OrcaRouter launches OrcaCyber Zero 1.5 with 1M-token context and reported 100% Cybench performance

OrcaRouter has released OrcaCyber Zero 1.5, a specialized artificial intelligence model designed for advanced cybersecurity research, vulnerability analysis, penetration testing, and multi-step security investigations. Announced on October 10, 2026, the model combines a 1-million-token context window, native function calling, structured outputs, and an API restricted to verified security researchers and organizations conducting authorized testing.

The launch is accompanied by substantial performance claims. According to OrcaRouter's evaluations, OrcaCyber Zero 1.5 achieved 100% on Cybench under unrestricted agent-execution conditions, 95.8% on an evaluable subset of CVE-Bench, 76.5% on SWE-bench Pro V2, and 93.9% on HumanEval+. These results suggest considerable capabilities in cybersecurity reasoning and software engineering, although they have not yet been independently reproduced, and the conditions under which each benchmark was conducted are essential to interpreting the scores.

OrcaCyber Zero 1.5 succeeds the earlier Zero 1.0 model, extending OrcaRouter's development of AI systems capable of investigating security weaknesses, reproducing vulnerabilities, and coordinating technical investigations through autonomous agents. The service is priced at $3 per million input tokens and $7.50 per million output tokens, with discounted cached-input processing, but access requires approval through the platform's Security Research tier.

The release raises two important questions for enterprise security teams: how much confidence should be placed in the reported benchmark results, and whether a cybersecurity-specialized model can improve real vulnerability detection and remediation without increasing operational or security risks.

··········

ORCACYBER ZERO 1.5 COMBINES LONG-CONTEXT REASONING WITH NATIVE AGENT CAPABILITIES.

The model is designed to examine complex software environments and coordinate multiple analytical steps, rather than operate exclusively as a conversational security assistant.

OrcaRouter describes OrcaCyber Zero 1.5 as a frontier cybersecurity model that has undergone additional training for security-specific workloads, including vulnerability discovery, reproduction of documented weaknesses, authorized penetration testing, security auditing, and advanced technical reasoning.

These activities require capabilities that extend beyond recognizing suspicious source-code patterns. A security investigation may involve understanding how application components interact, determining whether an identified weakness is reachable under realistic conditions, evaluating the evidence supporting a suspected vulnerability, and deciding which additional checks are necessary before a finding can be reported.

The model supports native tool calling, allowing an application to integrate it with appropriately controlled development and security tools. Structured outputs make the resulting information easier to incorporate into vulnerability-management systems, technical reports, or automated review pipelines, while its extended context capacity allows large volumes of code and documentation to be considered during the same analytical session.

The published specifications establish several operational characteristics, but OrcaRouter has not disclosed enough information to independently reconstruct the model's underlying architecture, parameter count, training data composition, or the complete post-training methodology used for Zero 1.5.

........

Technical characteristic	OrcaCyber Zero 1.5
Developer	Orca / OrcaRouter
Release date	October 10, 2026
Model identifier	orca/orcacyber-zero-1.5
Primary specialization	Cybersecurity research and authorized security engineering
Maximum context window	1,000,000 tokens
Maximum output	128,000 tokens
Input modality	Text
Output modality	Text
Tool and function calling	Supported
Structured outputs	Supported
API compatibility	OpenAI-compatible
Access requirements	Security Research approval
Publicly disclosed parameter count	Not specified
Published model weights	Not provided through the announced API offering

........

The absence of publicly disclosed architectural details matters when comparing Zero 1.5 with other security-oriented AI systems. Benchmark scores describe outcomes under particular conditions, but they do not establish how much of the performance comes from the underlying model, specialized post-training, agent orchestration, external tools, or the computational budget allocated to each task.

For enterprise deployments, those distinctions influence reproducibility, latency, infrastructure requirements, and the reliability of results outside controlled evaluations.

