Anthropic has launched stop task 5According to Anthropic and The Register, CloudFable 5 is claimed to offer token pricing at half the rate and near-marginal performance.
Opus 5 is priced at $5 per million input tokens and $25 per million output tokens, compared to $10 and $50 for Fable 5. Anthropic says the model “comes close to the marginal intelligence of the Cloud Fab 5 at half the price.” Opus 5 does not require data retention.
Anthropic has described the Opus 5 as its “most aligned model yet”. The company reports that it outperforms Opus 4.8 overall, with the largest improvements in agentic coding, computer use, and long-term knowledge work.
Pricing and benchmark status, shortened system prompts and changed reference guidance
Fable 5 is priced at $10 per million input tokens and $50 per million output tokens, while Opus 5 is priced at $5 and $25. For comparison, The Register notes that OpenAI’s GPT-5.6 Sol costs $5 per million input tokens and $30 per million output tokens.
Since token usage per task varies by model, cost per-task is a different metric. Artificial analysis reports the weighted average cost per Intelligence Index task as $2.75 for Fable, $2.03 for Opus 5 at maximum settings, $1.04 for GPT-5.6 Sol at maximum settings, and $0.95 for KM3. The Opus 5 leads the Artificial Analysis Intelligence Index with a score of 61, one point ahead of Fable.
Anthropic claims that the Opus 5 “offers significantly improved performance at the same cost as its predecessor, the Opus 4.8.” The system card states that it is comparable to Cloud Fable 5 and Cloud Mythos 5 and surpasses them in some respects.
Anthropic has reduced the size of the system prompt in its latest models. Cloud Code engineer Tharik Scheipper says 80 percent of the Cloud Code system prompts were removed, allowing customers to use more context. As a result, Anthropic has updated its guidance for preparing prompts, skills, and CLAUDE.md files.
“In your system prompts, skills, and CLAUDE.md files, you may need to simplify just like we did,” Shihipar said. claude doctor Can automatically optimize some signals and skills.
Register notes that Opus 5 responses can last longer than previous Opus models. Anthropic warns that “Cloud Opus 5’s default user-facing responses last longer than previous Opus models” and advises customers to “explicitly prompt for this” to control response length, as longer responses increase output token costs.
Cybersecurity, what Opus 5 means for API users, and availability
Opus 5 scores close to Mythos in finding vulnerabilities in open source code, with an OSS-Fuzz benchmark score of 80 percent compared to 79.4 percent for Mythos. It is significantly weaker at weaponizing those findings, succeeding in 4 out of 14 exploitation attempts compared to 13 out of 14 for Mythos.
Anthropic says the Opus 5’s cyber classifiers are proportionally less restrictive than the Fable 5’s. According to Anthropic, they “allow Opus 5 to find vulnerabilities in source code, but prevent ‘binary-based’ vulnerability scanning (a method more likely to be associated with malicious actors), penetration testing, and exploit generation.”
Register notes that the model is likely to be more useful for cybersecurity and biology work because it refuses to cooperate less frequently than some other models.
Anthropic is also introducing an automatic fallback function on its API. Requests deemed too dangerous to route on an Opus 5 or Fable 5 may be routed to a less capable model rather than being blocked directly.
For developers and enterprises using Anthropic API, there are practical points to evaluate:
- Just compare per-task costs rather than per-token pricing, as Opus 5 costs $2.03 per Intelligence Index task at maximum settings, compared to $2.75 for Fable, but GPT-5.6 Sol and KM K3 are lower on that measure.
- Budget for noted response verbosity by explicitly prompting for shorter responses where output token cost matters, since Opus 5 defaults to longer responses.
- Review existing system prompt, skill and CLAUDE.md files and simplify and run them
claude doctorAdapting cues and skills to new models. - If denials have previously blocked the workflow, evaluate the automated API fallback function, as it re-routes flagged requests rather than blocking them.
- Confirm data retention terms, as Opus 5 does not have a data retention requirement applicable to some covered models.
Anthropic released Opus 5 on Friday, July 24, 2026. According to The Register, the lack of data retention requirement is likely to be a sticking point for businesses.
Anthropic has not detailed how automated fallback routing selects less capable target models, or published a full list of task categories affected by binary scanning, penetration testing, and adjusted cyber classifiers beyond its named exploit generation restrictions.
Thanks for being a Ghax reader. The post Anthropic to release Cloud Opus 5 at half the price of Cloud Fable 5 appeared first on Ghacks.
