Claude Sonnet 4.8: What the Leak Actually Says, and Why the Pattern Doesn't Fit
A 'Sonnet 4.8' string appeared in a Claude Code source-map leak on March 31. Anthropic has shipped nothing. Polymarket is at 3% for today. And a version jump from 4.6 → 4.8 (skipping 4.7) would be unprecedented for Anthropic. Here's the honest read.
A “Sonnet 4.8” reference inside the Claude Code npm package — exposed when a source-map file accidentally shipped on March 31, 2026 — is the entire evidentiary basis for the Claude Sonnet 4.8 rumor that’s been building over the past eight weeks. Today, May 24, the Polymarket on a Sonnet 4.8 release by today closed at 3%. There is no Anthropic announcement, no model card, no API ID, no benchmarks, and — worth flagging upfront — no Sonnet 4.7 either.
The pattern break matters more than most of the coverage acknowledges. Below is a separation of what’s actually confirmed, what’s pattern-based speculation, and what the leak does and doesn’t suggest about Anthropic’s next Sonnet release.
What’s confirmed
| Item | Status |
|---|---|
| ”Sonnet 4.8” string exists in Claude Code source-map (Mar 31, 2026) | Confirmed |
| Source map: ~512K lines of TypeScript, accidentally shipped via npm | Confirmed |
Originating leak post carried a ? qualifier (low corroboration) | Confirmed |
| Anthropic has made no public statement about Sonnet 4.8 | Confirmed |
| No model card, no API ID, no pricing, no benchmarks published | Confirmed |
| Polymarket “Sonnet 4.8 by May 24” closed at 3% | Confirmed |
That’s the floor. Everything else is interpretation.
The pattern break
Anthropic has been remarkably consistent about pairing Opus and Sonnet minor versions:
| Opus | Sonnet | Pattern |
|---|---|---|
| Opus 4.5 | Sonnet 4.5 | Paired |
| Opus 4.6 | Sonnet 4.6 | Paired |
| Opus 4.7 (April 16, 2026) | — (no Sonnet 4.7) | Anomaly |
The current state is that Opus 4.7 has been out for five weeks with no matching Sonnet. The natural next move — by Anthropic’s own historical cadence — is Sonnet 4.7, not Sonnet 4.8. The Codersera analysis put it bluntly: “Skipping from Sonnet 4.6 straight to Sonnet 4.8 without a 4.7 would be a first.”
Three readings of the leaked string, in order of how much you should weight them:
- Planning artifact. Source-map leaks routinely surface internal version numbers that don’t correspond to actual upcoming releases. Engineering teams plan three or four versions ahead; product names get stamped into config files long before any model trains. “4.8” in a TypeScript bundle is consistent with a future-planning placeholder.
- Skipped-Sonnet hypothesis. Anthropic decides that Sonnet 4.7 isn’t worth shipping (maybe the training run didn’t separate cleanly from 4.6) and goes straight to a 4.8 that includes the Opus 4.7 work plus additional changes. This is plausible but unprecedented — Anthropic has historically shipped every Sonnet generation it trained.
- Sonnet 4.8 is imminent. The most aggressive read. The Polymarket 3% number is the market’s pricing on this scenario, and it’s pricing it close to zero.
The honest read: the most likely next Sonnet release is 4.7, matching the existing Opus 4.7. The 4.8 string is more likely a planning marker than a product release signal.
What the rumored features tell you (and don’t)
The leak’s downstream coverage has cited several feature names — KAIROS persistent agents, Undercover Mode, Mythos references — as evidence of what Sonnet 4.8 will include. Two notes on this:
- Names in a source bundle aren’t shipping features. A
KAIROSstring in compiled TypeScript could be a removed prototype, a feature-flagged experiment, a code name for an unrelated subsystem, or a planned capability shelved for later. The presence of a string in a source map is the lowest signal-to-noise channel that exists for understanding a product roadmap. - The pattern with previous Anthropic leaks. Internal names that surfaced in earlier source-map leaks — “Claude Mythos” being the most prominent — turned out to refer to the Opus line, not Sonnet. If KAIROS is the next persistent-agent product, it’s probably an Opus-tier or platform-level feature, not a Sonnet model change.
This doesn’t mean the rumored features are made up — it means they’re not specifically about Sonnet 4.8 in a way that would let you plan against them. Wait for the model card.
What Sonnet 4.6 is still doing for you today
Worth noting because the “Sonnet 4.8 is coming soon” narrative obscures the fact that Sonnet 4.6 is a remarkably capable model that’s been stable for three months:
- Pricing: $3 input / $15 output per 1M tokens (with 90% prompt-cache savings, 50% batch savings) — same as Sonnet 4.5
- Benchmark profile: 70% more token-efficient than Sonnet 4.5 on filesystem benchmarks with 38% accuracy improvement, per Anthropic
- Context window: 1M tokens flat (no surcharge)
- Coding default: Still the developer default for tool-driven coding workflows, with Gemini 3.5 Flash now competing on agent benchmarks but Sonnet 4.6 still leading on SWE-Bench Pro at the Pro tier
If your production stack is on Sonnet 4.6, there’s no reason to plan a migration today. The next thing you’ll actually evaluate against is Sonnet 4.7 (likely) or Sonnet 4.8 (less likely), and both will require their own benchmark validation when they ship.
What builders should actually do
Three concrete moves while the speculation cycle plays out:
- Don’t pin your eval timeline to leak speculation. The “Sonnet 4.8 is coming May 2026” framing has been around for eight weeks. We’re at the end of May. The Polymarket is at 3%. If your roadmap was waiting for Sonnet 4.8 to make a decision, you’re past the point where waiting pays off.
- Run your eval set against Opus 4.7 instead. It launched April 16, costs more ($5/$25 per 1M) but exists today. If your workload is reasoning-heavy and you’ve been holding off on a Sonnet upgrade, the model that’s actually available is Opus 4.7.
- Plan for “next Sonnet” not “Sonnet 4.8 specifically.” Whatever Anthropic ships next will need a system card review, your eval set run against it, and a deployment decision. The version number ends up being the least interesting part of that process.
When to actually pay attention
Three signals that would make this story load-bearing:
- An official Anthropic announcement. The Anthropic news feed is the only place worth watching. Source-map leaks have a near-zero correlation with release timing.
- A model ID appearing in the Claude API docs. Anthropic publishes model IDs as part of API documentation updates. Until
claude-sonnet-4-8-20XXXXXX(or similar) shows up there, the model doesn’t exist for API consumers. - A system card on Anthropic’s deployment safety hub — or its Anthropic equivalent. Every recent Claude release has shipped with documented evaluation data. No card means no release.
How to access Anthropic’s current models today
For LLM-side workloads, the WaveSpeedAI LLM endpoint gives you OpenAI-compatible access to current frontier text models — including Anthropic’s Claude family — behind a single API key. When Sonnet 4.7 (or 4.8, if Anthropic does break the pattern) actually ships, expect to compare it under that same endpoint within days.
Until then: Sonnet 4.6 and Opus 4.7 are the production options. The leak is a leak.
Sources: Codersera “Real vs Rumored”, Geeky Gadgets on the leak, Goldie Agency leak analysis, Anthropic’s Claude Sonnet 4.6 announcement, Claude API models overview.