Claude Sonnet 4.8 & Opus 4.8: What Actually Shipped (2026)

Last updated: June 2026

When the Claude Code source code leaked on March 31, 2026, the community immediately found what Anthropic tried hardest to hide. Not just unreleased features — but explicit references to future model versions and a complete internal codename system that maps Anthropic’s past, present, and planned AI lineup.

The irony is almost poetic: the leaked codebase contains an „Undercover Mode” that explicitly instructs Claude to never reveal internal model codenames in public git commits. The forbidden list includes „animal names like Capybara, Tengu” and „unreleased model version numbers (e.g., opus-4-7, sonnet-4-8).” Those exact strings — intended to be hidden — are now the most searched terms from the leak.

This article decodes the full codename system, maps what each animal name corresponds to, and analyzes the internal quality metrics that Anthropic never intended to publish.

What model codenames did the leak reveal?

Anthropic uses an animal-based codename system for internal model development. The leaked source code, analyzed by independent researchers and documented on GitHub, confirms the following mappings:

The naming convention is consistent: small-to-medium mammals and mythological creatures. Fennec (the fox), Capybara, Numbat — each one a development codename that gets replaced by the marketing name before public release. What makes this leak unique is that we can now trace the evolution path of models through the migration functions embedded in the code.

What do the migration functions tell us about Anthropic’s release history?

The leaked codebase contains model migration functions — code that handles transitioning users from one model version to the next. These functions act as a fossil record of Anthropic’s internal release timeline:

// Model migration history — extracted from Claude Code v2.1.88
migrateFennecToOpus        // Fennec → Opus (codename to product name)
migrateSonnet1mToSonnet45  // Sonnet 1M context → rebranded as Sonnet 4.5
migrateSonnet45ToSonnet46  // Sonnet 4.5 → Sonnet 4.6
resetProToOpusDefault      // Pro users reset to Opus as default model

Reading these in sequence tells a clear story. Fennec started as an internal Opus development build. The 1-million-token context window Sonnet was initially a separate variant that got folded into the Sonnet 4.5 release. When Sonnet 4.6 shipped on February 17, 2026, another migration function handled the transition for existing users. And at some point, Pro tier subscribers had their default switched to Opus — likely during the Opus 4.6 launch alongside Agent Teams in early February.

The most revealing detail? The Undercover Mode instructions list opus-4-7 and sonnet-4-8 as examples of „unreleased model version numbers” that must never appear in public commits. This isn’t hypothetical — it confirms these version numbers exist internally at Anthropic, even if the models behind them aren’t finished yet.

What is Capybara and how does it connect to Claude Mythos?

This is where two separate leaks converge. Five days before the Claude Code source leak, on March 26, 2026, a different incident exposed Anthropic’s internal content management system. A misconfigured data store left nearly 3,000 unpublished assets publicly accessible — draft blog posts, PDFs, images, and internal memos describing an unreleased model called Claude Mythos.

Anthropic confirmed that leak was real. A spokesperson acknowledged the company had completed training on Mythos, describing it as a „step change” and „the most capable we’ve built to date.” The leaked blog drafts positioned Mythos not as Opus 5, but as the first model in a new tier above Opus.

The Claude Code source leak now confirms the internal codename: Capybara = Mythos. The code references „Capybara” as a Claude 4.6 variant, but the Mythos leak documents describe it as sitting above the existing Opus tier. The most likely interpretation is that Capybara started as an experimental 4.6 variant and evolved into something significantly more capable during development — powerful enough to warrant its own tier.

This matches the leaked cybersecurity assessment that moved markets. Mythos was described internally as being „currently far ahead of any other AI model in cyber capabilities,” a statement that triggered 4–7% drops in cybersecurity stocks including CrowdStrike, Palo Alto Networks, and Zscaler on March 27–28.

What do the internal quality metrics reveal?

The most technically significant finding from the source leak isn’t a codename — it’s internal benchmark data that Anthropic would never willingly publish.

Comments attached to Capybara’s development code reveal that its eighth iteration (v8) carries a false claims rate of 29–30%. This is not a typo. It’s actually a regression from v4, which had a false claims rate of 16.7%. The model got almost twice as unreliable between versions.

This data point is extraordinary for several reasons:

It shows that more capable ≠ more reliable. The AI industry’s public narrative focuses on benchmark scores going up — SWE-bench, AIME, ARC-AGI. Anthropic’s internal data shows that as Capybara gained capabilities, its tendency to make false claims nearly doubled. This is the classic capability-reliability tradeoff that researchers have warned about but companies rarely quantify publicly.