··········

ORCAROUTER REPORTS 100% ON CYBENCH AND 95.8% ON A SELECTED CVE-BENCH SUBSET.

The headline results are substantial, but the published percentages do not demonstrate that the model can identify or exploit every real-world vulnerability.

Cybench is a cybersecurity evaluation framework originally developed around 40 professional-level Capture the Flag tasks drawn from four competitions. The challenges cover different technical difficulties and provide environments in which AI agents can execute commands, inspect results, and work toward verifiable objectives.

Unlike evaluations that simply ask a model to explain a security concept, Cybench measures whether an agent can complete tasks through a sequence of actions. Performance therefore depends on the model's reasoning capabilities, the tools available to it, how execution is coordinated, and the limits imposed on the evaluation environment.

OrcaRouter reports that Zero 1.5 achieved 100% on Cybench using unrestricted agent execution. That qualification is central to understanding the claim: unrestricted execution describes the evaluation configuration, not unrestricted commercial access to the model or proof that the system operates without safety controls.

The publicly available model listing does not provide a complete task-by-task record, including the precise set of challenges completed, the number of attempts permitted, the agent configuration, or the computation used for each result. Without this information, the reported percentage cannot be treated as independently verified performance across every version and configuration of Cybench.

CVE-Bench measures a different capability. Developed to evaluate AI agents against critical vulnerabilities in real web applications, it uses controlled environments that allow researchers to test whether an agent can demonstrate exploitation of known weaknesses under conditions resembling practical software security assessments.

OrcaRouter reports 95.8% on the evaluable CVE-Bench subset, rather than on the complete benchmark. This limitation must be preserved when interpreting the result, because restricting an evaluation to tasks that could be executed successfully within the testing environment changes the denominator and may affect how representative the score is.

The company also reports results on two software-engineering benchmarks, providing some indication of the model's capabilities outside explicitly cybersecurity-focused evaluations.

........

Benchmark	Reported score	Evaluation focus
Cybench	100.0%	Multi-step cybersecurity challenges under unrestricted agent execution
CVE-Bench	95.8%	Real-world vulnerability scenarios on an evaluable subset
SWE-bench Pro V2	76.5%	Software engineering and repository-level problem solving
HumanEval+	93.9%	Code generation evaluated through expanded functional tests

........

All four scores were published by OrcaRouter with an evaluation date of October 10, 2026. They are vendor-reported results, not measurements independently reproduced by Data Studios.

The distinction between these benchmarks also prevents direct comparisons of their percentages. Completing a CTF challenge, demonstrating a known web vulnerability, resolving a repository-level software issue, and passing code-generation tests are fundamentally different activities, with separate definitions of success.

Even within a single benchmark family, changes in agent scaffolding, available tools, retry limits, task selection, and evaluation procedures can materially affect results. The 76.5% figure on SWE-bench Pro V2, for example, should not automatically be compared with scores obtained on other SWE-bench variants.

A further methodological concern is the difference between reproducing known vulnerabilities and discovering previously unknown ones. Strong performance on benchmark tasks with established success criteria provides evidence that an agent can reason through those particular environments, but zero-day vulnerability discovery requires evaluating unfamiliar software without knowing in advance whether an exploitable weakness exists.

Real security investigations also produce false positives, incomplete evidence, non-exploitable findings, and technical observations whose consequences depend on deployment conditions. None of those challenges disappears because a model reports a perfect result on a selected benchmark.

An independent assessment would need access to the task-level results, evaluation harness, model configuration, tool permissions, retry policies, and ideally previously unseen vulnerabilities. Until those elements become available, the announced scores are useful indicators of potential capability rather than definitive measurements of real-world effectiveness.

··········

THE 1-MILLION-TOKEN CONTEXT WINDOW COULD SUPPORT REPOSITORY-SCALE SECURITY INVESTIGATIONS.

Long-context processing allows developers to supply extensive technical material during a single model interaction, although the maximum supported window does not guarantee consistent reasoning quality across its entire length.

