Anthropic dropped Opus 5 on Friday, its newest heavyweight AI model, and the numbers tell a surprising story: this smaller, cheaper sibling of Fable 5 actually beats the flagship on several key benchmarks. At half the price.
The model, available now on all Anthropic platforms, marks the latest step in a breakneck release cycle. Opus 4.8 only came out two months ago, on May 28. Since then, Anthropic has rolled out Mythos 5, Fable 5, Sonnet 5, and now Opus 5 — leaving only the lightweight Haiku model waiting for a 5-series upgrade. That pace highlights just how fast the AI market is moving. Anthropic, often seen as the safety-conscious player in the space, is now shipping major models at a clip that rivals any competitor.
What Opus 5 brings to the table
Opus 5 outperforms Fable 5 on Frontier-Bench v0.1, a software engineering evaluation, and on GDPval-AA for knowledge work tasks. On CursorBench 3.2 at max effort, Opus 5 lands within half a percent of Fable 5's peak score but costs half as much per task. That combination of near-frontier intelligence at a mid-range price point is what makes this release noteworthy.
Anthropic describes the model as "thoughtful and proactive," and early testers back that up. "I gave it a chief-of-staff role over my dev environments over one weekend," said Cristian Rivera, a Staff Software Engineer at an AI startup. "It built its own monitor, drove each box, and pulled me in only for the judgment calls." Another tester described Opus 5's approach to code review: "It verified the branches, checked the template, and thought through test implications so the handoff was clean. The older models tended to jump ahead and get caught on our checks."
The model is priced at $5 per million input tokens and $25 per million output tokens — identical to Opus 4.8. There's also a Fast mode that runs about 2.5 times the default speed, available at double the base price on the Claude Platform and through Claude Code usage credits. For developers who need quick iterations, Fast mode could be the difference between a usable assistant and a bottleneck.
On knowledge work tasks, Opus 5 shows a clear generational step. It improved by 8 percent on the Box enterprise content benchmark and posted an 11 percent gain on data analysis tasks. Due diligence workflows saw a 17 percent improvement over Opus 4.8, according to Box CTO Ben Kus.
Fewer restrictions, smarter safeguards
One of Opus 5's biggest selling points is freedom from the tight guardrails that have frustrated Fable 5 users. Opus 5 does not carry the 30-day data retention policy that covers Fable and Mythos, a change that privacy-conscious users had pushed for. That means conversations with Opus 5 are not retained on Anthropic's servers for the same window, giving enterprises more confidence when handling sensitive data.
The safety classifiers on Opus 5 are proportionally less restrictive than Fable 5's. Anthropic expects them to engage about 85 percent less often. The model can find vulnerabilities in source code — useful for defensive security work — but still blocks binary-based vulnerability scanning, penetration testing, and exploit generation. This is a deliberate design choice: Opus 5's cyber classifiers allow white-hat security research while preventing the most dangerous misuse cases.
There's a new beta feature called Automatic Fallbacks. When a prompt triggers the safety classifier, the system routes the request to a less powerful model instead of returning an error. API users get a functional response rather than a dead end. That alone could save developers real frustration — nothing stalls a workflow like a sudden refusal from the model you are counting on.
The performance story
On Frontier-Bench v0.1, Opus 5 more than doubles Opus 4.8's performance at a lower cost per task. The improvement is even steeper on scientific research. On organic chemistry tasks — inferring molecular structures from spectroscopy data — Opus 5 scores 10.2 percentage points higher than Opus 4.8. On protein-related benchmarks where the model predicts how variations in a protein sequence affect its function, it is up 7.7 points.
"The model behaves more like a careful scientist than any model we've run," said Alfredo Andere, CEO of a genomics analysis firm that tested Opus 5 early. "It reaches for the right statistical tests to rule out confounders, cross-checks its own results by independent methods, and stays on track through long multi-step analyses."