It reveals the „assertiveness counterweight.” Alongside the false claims metric, the code references a mechanism designed to prevent the model from becoming „too aggressive in its refactors.” This suggests that as Capybara became more powerful in code generation, it also became more confident in making changes — including wrong ones. Anthropic built a dedicated counterweight to throttle this tendency, which implies the unweighted model was aggressively rewriting code in ways that introduced errors.

It sets a benchmark for the entire industry. Before this leak, nobody outside Anthropic (or its competitors) knew what internal false claims rates look like for frontier large language models. A 29% false claims rate in a model described as a „step change” suggests that even the most advanced systems hallucinate at rates that would be unacceptable in high-stakes applications without human oversight.

What does the „assertiveness counterweight” mean in practice?

One of the subtler findings from the leak is a mechanism called the assertiveness counterweight. This is a guardrail specifically designed to prevent Capybara from being too aggressive when modifying user code.

The problem it solves is straightforward: as deep learning models become more capable at code generation, they also become more confident in making changes — including changes that are wrong. A model that hesitates before rewriting a function is annoying but safe. A model that confidently rewrites an entire module and introduces subtle bugs is dangerous.

The assertiveness counterweight suggests Anthropic discovered this failure mode during internal testing of Capybara and had to engineer a dedicated mechanism to throttle it. This isn’t a prompt-level instruction like „be careful with code changes” — it’s a system-level architectural constraint built into the model’s operating logic.

For developers building with AI agents, this is a critical design pattern to understand. More capable models need more sophisticated restraint mechanisms. The pattern holds whether you’re building a coding agent, a trading bot, or any autonomous system: capability without calibrated confidence is a liability.

What does the „over-commenting” problem reveal?

A less dramatic but equally telling detail from the leak: developers noted that Capybara has an over-commenting tendency. The model adds excessive code comments — explaining obvious operations, annotating every function, padding code with verbose documentation that no experienced developer would write.

This matters because over-commenting is a classic signal of AI-generated code. In a world where Anthropic’s Undercover Mode explicitly instructs Claude to hide its AI identity in open-source contributions, a model that compulsively comments every line of code is undermining its own stealth. It’s the AI equivalent of a spy who can’t stop explaining what they’re doing.

More importantly, over-commenting increases the token count of every response, raising costs for API users. For high-volume users of machine learning pipelines that use Claude as a coding agent, even a 20% increase in output tokens translates directly to 20% higher bills. This isn’t a style preference — it’s a cost problem at scale.

How does Anthropic’s model versioning compare to competitors?

The leak provides an unusually clear picture of how Anthropic stages model releases internally. Comparing this to what we know about OpenAI and Google:

The most significant structural difference is the new Mythos tier. Both OpenAI and Google have maintained three-tier hierarchies (cheap/balanced/powerful). Anthropic introduced a fourth tier — a „beyond Opus” class that sits above their previous best, which shipped as Fable 5 and Mythos 5 on 9 June 2026. It represents a genuine architectural leap rather than an incremental version bump.

What should developers and AI practitioners expect next?

The combination of the Mythos CMS leak (March 26) and the Claude Code source leak (March 31) provides an unusually complete picture of Anthropic’s roadmap. Three conclusions stand out:

The Opus strings were real; the “Sonnet 4.8” string was not (update). Opus 4.7 shipped and was quickly superseded by Opus 4.8 (GA 28 May 2026). But despite appearing in the same forbidden-strings list, no Sonnet 4.8 ever launched — Anthropic’s Sonnet line still sits at 4.6, and the next leap went to the Mythos-class tier instead. The lesson: a version string in operational code signals intent, not a guaranteed release.

Capybara/Mythos was the real next-generation leap — and it landed (update). On 9 June 2026 Anthropic launched the Mythos-class tier as Claude Fable 5 (generally available) and Mythos 5 (restricted, then suspended). Exactly as the codenames implied, this was not “Opus 5” but a qualitatively new tier sitting above Opus. The leak’s framing of Capybara as a generational jump rather than an increment proved correct.

The false claims regression is the most honest AI metric ever accidentally published. Anthropic’s internal acknowledgment that their most capable model hallucinates at nearly 30% — and that this is worse than a less capable version — is the kind of data the industry needs more of. It confirms what practitioners building with RAG systems and context engineering already suspect: scaling capabilities does not automatically scale reliability, and external verification layers remain essential.

Related Guides

Building with these Claude models? Start here:

• Build AI Agents with Claude — the Agent SDK and Skills for shipping production agents on Opus and Sonnet.
• Best AI Coding Assistants 2026 — where Claude ranks against Cursor, Copilot and the rest.
• Best AI Agent Frameworks 2026 — the frameworks to orchestrate Claude-powered agents.