Anthropic is using an unreleased artificial intelligence model internally that is reportedly more capable overall than every publicly available version of Claude. The model, referred to as “Model 2” in Anthropic’s August 2026 Risk Report, belongs to the company’s Mythos class and is already being used for coding, data generation, research and engineering work.
The disclosure provides a rare look at a frontier AI system that Anthropic has no current plans to release publicly. According to the company’s report, Model 2 is slightly stronger overall than Claude Mythos 5, although it is weaker in some areas and does not represent a major capability jump over its predecessor. Anthropic’s internal evaluation places Model 2 about 1.5 points above Mythos 5 on its AECI capability index.
Anthropic’s Model 2 Is More Capable Than Mythos 5
Model 2 appears to be an internal successor within Anthropic’s Mythos model class.
The company’s Risk Report says the model is slightly stronger overall than Claude Mythos 5. However, the improvement is relatively modest compared with previous jumps between major model generations.
On Anthropic’s internal AECI capability index, Model 2 reportedly scores about 1.5 points higher than Mythos 5.
Model 2 Vs Mythos 5
| Metric | Model 2 | Claude Mythos 5 |
|---|---|---|
| Publicly available | No | Yes |
| Overall capability | Slightly higher | Lower than Model 2 |
| AECI score | ~1.5 points higher | Baseline |
| Mythos class | Yes | Yes |
| Strengths | Coding, research, engineering, data generation | Broad frontier capabilities |
| External release planned | No current plans | Publicly available |
Anthropic also says Model 2 is weaker than Mythos 5 in some areas, meaning its advantage is not universal across every capability. :contentReference[oaicite:2]{index=2}
Model 2 Does Not Represent A Massive Capability Jump
The performance difference between Model 2 and Mythos 5 appears relatively small compared with earlier advances in Anthropic’s model lineup.
The company says Model 2 does not show the kind of major capability improvement seen when it moved from Opus 4.6 to the Mythos class.
Its approximately 1.5-point advantage over Mythos 5 on the AECI index is also smaller than the improvement from Mythos Preview to Mythos 5.
Relative Capability Progress
| Model Transition | Reported Improvement |
|---|---|
| Opus 4.6 → Mythos | Large capability jump |
| Mythos Preview → Mythos 5 | Larger than Model 2’s reported gain |
| Mythos 5 → Model 2 | ~1.5-point AECI improvement |
This suggests Anthropic is continuing to improve its frontier models, but the latest internal system may represent an incremental rather than revolutionary step.
Anthropic Is Already Using Model 2 For Coding
One of the most significant details in the disclosure is how extensively Anthropic is already using Model 2 internally.
The company relies on the model for coding, data generation, research and engineering tasks. In some cases, these tasks are handled through agents that can operate continuously.
Anthropic also says Claude now writes most of the code in its production systems.
This indicates that AI is becoming deeply integrated into Anthropic’s own software-development process rather than being used only as an optional coding assistant.
Internal Uses Of Model 2
| Use Case | Reported Role |
|---|---|
| Coding | Major internal application |
| Data generation | Used internally |
| Research | Used for research tasks |
| Engineering | Supports engineering work |
| AI agents | Can operate continuously |
| Production software | Claude writes most production code |
The development is particularly notable because Anthropic is effectively using its own frontier models to accelerate the development of future AI systems and infrastructure.
AI Agents Can Run Continuously
Anthropic’s report indicates that Model 2 is sometimes used through agents that operate continuously.
This is different from a conventional chatbot interaction where a user asks a question and waits for an answer.
An AI agent can instead be assigned a larger objective and continue working through multiple steps.
Traditional AI Vs Agentic AI
| Traditional AI | Agentic AI |
|---|---|
| Responds to individual prompts | Works toward larger objectives |
| Short interaction | Multi-step workflow |
| Human initiates each task | Agent can continue operating |
| Limited task duration | Can run continuously |
| Generates individual outputs | Can execute sequences of actions |
The use of Model 2 in continuous agents suggests Anthropic is testing increasingly autonomous approaches to software engineering and research.
Model 2 Went Through A Safety Review
Despite being kept internal, Model 2 underwent a review before being deployed within Anthropic.
However, the company says the model was not tested as thoroughly as Mythos 5.
Anthropic reported that its review did not identify new or more concerning forms of misalignment. The company currently rates the overall risk from misalignment as “low.”
Model 2 Safety Assessment
| Area | Anthropic’s Reported Assessment |
|---|---|
| Internal review | Completed |
| Testing compared with Mythos 5 | Less extensive |
| New worrying misalignments | None identified |
| Overall misalignment risk | Low |
| External release | No current plans |
The distinction between “not more concerning” and “fully risk-free” is important. Anthropic’s assessment is a risk classification, not a guarantee that the model cannot behave unexpectedly.
Why Anthropic Is Keeping Model 2 Private
Anthropic currently has no plans to release Model 2 externally.
The company has not indicated that the model is being withheld because of one specific newly discovered safety issue. Instead, the model remains part of Anthropic’s internal development and research environment.
The decision gives Anthropic additional time to evaluate the system and determine whether its capabilities and risks justify a public deployment.
This approach also allows the company to use the model internally without immediately exposing its most advanced capabilities to the wider market.
Anthropic’s Internal AI Development Loop Is Getting Stronger
The Model 2 disclosure highlights a potentially important feedback loop.