Opus 5 also shows strong gains on financial modeling. One early adopter reported 9 percentage points higher accuracy with a third fewer turns and tool calls, at 60 percent less time. Another tester at a trading firm said Opus 5 achieved the best scores on their internal benchmark using roughly a seventh of the reasoning tokens and under half the latency of Opus 4.8.
For legal document work, Opus 5 scored the highest of any model tested on first-turn redlines — nearly double Opus 4.8's score. "Commenting is better too," said Ryan Tanenholz, a member of technical staff at a legal AI firm. "On NDAs it gets to the redline in less time and with fewer passes, with accuracy maintained or better."
On Zapier's AutomationBench, Opus 5 topped the leaderboard. "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," the Zapier team reported. "Previous models didn't pass; Opus 5 hit 100 percent."
How Opus 5 fits in Anthropic's lineup
Anthropic now has five models in its 5-series: Mythos 5 (the most capable), Fable 5 (the flagship with the strongest safeguards), Sonnet 5 (balanced), Opus 5 (the efficient heavyweight), and the yet-to-launch Haiku 5 (lightweight). Opus 5 sits in an interesting middle ground — close enough to Fable 5 on benchmarks to serve most enterprise use cases, but without the heavy restrictions that make Fable hard to deploy at scale.
It is the default model on Claude Max and the strongest model available on Claude Pro. For API users, Opus 5 offers a clear value proposition: near-frontier intelligence at roughly half the operational cost of Fable 5. For biology research, Opus 5 is now Anthropic's most capable generally available model, since Fable 5 blocks many biology-related requests and Mythos 5 is more restricted.
Safety by design
Anthropic says Opus 5 does not advance the frontier in risky dual-use capabilities. In evaluations with private-sector and government partners, it remains behind Mythos 5 in both biology research and offensive cybersecurity. The model was intentionally not trained on cyber tasks, though it has improved on them anyway as a side effect of becoming generally more capable.
The company's automated behavioral audit found Opus 5 to be its most aligned model to date. It adheres to Claude's Constitution better than Opus 4.8, Sonnet 5, or Fable 5, exhibits the lowest rates of deceptive behavior, and is the least susceptible to being tricked into misuse. "It thinks harder before it writes a single line, catches its own logical faults during planning rather than after the fact, and reasons about why an answer is right, not just whether it works," said Denis Shiryaev, Head of AI at JetBrains.
On OSS-Fuzz, an evaluation that tests how well models can find and then exploit vulnerabilities, Opus 5 identifies vulnerabilities at a similar rate to Mythos 5 but falls far behind when it comes to actually exploiting them — exactly the profile Anthropic wants for a safe, deployable model.
The bigger AI picture
The launch comes during a busy week for AI news that underscores just how competitive this market has become. Google announced that Gemini has crossed 950 million monthly users, up from 750 million in February, and is closing in on ChatGPT's 1 billion user milestone. Sensor Tower data shows Gemini's market share among AI assistants rose to 27.7 percent in the first half of 2026, while ChatGPT's share dipped below 50 percent for the first time.
A new AI lab called Prentis, co-founded by Reid Hoffman and Mark Pincus, is separately in talks to raise $100 million at a $1 billion valuation. Prentis is betting that automating office workflows — handling insurance claims, customs duty refunds, and other document-heavy tasks — will outpace coding as AI's biggest commercial use case. Its Hive-32B model claims to outperform GPT-5.4 and Opus 4.6 on computer-use benchmarks at roughly a tenth of the cost.
On the hardware side, AMD launched its Helios rack-scale AI system at the AAI 2026 conference, taking direct aim at Nvidia's data center dominance. And AI chip startup Etched hit a $10.3 billion valuation from big-name investors, signaling that the appetite for AI compute investment shows no signs of cooling.
Opus 5 is available today on Claude.ai, the Claude API, and through Claude Code. For enterprise users in Anthropic's Cyber Verification Program, a version with fewer security restrictions is available immediately.