Added to our archive on 1 October 2026. This report is dated to the day of the event.

Less than seven weeks after Anthropic put a price of USD 10 per million input tokens on its most capable public model, it released one that it says does most of the same work for half as much. Claude Opus 5 went on sale on 24 July 2026 at USD 5 per million input tokens and USD 25 per million output tokens, unchanged from Claude Opus 4.8, which had arrived on 28 May.

“It's a thoughtful and proactive model that comes close to the frontier intelligence of Claude Fable 5 at half the price,” Anthropic wrote in its announcement. For businesses the bigger change may be what Opus 5 leaves out: Fable 5's strict cyber filters, and the 30 day data retention that Anthropic requires on its most capable models. Opus 5 is available on every Anthropic platform, became the default model on Claude Max and is the strongest model offered on Claude Pro.

  • What: Claude Opus 5, successor to Opus 4.8, on the Claude apps, the Claude API, Amazon Bedrock, Google Cloud and Microsoft Foundry.
  • Price: USD 5 per million input tokens and USD 25 per million output tokens; Claude Fable 5 costs USD 10 and USD 50.
  • Why it matters: close to top tier results without Fable's heavy filters or its data retention rule.

What changed against Opus 4.8

All performance figures here come from Anthropic and its early access customers. On Frontier-Bench v0.1, a coding test, Anthropic says Opus 5 more than doubled Opus 4.8's score at a lower cost per task. On CursorBench 3.2 at maximum effort it landed within 0.5 percent of Fable 5's best score at half the cost per task. On ARC-AGI 3, which tests solving unfamiliar problems, its score was three times that of the next best model, and on the OSWorld 2.0 computer use test it beat Fable 5's best result at just over a third of the cost. In life sciences it scored 10.2 percentage points higher than Opus 4.8 on working out molecular structures from spectroscopy data.

The company's main claim is about behaviour: Opus 5 is “much stronger at verifying its work and iterating carefully until it succeeds.” In one test it had to rebuild a machine part in the FreeCAD design program from a drawing it had no way to view directly, so it wrote its own computer vision pipeline to read the geometry from the raw pixels. Sualeh Asif, a co-founder of the coding tool Cursor, summed up the early access verdict: “Claude Opus 5 delivers near Fable 5 intelligence at Opus speed and cost.”

The technical envelope is a 1 million token context window, up to 128,000 output tokens per request and reliable knowledge up to May 2026. An effort setting lets users trade intelligence against tokens and speed, a feature Fortune highlighted in its launch report. A fast mode runs about 2.5 times quicker at twice the base price. Two beta features shipped alongside: changing which tools a conversation can use without invalidating the prompt cache, and automatic fallbacks that send requests flagged by safety classifiers to another model instead of blocking them.

Lighter filters, no retention rule

Anthropic's announcement describes Opus 5's safeguards as similar to those of Opus 4.8, with stronger guardrails on a narrow range of cyber tasks; the system card instead says they match those of Fable 5, except that finding vulnerabilities in source code is now allowed at all access levels. Opus 5's cyber classifiers, separate AI systems that screen requests, let it find vulnerabilities in source code but block scanning compiled binaries for flaws (a method Anthropic links more closely to attackers), penetration testing and writing exploits. Anthropic expects them to step in around 85 percent less often than on Fable 5. In Claude.ai, Claude Code and Claude Cowork, flagged requests fall back to Opus 4.8 by default. Members of Anthropic's Cyber Verification Program get a version with fewer restrictions.

In biology the looser safeguards make Opus 5 “our most capable generally available model for scientific research,” the company wrote, and biology requests blocked on Fable 5 now go to Opus 5 instead of Opus 4.8. Like earlier Opus models, it has no data retention requirement for general access, while traffic on Fable 5 and Mythos 5 is kept for 30 days. TechCrunch noted that the retention rule had raised concerns among some privacy conscious users.

85%
How much less often Anthropic expects Opus 5's cyber classifiers to intervene than those on Claude Fable 5 (Anthropic, 24 July 2026).

What the safety report says

Anthropic says it deliberately did not train Opus 5 on cyber tasks, yet the model improved on them anyway as it became more capable in general. The system card puts its cyber skills above Opus 4.8 but below Mythos 5: close at finding vulnerabilities, but “substantially behind Mythos 5 in its ability to exploit them.” Under its Responsible Scaling Policy, Anthropic places the model at its CB-1 level for chemical and biological risk (help with known weapons) but not CB-2 (novel weapons), applies the same ASL-3 protections as on Opus 4.8, and rates overall alignment risk as very low. On the company's automated behavioural audit, Opus 5 scored 2.3 for overall misaligned behaviour, the lowest of its recent models by Anthropic's own measure.

What happened next

On 22 September 2026 Anthropic released Claude Opus 5.5 at USD 4 per million input tokens and USD 20 per million output tokens, saying that at default settings it would cost 40 percent less than Opus 5 on typical workloads.