Anthropic presents Claude Opus 4.7 in code, an improvement of visual thinking

Anthropic PBC today open access in Claude Opus 4.7, the latest addition to its popular line of major language models.

The company claims that the LLM is better than its predecessor for coding jobs. Opus 4.7 scored 64.3% on the SWE-Bench Pro benchmark, about 10% higher than Opus 4.6. The new model also solved many tasks on the Terminal-Bench 2.0 dataset, including coding challenges involving the command line.

Although Opus 4.7 is better than its predecessor in many aspects, it is not the most capable LLM for Anthropic. Last month, the company previewed a model called the Claude Mythos that is highly capable of coding. The company did not make the final LLM widely available due to concerns that it could be misused by hackers.

Opus 4.7 includes a mechanism that detects attempts to use the cyber attack model. According to Anthropic, its engineers will collect data about the machine’s performance and use the findings to create Mythos monitoring systems. It hopes those precautions will allow the company to safely make “Mythos-class models” more widely available to customers.

Cybersecurity experts often research threats by imitating the tactics of hackers. As a result, the commands they send to Opus 4.7 have a good chance of being blocked by Anthropic. The company plans to address this issue with a new program called the Cyber ​​Verification Program. It will see Anthropic relax the lines of surveillance on the accounts of cybersecurity experts to allow for broader information.

Coding isn’t the only area where the Opus 4.7 performs better than the company’s previous models. According to Anthropic, it is also better for visual thinking tasks. Opus 4.7 can “see images at high resolution” and is very capable of producing visual assets such as user interface designs.

The model performs some functions almost as well as the Mythos. Opus 4.7 came within 1% of the frontier model score on GPQA Diamond, a set of undergraduate-level science questions. OpenAI Group PBC’s GPT-5.4, on the other hand, tops Mythos’ score on BrowseComp, a benchmark designed to test the online research skills of LLMs.

Anthropic is launching Opus 4.7 alongside a number of other product updates.

The company’s application programming interface enables developers to set so-called effort levels for its LLMs. Increasing the level of effort increases the quality of the output and the cost of disruption. Anthropic today introduced a new tier called xhigh that sits between the highest and second highest tiers. According to the company, this addition will enable developers to optimize their workload cost performance measurement.

Anthropic also added a second expense management feature to its API. Clients can now set job budgets, parameters that define the maximum number of tokens Claude may process while performing a job. The use of tokens directly affects the cost of speculative innovation.

Claude Code, an Anthropic program assistant, discovered a cutting command called ultrareview. Instructs the tool to scan the code file for bugs and other issues. Claude Code customers with a Max subscription can use the feature alongside another recently added automation capability, auto mode, which enables the assistant to complete long-running programming tasks much faster.

Image: Anthropic