Anthropic develops increasingly capable models, uses them to perform coding and research internally, and then uses the resulting productivity gains to support further AI development.
That can accelerate the development process itself.
AI Development Feedback Loop
| Stage | Role |
|---|---|
| Model development | Build more capable AI |
| Internal deployment | Use models on company tasks |
| Coding assistance | Accelerate software development |
| Research support | Speed up experimentation |
| Data generation | Produce useful training material |
| Engineering productivity | Increase development capacity |
| Next-generation models | Feed improvements into future systems |
If frontier models become increasingly capable at software engineering and research, AI companies could potentially shorten their own development cycles.
Anthropic Is Becoming A Heavy User Of Its Own Models
The fact that Claude writes most of Anthropic’s production code is significant beyond Model 2 itself.
It suggests that Anthropic is using AI extensively throughout its own engineering organization.
The company’s internal use of Model 2 could therefore provide a real-world testing environment where researchers can observe how a more capable model performs on complex engineering tasks.
Anthropic’s Internal AI Use
| Function | AI Involvement |
|---|---|
| Production coding | Most code reportedly written by Claude |
| Engineering | Model 2 used internally |
| Research | Model 2 used |
| Data generation | Model 2 used |
| Agent workflows | Continuous operation in some cases |
This type of internal deployment can also reveal problems that may not appear in conventional benchmark evaluations.
Model 2’s Internal Status Shows The Gap Between Public And Frontier AI
The disclosure also illustrates an increasingly important characteristic of the AI industry: companies may have models internally that are ahead of what consumers can access.
Public users typically interact with models that have gone through extensive safety testing, product integration and operational preparation.
Internal research systems can be more experimental and may not be ready for broad deployment.
Internal Vs Public AI Models
| Internal Frontier Model | Public Model |
|---|---|
| Experimental | Productised |
| Limited access | Broad access |
| Ongoing evaluation | Extensive deployment testing |
| May have incomplete safety testing | More mature safety process |
| Used for research | Used by customers |
| Can remain unreleased | Commercially available |
Model 2 appears to occupy the first category.
Anthropic’s Risk Assessment Is Becoming More Important
The Model 2 disclosure comes from Anthropic’s Risk Report rather than a conventional product announcement.
That means the company’s description is focused heavily on capabilities, risks and safety assessments rather than consumer-facing features.
The report provides insight into how Anthropic evaluates increasingly capable AI systems before deciding whether they should be deployed more broadly.
This approach is particularly relevant as models become capable of performing more complex coding, research and agentic tasks.
The Cybersecurity Question Is Becoming More Important
Anthropic’s latest risk reporting has also highlighted growing concerns around the cybersecurity capabilities of advanced AI systems.
Separate reporting on the company’s August Risk Report said Anthropic raised its qualitative assessment of catastrophic misalignment risk in high-stakes situations from “very low” to “low,” citing increased uncertainty associated with recent cybersecurity evaluation incidents.
That broader context helps explain why the capabilities of unreleased frontier models are being scrutinised closely.
More capable coding and research systems can create substantial productivity gains, but the same capabilities can potentially increase the consequences of misuse or unexpected behaviour.
Model 2 Could Be A Preview Of Future Claude Systems
Although Anthropic has no current plans to release Model 2, the model could provide a glimpse into the direction of future Claude systems.
The most notable areas are not simply conversational ability but coding, autonomous agents, research and engineering.
If Anthropic eventually determines that the model can be deployed safely, some of its capabilities could potentially influence future public Claude releases.
However, there is currently no confirmed timeline for an external release.
What Is Known About Model 2
| Question | Current Answer |
|---|---|
| Does Model 2 exist? | Yes, according to Anthropic’s Risk Report |
| Is it public? | No |
| Is it stronger than Mythos 5 overall? | Slightly |
| Is it stronger in every area? | No |
| Is it used internally? | Yes |
| Coding use | Yes |
| Research use | Yes |
| Continuous agents | Yes, in some cases |
| Safety review completed? | Yes |
| Tested as extensively as Mythos 5? | No |
| Misalignment risk rating | Low |
| External release planned? | No current plans |
The Bigger Picture
Anthropic’s disclosure of Model 2 offers a rare glimpse into the company’s internal frontier-model development. The system is reportedly slightly more capable overall than Claude Mythos 5 and scores around 1.5 points higher on Anthropic’s internal AECI capability index. Yet the improvement is described as incremental rather than a major capability leap
More important may be how Anthropic is using the model. Model 2 is already being applied internally to coding, data generation, research and engineering, including workflows where AI agents can operate continuously. Claude also reportedly writes most of Anthropic’s production code, demonstrating how deeply AI-assisted development has become embedded in the company’s operations.
Looking Ahead
Anthropic’s decision not to release Model 2 means the model will remain primarily an internal tool for now. The company can continue evaluating its capabilities, studying its behaviour and using it to accelerate research and engineering without immediately exposing the system to millions of external users.
The larger significance is the widening gap between publicly available AI and the systems frontier labs are developing internally. As these models become better at coding, research and autonomous work, companies such as Anthropic may increasingly use AI to build and improve AI itself. How safely that development loop can operate will become one of the defining questions for the next generation of frontier models.
Get the day’s top stories in your inbox
One concise email. No spam, unsubscribe anytime.
