TokenMix Research Lab · 2026-05-25

Claude Sonnet 4.8 Release Date: What the Leak Proves vs Doesn't

Claude Sonnet 4.8 has not shipped. As of 2026-05-25, Anthropic has issued no announcement, no model card, no API ID. The case for its existence rests entirely on a 59.8 MB JavaScript source-map file that shipped accidentally with @anthropic-ai/claude-code npm v2.1.88 on 2026-03-31, which contained the string sonnet-4-8 in a security-filter list alongside opus-4-7 and mythos. Polymarket closed at 3% probability on a mid-May ship date.

That's the entire evidence base. Everything else — KAIROS, Undercover Mode, the 82-84% SWE-Bench projection, the $3/$15 price floor, the early-May release window — is inference layered on inference. This post separates what the leak actually proves from what it doesn't, and gives developers a practical decision framework for what to do while Anthropic stays silent. All claims verified 2026-05-25 against source documents.

Table of Contents

• Quick Verdict — Confirmed / Likely / Speculation
• What the Leak Actually Contains
• Why the 4.6 → 4.8 Skip Breaks Pattern
• Expected Pricing If It Ships
• Expected Benchmark Improvements (Speculative)
• Production Decision Framework
• What to Do While You Wait
• FAQ
• Sources

Quick Verdict — Confirmed / Likely / Speculation

The hard data is small. The interpretation around it is large.

What the Leak Actually Contains

The factual base, verified against multiple coverage sources:

The artifact. On 2026-03-31, version 2.1.88 of @anthropic-ai/claude-code shipped to npm with a 59.8 MB JavaScript source map (cli.js.map) intended for internal debugging. Security researcher Chaofan Shou flagged it at approximately 4:23 AM ET on X. The file contained roughly 1,900 source files. GitHub mirrors appeared within hours.

What's in the source map. The leak exposed approximately 512,000 lines of TypeScript. Among the discoveries:

• A security filter list of "forbidden" model strings — names that Anthropic-internal builds of Claude Code must not surface in user-facing output. The list includes opus-4-7 (which subsequently shipped 2026-04-16), mythos, and sonnet-4-8.
• References to KAIROS, a "persistent background agent" mentioned over 150 times. Described in code comments as a daemon-mode capability that receives periodic tick prompts and can initiate actions like notifications or GitHub webhook handling.
• Undercover Mode, a feature that activates when Anthropic employees use Claude Code on non-Anthropic repositories. It strips Co-Authored-By attribution, blocks references to internal Slack channels, and forbids mentioning unreleased model names. This is the security layer that the forbidden-strings list serves.

What the leak does not contain. No model card. No benchmark results. No pricing table for unreleased models. No release-date metadata. No documentation describing what sonnet-4-8 actually is — whether it's a planned product, an internal training checkpoint, a reserved string for a future iteration, or a typo. The source-map context is a string in a filter list, nothing more.

WaveSpeed's editorial framing is the cleanest summary: "The presence of a string in a source map is the lowest signal-to-noise channel that exists for understanding a product roadmap."

Why the 4.6 → 4.8 Skip Breaks Pattern

Anthropic has shipped paired Opus and Sonnet minor versions consistently:

The skeptical case is structural: Anthropic has never skipped a Sonnet minor version. A jump from 4.6 to 4.8 would be the first time in the model family's history. Sonnet versions ship 1-4 weeks after the corresponding Opus — and the corresponding Opus to 4.8 would presumably be Opus 4.8 (not yet announced).

There are three plausible readings:

Reading 1 — The leak is correct. Anthropic deliberately skipped 4.7 for Sonnet because the internal training checkpoint named "sonnet-4-7" did not meet quality bars, so the next public release got bumped to 4.8 to maintain version alignment with a future Opus 4.8. This requires assuming Anthropic broke its own convention for unstated quality reasons.

Reading 2 — The string is a planning artifact. The sonnet-4-8 string in the security filter is a reserved placeholder for a far-future release, not the next Sonnet. The next Sonnet release will be 4.7, matching the established pattern. This is the WaveSpeed reading.

Reading 3 — The string is internal nomenclature. sonnet-4-8 could refer to an internal model variant that won't ship under that public name. Anthropic regularly trains and discards checkpoints; not every internal model ID becomes a product.

The honest position: all three are plausible. No external evidence resolves the ambiguity. Developers building production systems should not assume reading 1 is correct until Anthropic confirms it.

Expected Pricing If It Ships

Sonnet pricing has been remarkably stable across generations. If Sonnet 4.8 (or 4.7) ships, the working assumption is the same floor:

Cache pricing is 90% off input on hits and 25% premium on writes — Anthropic's standard. Long-context surcharge (>200K input) typically scales 2x per Anthropic's published policy.

This stability is itself useful intelligence. It means a Sonnet 4.8 launch should not shift cost models for teams already running on 4.6. The pricing equilibrium against competitors stays put: Sonnet remains roughly 9x more expensive than DeepSeek V4-Pro on output ($15 vs $0.87 per MTok post-permanent-cut) and roughly 8x cheaper than Opus 4.7 ($15 vs $25 output).

Expected Benchmark Improvements (Speculative)

Anthropic's pattern across Opus generations is approximately +2-4 points on SWE-Bench Verified per minor revision, with vision improvements compounding faster. Applied to Sonnet 4.6's baseline:

These are the published predictions. None are confirmed. If Sonnet 4.8 ships at the projected SWE-Bench Verified score (82-84%), it would land between Qwen 3.6-Plus (78.8) and the open-weights Kimi K-class top tier, and roughly parity with the speculative Opus 4.7 internals.

