harness since no live feed existed.
The implications for enterprise deployment are obvious. Human review is expensive. A model that can validate its own work at scale chips away at that hidden cost.
Safety and Fallback Behavior
Misalignment scores sit at 2.3 on Anthropic's behavioral audit, the lowest of any recent model. Deceptive behavior is down. Susceptibility to trickery is down. The model still improved on cybersecurity tasks through general capability gains, but Anthropic intentionally avoided training it on cyber-specific datasets.
Vulnerability identification sits around 79.4% on OSS-Fuzz. Exploit development is limited to four challenges solved, compared to Mythos 5's thirteen. By design. Cyber classifiers intervene 85% less often than they do on Fable 5.
When requests get flagged, they fall back to Opus 4.8. Users see a message explaining the downgrade. Lower capability, lower risk.
Biological research requests that were previously blocked from Fable 5 now route to Opus 5 instead of Opus 4.8. Still limited on long-running autonomous bio tasks, but a meaningful upgrade for scientific research.
New Capabilities and What Comes Next
Fast mode delivers roughly 2.5× speed at 2× the base price. Automatic fallback routing kicks in when classifiers flag problematic prompts. Tool changes mid-conversation no longer invalidate the prompt cache, which should matter for agent developers running long sessions.
The launch drew 1,461 points and 805 comments on Hacker News within hours of posting. It also hit number one on the Artificial Analysis Intelligence Leaderboard.
Opus 5 isn't chasing peak capability. It's chasing economic efficiency for bounded work. The pricing stays flat. The token counts drop. You can run it harder without breaking the bank.
Head here to explore the Claude API. Keep in mind that Fast Mode and fallback routing are available immediately across all supported platforms. The model ID is claude-opus-5.
