Anthropic already has the top two models on the Artificial Analysis Intelligence Index, but it has now also spoken about some unreleased ones that are even more capable.
In a new risk report covering the period up to July 15, 2026, the company disclosed the existence of an internal model called Model 2, describing it as somewhat more capable than Claude Mythos 5 and used heavily inside Anthropic for coding, data generation, and other agentic tasks.
The report states that Mythos 5 and Model 2 are being used extensively for research and engineering within Anthropic, both through interactive sessions and persistent agent deployments that run largely unsupervised. Claude, the company says, now authors a large majority of the code that gets merged into Anthropic’s own production codebases. Anthropic has long spoken about all internal code being written by AI, and Model 2 seems to have continued the trend.
Anthropic is careful about how far it takes the acceleration claim, though. The report says the company believes its internal AI R&D efforts are meaningfully faster with AI assistance than they would be without it, but stops short of calling it a doubling. “We are uncertain and measurement is difficult,” the report notes, an admission that sits somewhat awkwardly next to the rest of the disclosure, given how central “recursive self-improvement” has become to how Anthropic frames its own progress to regulators and the public.
Where Model 2 sits in Anthropic’s lineup #
The risk report identifies three internal models with frontier or near-frontier capabilities that hadn’t been publicly released as of the coverage date. One of them, Claude Opus 5, was already deployed internally at the time and has since gone public — Anthropic notes its capabilities are generally lower than those of Mythos 5 or Fable 5, which lines up with how Opus 5 has actually performed since launch, trading blows with Fable 5 at a lower cost rather than beating it outright.
The second is a model Anthropic refers to only as Model 1, which it says is broadly similar in capability to Mythos Preview and Mythos 5. Anthropic admits it hasn’t run extensive risk assessments on Model 1, and doesn’t believe it’s meaningfully more capable or differently aligned than its other current models. Internal staff, given the choice, largely preferred working with Mythos 5 instead, and usage of Model 1 was already low and falling by the time the report was compiled. Anthropic says it doesn’t expect the model to see a wider release, internally or externally, and treats any risk analysis of Mythos 5 as an effective ceiling for Model 1’s risk as well.
Model 2 is the one Anthropic actually spends time on in the report. The company describes it as a noticeable improvement over Mythos 5 on many tasks relevant to internal use, though not the kind of capability jump it saw going from Claude Opus 4.6 to Mythos Preview earlier this year. There are no current plans to release Model 2 externally, and Anthropic says it hasn’t put the model through its typical full suite of predeployment testing, which is why its confidence in Model 2’s capability profile is lower than what it holds for models it has actually shipped. Even so, the report goes on to discuss how Model 2 applies to each of Anthropic’s major RSP threat models, treating the coverage in later sections more briefly than it does for models the public can actually access.
What Model 2 says about Anthropic’s lead #
The way Anthropic is describing Model 2 is worth putting next to where the rest of the field actually stands right now. On the Artificial Analysis Intelligence Index, Claude Opus 5 currently leads at 63, with Fable 5 just behind at 62. But GPT-5.6 Sol and Grok 4.6 are tied at 61, a single point behind Fable 5 and only two behind Opus 5 itself. Kimi K3, an open-weights model out of Moonshot AI, sits at 60, within striking distance of all of them. Six months ago, the gap between Anthropic’s frontier and everyone else’s was wide enough to be the entire story of a model launch. Right now it’s closer to a rounding error, and it shows up across coding and agentic benchmarks too, not just the headline index.
That’s the backdrop against which Anthropic is choosing to describe Model 2 as a “noticeable improvement” over Mythos 5, explicitly not the kind of jump it saw going from Opus 4.6 to Mythos Preview. Read against a competitive field that’s basically caught up, that framing reads less like modesty and more like an honest account of where Anthropic’s internal roadmap actually is. The company that spent the better part of the past year describing its unreleased models as generations ahead of anything public is now describing its newest one as a solid step forward, tested less thoroughly than usual, with no plans to ship it. Whether that’s Anthropic being conservative about a model it hasn’t fully vetted, or a sign that the pace of internal progress has genuinely slowed as OpenAI, xAI, and Moonshot close in, is the question this disclosure leaves open rather than answers.