The realistic caveat: even Anthropic's own version increments aren't monotonic. Some minor revisions improve coding while regressing on reasoning, or vice versa. Until the model card and independent benchmarks land, treat every projection as a target, not a delivered capability.

Production Decision Framework

For teams running on Sonnet 4.6 today and asking "should I plan around 4.8?":

Decision heuristic: if a Sonnet 4.8 launch in the next 4 weeks would force you to refactor your stack, your stack is too tightly coupled to model versions. The right architecture supports a config-flip upgrade — not a code change.

What to Do While You Wait

Three practical moves that pay off whether Sonnet 4.8 ships in 2 weeks or 2 months:

1. Decouple model selection from code. Move model IDs into environment variables or a configuration service. When Sonnet 4.7 or 4.8 (or whatever the next release is named) lands, a single env-var change should be the only modification needed to roll out.

# Bad
model = "claude-sonnet-4-6"

# Good
model = os.environ.get("CLAUDE_SONNET_MODEL", "claude-sonnet-4-6")

2. Build a fallback chain. Anthropic launch-day capacity is regularly constrained for the first 24-72 hours. A graceful fallback to the previous version (or an alternative provider) preserves uptime.

3. Run the eval suite you wish you had ready. Most teams discover during a model upgrade that their evaluation harness measures the wrong things. Build (or borrow) a workload-specific eval now, against Sonnet 4.6, so the day Sonnet 4.7/4.8 ships you can quantify whether the upgrade is worth taking immediately or worth delaying.

The Claude API pricing guide on TokenMix.ai maintains current pricing and cache math for all live Anthropic models — useful for cost re-baselining when the next version lands.

FAQ

When will Claude Sonnet 4.8 be released? No confirmed date. NxCode projects early-to-mid May 2026 based on Opus 4.7's April 16 release; that window has now passed. WaveSpeed argues the next release will be 4.7, not 4.8. Polymarket-listed odds on a Sonnet 4.8 mid-May ship date closed at 3%. As of 2026-05-25, Anthropic has issued no statement.

Is Sonnet 4.8 confirmed real? Only the string sonnet-4-8 in leaked source-map code is confirmed. Whether it corresponds to a planned public release, an internal training checkpoint, or a placeholder is unverified. The Anthropic news feed is the authoritative source.

Will Sonnet 4.8 pricing change from $3/$15? Likely not. Anthropic has held Sonnet input/output at $3/$15 per MTok across at least five minor revisions. Cache pricing (90% off input on hits, 25% premium on writes) has been equally stable.

What is KAIROS? A persistent background-agent capability referenced over 150 times in the leaked code. Described as a daemon-mode feature that receives periodic tick prompts and can take actions like sending notifications or processing GitHub webhooks. Whether it ships with Sonnet 4.8, a later Claude Code update, or never is unverified.

What is Undercover Mode? A Claude Code feature that activates when Anthropic employees use Claude Code on non-Anthropic repositories. It strips Co-Authored-By attribution from commits, blocks Slack-channel references, and enforces the forbidden-model-names filter that contains sonnet-4-8. This is real and shipping; whether end-users will see any equivalent is unclear.

Should I migrate from Sonnet 4.6 in anticipation of 4.8? No. Anthropic launches Sonnet versions backward-compatible on the API; migration is typically a model-ID change, not a refactor. Stay on 4.6 until the model card and benchmarks land, then evaluate against your specific workload.

What's the cheapest alternative to Sonnet 4.6 today? DeepSeek V4-Pro at $0.435/$0.87 per MTok (post-permanent-cut, 17x cheaper on output) and Qwen 3.6-Plus at $0.325/$1.95 per MTok (8x cheaper on output) are the two strongest cost-per-task alternatives. Both have published benchmarks competitive with Sonnet on coding workloads. See TokenMix.ai's cost-per-task analysis for the full math.

Will Sonnet 4.8 support 1M context like Opus 4.7? The leaked source map does not specify. Sonnet 4.6 already supports a separately-priced 1M context tier. Whether Sonnet 4.8 inherits the default 200K + 1M premium tier structure or unifies is unknown.

Sources

• Anthropic Claude Sonnet 4.6 (current GA) — Anthropic
• Anthropic Claude Pricing — Anthropic
• NxCode — Claude Sonnet 4.8 release date features what to expect 2026 — pro-leak interpretation
• WaveSpeed — Claude Sonnet 4.8 leak vs reality — skeptical interpretation
• Claudefast — Claude Code source leak everything found — leak inventory
• Medium — The Claude Code Leak (Mfierce) — narrative analysis
• TokenMix.ai Claude API Pricing tracker
• TokenMix.ai DeepSeek V4-Pro migration framework
• TokenMix.ai Qwen 3.6 tier picker

Related Articles

• Claude API Pricing 2026: Opus, Sonnet, Haiku Costs Compared
• Claude API Cache Pricing 2026: 90% Input Savings Explained
• Cheapest Frontier LLM API 2026: DeepSeek vs Claude vs GPT Cost
• DeepSeek V4-Pro 75% Cut: When to Migrate from Claude or GPT
• Qwen 3.6 Tier Picker 2026: Max-Preview vs Plus vs Flash vs 35B

Author: TokenMix Research Lab · Last Updated: 2026-05-25 · Data Checked: 2026-05-25