Claude Sonnet 4.8: What the March 2026 Leak Proves — and What Remains Pure Speculation

On March 31, 2026, Anthropic published version 2.1.88 of the npm package @anthropic-ai/claude-code — with a 59.8 MB source map file that should never have become public. The reason: someone in the build process had forgotten to add *.map to the .npmignore. Boris Cherny, head of Claude Code, later publicly confirmed that it was a "plain developer error". The result: 512,000 lines of TypeScript source code, 1,900 files, and thereby the largest accidental leak in the history of AI tools to date.

Since then, countless speculations about an allegedly imminent Claude Sonnet 4.8 have been circulating in tech forums, on X, and on English-language specialist blogs. The problem: most of these reports mix verifiable facts from the leak with wishful thinking. Any developer or CTO who makes budget or architecture decisions on this basis risks misinvestment. This article draws a strict line between what the source code objectively proves and what remains interpretation.

What the leak objectively reveals about Sonnet 4.8

The only verifiable hint of a model called "Sonnet 4.8" is found in the so-called Undercover Mode configuration of the Claude Code client. This mode activates automatically when an Anthropic employee (USER_TYPE === 'ant') uses Claude Code in a public repository that does not belong to Anthropic. Its function: it strips Co-Authored-By attributions from Git commits, blocks internal Slack channel names and — crucial for our analysis — forbids the mention of not-yet-released model versions.

In the list of forbidden strings that must be filtered in this Undercover Mode, the following entries appear:

• opus-4-7 — since released (April 16, 2026)
• sonnet-4-8 — unreleased to this day
• mythos — since released as Mythos Preview (April 7, 2026)
• capybara — internal codename, presumably belonging to Mythos
• numbat — unreleased, unknown model

That is the entire concrete proof. The mere existence of the string sonnet-4-8 in an internal filter list proves that Anthropic uses this version number internally — nothing more. In the entire leak there is no model card, no API documentation, no benchmarks, no feature flags that specifically reference Sonnet 4.8, and no migration function targeting this version string. The migration history in the code ends at migrateSonnet45ToSonnet46.

What we explicitly do not know

The clarity of what we do not know is just as important as the facts themselves. In the context of AI roadmaps, half-information quickly tempts people into strategic missteps. Here is a list of what the leak does not provide:

No release date. Anthropic has communicated no timeline for Sonnet 4.8. The prediction markets (Manifold, Polymarket) priced a release before May 24, 2026 at roughly 3 percent — so the market bet heavily against a near-term release. The pattern of past releases (Sonnet 4.5 in September 2025, Sonnet 4.6 in February 2026) suggests a gap of three to four months between point releases instead. Even if Anthropic is working on a Sonnet update, the version number 4.8 is no guarantee that the next public Sonnet model will not be named differently internally or even skipped.

No specific features or capabilities. Other blogs speculate about "vision upgrades", "better coding benchmarks" or "higher image resolution" for Sonnet 4.8. These guesses derive solely from the fact that Opus 4.7 received improvements in exactly these areas in April 2026 (2,576 pixel edge length, 98.5 percent visual acuity, Task Budgets, file-system memory recall). However, there is not a single hint in the leak that these features will be carried over to the Sonnet tier — let alone in a version 4.8. That is reasoning by analogy, not evidence.

No pricing information. Sonnet 4.6 currently costs 3 US dollars per million input tokens and 15 US dollars per million output tokens. Some sources claim Sonnet 4.8 will "probably" have the same price. That is plausible, because Anthropic has kept Sonnet prices stable across several generations — but it is not information derived from the leak, it is an extrapolated guess.

No API model identifier. Anthropic documents its API models in a public list. For Sonnet 4.8 no entry exists there. The leak likewise contains no API endpoint or internal routing code that references claude-sonnet-4-8 or a similar ID.

The leak hierarchy: what else is in the code

To place Sonnet 4.8 correctly, it is worth looking at the other finds from the March leak. They show what Anthropic is working on in parallel — and how far advanced these features are compared to a possible new Sonnet version.

KAIROS — the always-on background agent. With more than 150 references in the source code, KAIROS is the most advanced unreleased component. It is a persistent daemon that runs even when the developer is not at the machine. KAIROS receives periodic <tick> prompts, decides autonomously whether it should act, and has exclusive tools such as push notifications, file delivery and GitHub webhook subscriptions. A 15-second blocking budget prevents resource monopolization. The feature is fully implemented, but hidden behind compile-time feature flags. The architectural relevance is enormous: KAIROS would upgrade Claude Code from a reactive chat tool to a proactive, self-directed system partner. For Sonnet 4.8 there is no comparable evidence within the leak.