Cybersecurity investigations often require evidence distributed across many files and systems. An authorization weakness, for example, might involve request-routing logic, identity-management components, configuration files, middleware, and dependencies maintained by different development teams.

Analyzing these elements individually can obscure relationships between components, especially when the relevant behavior depends on interactions that are not apparent from any single file.

OrcaCyber Zero 1.5's 1-million-token context window provides capacity for much larger working sets, potentially incorporating source-code segments, internal technical documentation, architecture descriptions, configuration information, and previous investigation results without repeatedly discarding context.

The model also supports outputs of up to 128,000 tokens, enabling lengthy structured responses when required. This output limit is separate from the context specification and should not be interpreted as a guarantee that the maximum input and maximum output can always be combined in a single request without additional constraints.

For security teams, long-context support can be useful in three broad categories of work.

First, repository-level investigation can involve examining how separate components implement authentication, authorization, input handling, and access to sensitive resources. Having the relevant materials available together may reduce the need to reconstruct relationships across numerous disconnected requests.

Second, vulnerability triage can benefit from combining technical advisories, affected-version information, code changes, and deployment-specific details. Rather than relying only on a generic severity classification, the model can assist analysts in assessing whether the documented conditions appear relevant to their environment.

Third, incident investigation may require correlating logs, configuration changes, application behavior, and technical documentation. A large context window allows more evidence to be considered together, provided that the information is appropriately filtered and handled according to organizational security requirements.

However, increasing context capacity creates its own engineering problems. Processing large inputs consumes additional tokens, increases the amount of potentially irrelevant information presented to the model, and may introduce conflicting or outdated evidence.

Furthermore, published context capacity is not equivalent to independently verified long-context retrieval accuracy. A model may accept a million tokens while performing less reliably when the relevant evidence is buried among hundreds of thousands of unrelated tokens.

Security teams should therefore test retrieval accuracy, evidence attribution, and reasoning consistency at different context lengths rather than assuming that submitting an entire repository will automatically produce a more reliable assessment.

··········

ACCESS IS RESTRICTED TO AUTHORIZED SECURITY RESEARCH THROUGH AN OPENAI-COMPATIBLE API.

OrcaCyber Zero 1.5 is available as a hosted service, but ordinary API registration does not automatically authorize use of its advanced cybersecurity capabilities.

The model is served through OrcaRouter's OpenAI-compatible API, using the identifier orca/orcacyber-zero-1.5 and the endpoint https://api.orcarouter.ai/v1.

Developers already using an OpenAI-compatible client can integrate the model without adopting an entirely different request interface. Support for function calling and structured outputs also allows it to participate in existing agent orchestration systems, although compatibility with a familiar API format does not establish that every provider-specific feature behaves identically.

Access is controlled through the Security Research tier, which is intended for trusted security researchers, professional red teams, and organizations conducting authorized security assessments.

OrcaRouter describes the access process as requiring information about the engagement, a passkey, and acceptance of the relevant terms. The platform can therefore restrict use of the model according to the purpose of the investigation rather than making its capabilities automatically available to every account holder.

This distinction is particularly relevant because some of the model's advertised functions can also be used offensively. Vulnerability research and penetration testing are legitimate activities when conducted with appropriate authorization, but the same underlying capabilities may be misused against systems whose owners have not granted permission.

The availability of a specialized model consequently creates responsibilities at both the platform and application levels. Access verification limits who can use the service, while the organizations deploying it remain responsible for defining the permitted scope, controlling credentials, restricting external actions, and preserving evidence of what their automated systems have done.

For enterprise security teams, a controlled implementation could connect the model to internal source repositories, approved testing environments, and vulnerability-management workflows while requiring human authorization before consequential operations.

Model-level restrictions and API access controls do not replace application-level authorization. A security agent should not gain permission to modify systems, access production data, or execute sensitive actions solely because its underlying model can reason about those operations.

