Anthropic Model 2 Tops Mythos 5 but Stays Internal

Anthropic Model 2 Is Somewhat More Capable Than Claude Mythos 5

In the executive summary of its report, Anthropic calls Model 2 an unreleased internal model that is somewhat more capable than Mythos 5, and Section 1.4 gives the details. For many internal tasks, Anthropic's rough qualitative sense is that Model 2 is a noticeable step up from Mythos 5, but not a jump like the one from Claude Opus 4.6 to Mythos Preview.

Anthropic has no current plans to offer it outside the company. Staff use Model 2 together with Mythos 5, and both see heavy use for coding, data generation and other agentic work. According to the report, Claude now writes a large majority of the code merged into Anthropic's production codebases.

Where Mythos 5 Is Available: Claude Fable 5 and Project Glasswing

You cannot use Model 2, but you can use Mythos 5 in one of two ways. Some customers reach Mythos 5 through Project Glasswing, while general users get it as Claude Fable 5, with extra safeguards added.

Two other models appear in the report. Claude Opus 5 was deployed internally on the report's coverage date and has since been released to the public, with capabilities generally below Mythos 5 and Claude Fable 5. Model 1 is roughly as capable as Claude Mythos Preview and Mythos 5, saw relatively little internal use, and had not been offered externally as of the coverage date.

What the August 2026 Anthropic Risk Report Says About Model 2

Anthropic expects Model 2 to behave much like Mythos 5 in both what it can do and what it tends to do. The two are the report's focus models: the most capable ones, and the ones used most inside the company. That expectation comes from Model 2's pre-internal-deployment review, though as of the coverage date the model had seen less use and testing than Mythos 5.

Anthropic also tried out a staged deployment with Model 2. It first put the model on internal surfaces with tighter blocks on dangerous actions, to collect real usage data in a safer setting before opening it up to unrestricted internal use. The approval process found no misalignment that was new or more worrying than what the report already covers for Mythos 5.

Where the Report Fits in the Anthropic Responsible Scaling Policy

The report is one of the ways Anthropic carries out version 3.4 of its Responsible Scaling Policy (RSP), and it covers more than one model. Anthropic aims to publish one every 3–6 months. A system card comes out with each model release, while this report looks at the company's activity as a whole, internal-only models included.

The overall risk rating goes up from “very low” to “low” in this report. Anthropic says its evidence probably still supports “very low,” but cites greater overall uncertainty as the reason for the change.

How Model 2 and Claude Mythos 5 Compare on SHADE-Arena and AECI

Model 2 edges out Mythos 5 on SHADE-Arena and appears to sit about 1.5 points higher on AECI. SHADE-Arena is scored as a stealth success rate, and Model 2 lands slightly above Mythos 5 and well below Mythos Preview.

For scale, many models, Mythos 5 included, stay under 1% stealth success with extended thinking turned on. Claude Mythos Preview does best on the default versions of these tests, and even it only reaches 6–10%.

AECI is Anthropic's internal version of Epoch AI's Epoch Capability Index (ECI). On limited data, Model 2 looks about 1.5 points higher than Mythos 5 there, with wide error bars — a smaller gain than the step from Mythos Preview to Mythos 5. Model 2 went through fewer chemical and biological evaluations, and most of them showed results equal to or weaker than Mythos 5's.

CoBench Tests an Anthropic Internal Model on Real R&D Tasks

CoBench is a newer internal test made of real Anthropic R&D tasks instead of stand-ins. One part of the report asks whether AI models could take the place of Anthropic's own technical staff, and Model 2 is one of the models it covers (Section 3.3). Anthropic's older task-based AI R&D suite had already reached the point where frontier models beat the human baseline on most tasks.

CoBench drops a model into Anthropic's infrastructure as it stood at some point in the past. The model then has to find the root cause of a problem Anthropic engineers really did solve. The version behind the published results contains 449 such problems.

According to the Figure 3.4.3.A caption, Mythos-class models do substantially better than other recent models but still fall short of what Anthropic would expect from a system able to fully replace its technical staff. The caption treats Mythos-class models as one group and gives no CoBench score for Model 2 by name.

Why CoBench Matters When Model 2 Isn't Released

With Model 2 unreleased and CoBench internal, outside readers have only the report's own description to work from. It gives three facts: Model 2 is covered by the substitution assessment, CoBench tests real work instead of stand-ins, and the Figure 3.4.3.A caption says Mythos-class models cannot yet fully replace technical staff.

What Was Redacted From the Anthropic AI Safety Report

At least three kinds of material are missing from the public version, and the report names each one itself. The file is called Redacted Risk Report August 2026, and its cover says “Risk Report: August 2026”.

• Leading indicators. Anthropic has been studying internal metrics it thinks could forecast later moves in AECI. It treats what they are and how they move as sensitive, and keeps them out of the public version (Section 3.5).
• Monitoring vulnerabilities. The report explains how internal usage is monitored, but the detailed weak points appear only in the unredacted version, so they do not leak into future search results and training data.
• Two appendices. The contents page lists Appendices 6.2 and 6.3 as redacted.

The contents page also lists Section 1.3.4, “Redaction scope and transparency”. As its title suggests, that section deals with the same question of what the public version omits.

Summary: Model 2 stays internal, and the report shows only part of it

Model 2 is an internal-only model that Anthropic rates as somewhat more capable than Mythos 5, with no current plans to release it. The report covers its review, its staged rollout and Anthropic's expectation that it behaves much like Mythos 5. It also gives two measurements, SHADE-Arena and AECI.

CoBench, with 449 problems, is the real-task test behind the report's discussion of replacing technical staff. The public version holds back leading indicators, monitoring vulnerabilities and two appendices. Related reading: OpenAI Preparedness Team: What the FT Reported.