Capybara and Mythos — the new performance tier above Opus. Capybara is the internal codename for what Anthropic publicly calls Claude Mythos — a model explicitly positioned above Opus and which does not simply mean "Opus 5". The leak contains internal quality metrics that are remarkably transparent for the industry: Capybara v4 had a false-claims rate of 16.7 percent, in v8 this rose to 29 to 30 percent. In other words: even the most capable internal iteration hallucinates in almost every third relevant test case. The code also documents an "assertiveness counterweight" meant to prevent the model from rewriting user code too aggressively. This data is more valuable than any marketing benchmark, because it consists of genuine internal engineering notes.

Numbat — the unknown model. Alongside Capybara, another tier codename exists in the code: Numbat. One comment reads verbatim "Remove this section when we launch numbat". This proves that Numbat is a standalone model in the pre-launch phase — but there is zero information on tier, capabilities or timeline.

ULTRAPLAN, Coordinator Mode, Buddy System. The leak contains at least 44 compile-time feature flags and 22 runtime gates. ULTRAPLAN offloads complex architecture planning into remote cloud containers with up to 30 minutes of compute time. Coordinator Mode enables multi-agent swarms in which one Claude instance can spawn and manage parallel worker agents. The Buddy System is a Tamagotchi-like ASCII companion in the terminal — apparently conceived as an Easter egg for April 1, 2026. All of these features are fully present in the code, but set to false for external builds.

Why the version number 4.8 instead of 4.7 confuses

One of the most frequently asked questions in the forums is: why is there no Sonnet 4.7? Anthropic released Opus 4.7 on April 16, 2026. Historically, the Sonnet version followed the Opus release within one to four weeks. Some analysts therefore speculated that a Sonnet 4.7 would appear shortly after Opus 4.7. Instead, only the string "sonnet-4-8" was found in the leak.

There are several possible explanations for this, all equally unproven: First, Anthropic could have decided to jump the Sonnet series directly from 4.6 to 4.8 to restore parity with the Opus version (Opus 4.7 exists, so Sonnet 4.8 will be its counterpart). Second, there could have been an internal Sonnet 4.7 that was never released and whose features were merged into 4.8. Third, the numbering could be arbitrary and say nothing about the actual feature scope. None of these hypotheses is confirmed by the leak.

What developers and companies should do now

From our perspective as an agency that advises mid-market companies on AI integration, three concrete recommendations can be derived — regardless of whether Sonnet 4.8 appears in two weeks or in six months.

First: do not make architecture decisions based on leaks. Sonnet 4.6 is an excellent model that has been running stably since February 2026 and shines with a one-million-token context window and 70 percent higher token efficiency over 4.5. Anyone building production code on Sonnet 4.6 today loses nothing when 4.8 appears — thanks to abstraction layers in the API integration, a model switch is usually a configuration change, not a rewrite. Leak-based pre-emptive migrations are technical debt without return.

Second: keep an eye on the agentic features. If the leak makes one thing clear, it is that Anthropic is shifting its focus massively from pure model upgrades to agentic capabilities. KAIROS, Coordinator Mode and ULTRAPLAN are not marginal experiments but fully implemented subsystems. For developers this means: the question will no longer be "Which model do I use?" but "How do I orchestrate multiple agents working autonomously?". Anyone optimizing their infrastructure for pure prompting today is missing the real paradigm shift.

Third: train source literacy. Most articles about Sonnet 4.8 in the German- and English-speaking world reinforce one another without going back to primary sources. The original tweet that discovered the string itself carried a question-mark qualifier. Wavespeed.ai described the leak's credibility as "weak" — and rightly so. We recommend that every technology decision-maker check at least two independent primary sources before strategic decisions. For this leak those are the GitHub mirror of the original source maps (by now largely removed through DMCA takedowns) and the detailed analyses by DecodeTheFuture.org and Kuber Studio, which documented the original code promptly.

Conclusion

Claude Sonnet 4.8 exists as an internal version string at Anthropic. That is all the Claude Code leak from March 31, 2026 objectively proves. Every statement about features, prices, release dates or benchmarks is speculation — partly plausible, but never derived from the leak. The true value of the leak lies elsewhere: it shows that Anthropic already possesses models, agent frameworks and safety mechanisms that could push the public perception of "AI assistance" forward by a decade.

For companies the consequence is clear: invest in robust, model-agnostic architectures and in understanding agentic workflows — not in waiting for the next version number. Sonnet 4.6 is production-ready today, KAIROS and Mythos show where things are heading. Whoever can tell the difference wastes no budget on rumors.

Frequently asked questions about Claude Sonnet 4.8