OrcaRouter's decision to gate Zero 1.5 reflects the operational sensitivity of cybersecurity-focused AI, particularly as autonomous agents become more capable of coordinating technical investigations over extended periods.

··········

PRICING IS $3 PER MILLION INPUT TOKENS AND $7.50 PER MILLION OUTPUT TOKENS.

The model uses conventional token-based pricing, with a discounted cache-read rate that can reduce the cost of repeated analysis involving shared technical material.

OrcaRouter lists OrcaCyber Zero 1.5 at $3 per million input tokens and $7.50 per million output tokens. Cached input tokens are charged at $0.75 per million when eligible for the published cache-read rate.

The prices match those listed for OrcaCyber Zero 1.0, meaning the successor does not currently introduce a higher base token rate despite its additional advertised capabilities.

For security applications, however, the cost of an individual request is often less informative than the expense of completing an entire investigation. Agentic workflows may repeatedly inspect evidence, evaluate alternative explanations, call external tools, and revise their conclusions before a finding can be accepted.

An investigation containing dozens of model calls can therefore consume substantially more tokens than a single prompt, particularly when large sections of source code or technical documentation are repeatedly included.

The following calculations illustrate the direct inference cost under different hypothetical workloads, using OrcaRouter's published standard rates and excluding caching.

........

Workload	Input tokens	Output tokens	Estimated cost
Brief security assessment	20,000	5,000	$0.0975
Repository investigation	100,000	20,000	$0.45
Extended analysis	500,000	50,000	$1.875
100 extended analyses	50 million	5 million	$187.50

........

These figures are derived estimates rather than measured costs from actual OrcaCyber security investigations. They assume that the stated token volumes are processed at standard rates and exclude additional infrastructure, external services, tool execution, storage, orchestration, and human review.

Caching changes the calculation when an application repeatedly reuses the same qualifying context.

For example, consider an investigation involving 100,000 input tokens and 20,000 output tokens. At standard rates, the input costs $0.30 and the output costs $0.15, producing a total direct inference expense of $0.45.

If 80,000 input tokens qualify for the advertised cache-read rate while the remaining 20,000 are billed at the standard input rate, the input expense falls to $0.12 and the combined cost becomes $0.27, assuming the same output length and excluding any separate cache-creation charges.

The saving is meaningful for repeated assessments involving a stable repository, shared technical documentation, or recurring analysis of the same software components, but it depends on actual cache eligibility and usage patterns.

A large context window can also increase expenditure if applications submit the maximum available material indiscriminately. The economic benefit of long-context processing is therefore strongest when it reduces repeated work or improves decision quality sufficiently to justify the additional input tokens.

The relevant purchasing metric is ultimately cost per validated security finding, not simply cost per million tokens. An inexpensive model that generates large numbers of unreliable findings may impose substantial downstream review costs, while a more expensive system can still be economically preferable if it consistently produces reproducible evidence and reduces analyst workload.

OrcaRouter's public performance indicators also require caution. Time-to-first-token measurements are displayed on the model page, but their values are dynamic and do not establish end-to-end completion times for lengthy security investigations.

Model response latency, total token-generation time, tool execution, and the time spent validating results must be considered separately when estimating operational performance.

··········

ORCACYBER ZERO 1.5 EXTENDS ORCAROUTER'S EARLIER WORK ON AUTOMATED VULNERABILITY REPRODUCTION.

The new release follows OrcaCyber Zero 1.0, which was already being evaluated through specialized agent systems for reproducing known software vulnerabilities.

OrcaRouter introduced the earlier model as a security-focused system intended to support vulnerability analysis, reproduction, and research. Its initial distribution was limited to approved testing, and the company subsequently described an agentic evaluation framework combining model routing, specialized memory, verification stages, and controlled execution environments.

In an October technical report, the OrcaCyber team described testing an agent system on CyberGym Level 1, which contains 1,507 vulnerability-reproduction tasks covering 188 open-source projects.

The reported system reproduced 1,477 of those vulnerabilities under the specified evaluation rules, achieving a 98.01% pass@1 result. The framework combined OrcaCyber Zero 1.0 with other model and agent components, meaning the reported performance belongs to the complete evaluation system rather than to the base model operating independently.

This is an important distinction when interpreting the newer release.

OrcaCyber Zero 1.5's announced Cybench and CVE-Bench results should not be presented as direct improvements over the earlier CyberGym evaluation, because the benchmarks assess different tasks and may use different agent configurations.

Nevertheless, the sequence of releases provides useful context for OrcaRouter's development strategy. The company has concentrated on security tasks with observable outcomes, where researchers can test whether an identified vulnerability can actually be reproduced instead of relying entirely on the model's written explanation.

For commercial security operations, this approach offers a clearer path to measuring usefulness. A model-generated description of a possible flaw has limited value unless the underlying weakness can be confirmed, its effect assessed, and the relevant correction identified.

The practical performance of Zero 1.5 will therefore depend on whether it can improve the reliability and efficiency of those stages across software that differs from the benchmarks used during development.

Independent testing should assess at least three outcomes: the proportion of suspected vulnerabilities that prove genuine, the time and expense required to establish reproducibility, and the proportion of findings that result in validated remediation.

Such measurements would provide a more useful basis for comparing specialized security models than headline benchmark percentages alone.

··········

THE COMMERCIAL OPPORTUNITY DEPENDS ON TURNING SECURITY REASONING INTO RELIABLE REMEDIATION.

OrcaCyber Zero 1.5 arrives as security organizations increasingly experiment with AI systems capable of investigating vulnerabilities and coordinating parts of the software assurance process.

The commercial appeal is straightforward. Large organizations maintain extensive software portfolios, depend on numerous open-source components, and face a continuous flow of vulnerability reports, code changes, and security alerts. Human analysts must determine which findings are credible, which systems are exposed, and which corrections can be applied without creating operational problems.

A specialized AI model could help reduce the effort required to examine this information, particularly when its tool integrations provide reproducible evidence rather than unsupported conclusions.

For software vendors, the strongest applications may involve reviewing authorized codebases before release, correlating vulnerability reports with affected components, and assisting developers in understanding whether a proposed correction addresses the underlying weakness.

Security consultancies and internal red teams may also benefit from models that can coordinate complex assessments within approved environments, provided that their activities remain constrained by explicit engagement rules and operational safeguards.

There are, however, reasons to avoid treating the launch as proof that autonomous cybersecurity research has been solved.

First, the reported benchmark results require independent reproduction. Second, the model's architecture, training methodology, and task-level evaluation details remain insufficiently documented for comprehensive technical assessment. Third, the commercial service is access-restricted, limiting the ability of ordinary developers to conduct broad comparative testing.

The ability to investigate a suspected vulnerability also does not imply that a model can safely deploy a correction. Security remediation requires understanding compatibility constraints, testing changes against affected software versions, assessing possible regressions, and coordinating deployment with the responsible engineering teams.

Those requirements become more demanding in critical infrastructure, healthcare, financial systems, and other environments where operational interruptions can produce consequences beyond the software itself.

The model's restricted-access distribution may help address some misuse concerns, but responsible adoption will still require independent authorization checks, isolated execution environments, logging, human oversight, and evaluation against the organization's own threat model.

The October 10 release establishes a concrete offering with clearly stated token prices, a large context window, and published cybersecurity benchmark claims. Whether OrcaCyber Zero 1.5 represents a substantial advance in practical security engineering will depend on reproducible testing and the quality of verified findings produced under real operational constraints.

For organizations evaluating the model, the decisive result will not be a perfect benchmark percentage, but the ability to discover genuine weaknesses, document them accurately, and support reliable remediation at a measurable cost.

··········

FOLLOW US FOR MORE.

DATA STUDIOS

datastudios.org