Back to Blog
Anthropic

Claude Sonnet 4.8: What Leaks Reveal About Anthropic's Next Major Model

The Mystery of the Version Number: Why Skip 4.7?

The Anthropic Sonnet series has progressed at a steady pace:

ModelReleaseAPI String
Claude 3.5 SonnetJune 2024claude-3-5-sonnet-20240620
Claude 3.7 SonnetFebruary 2025claude-3-7-sonnet-20250219
Claude 4 SonnetMay 2025claude-sonnet-4-0
Claude Sonnet 4.5September 2025claude-sonnet-4-5
Claude Sonnet 4.6February 2026claude-sonnet-4-6
Claude Sonnet 4.8May 2026 (Expected)claude-sonnet-4-8

Skipping version 4.7 suggests this is not a simple port from Opus 4.7. It likely originates from a different development timeline and training run.

Expected Specifications

Based on leaked information and extrapolating from Opus 4.7's improvements:

FeatureSonnet 4.6 (Current)Sonnet 4.8 (Expected)
SWE-bench Verified79.6%82-84%
GPQA Diamond74.1%76-78%
Image Resolution Limit~1.25MP3.75MP (3x)
Context Window1M (beta)1M (possible GA)
Price (Input/Output)$3/$15$3/$15 (likely unchanged)
Knowledge CutoffAugust 2025Late 2025 – Early 2026

SWE-bench Verified: Basis for the 82-84% Estimate

Sonnet 4.6 scored 79.6% on SWE-bench Verified, nearly matching Opus 4.6 (80.8%). Opus 4.7 reached 87.6%, showing a ~7-point improvement over Opus 4.6.

If Sonnet 4.8 achieves a similar improvement margin (6-8 points) from Sonnet 4.6, 82-84% becomes a reasonable estimate.

For context, the top SWE-bench Verified scores as of May 25, 2026:

RankModelScore
1Claude Mythos Preview93.9%
2Claude Opus 4.787.6%
3GPT-5.3 Codex85.0%
4Claude Opus 4.580.9%
5Claude Opus 4.680.8%
5DeepSeek V4 Pro (Max)80.6%
7Gemini 3.1 Pro80.6%
8Kimi K2.680.2%
8MiniMax M2.580.2%
10Claude Sonnet 4.679.6%

Achieving 82-84% would place Sonnet 4.8 ahead of Opus 4.5 and 4.6, and close to Opus 4.7. This would deliver near-Opus level coding performance at the Sonnet price tier ($3/$15).

Vision: 3.75MP Image Support

Opus 4.7 achieved "98.5% visual accuracy" and supports 3.75MP image resolution. Sonnet 4.8 is expected to bring this capability to the mid-tier price point.

Moving from Sonnet 4.6's ~1.25MP to 3.75MP—a 3x increase in resolution—would significantly boost practicality for document analysis, UI mockup understanding, and chart interpretation.

New "xhigh" Effort Level

The "xhigh" effort level (between "high" and "max") introduced with Opus 4.7 is also expected to be ported to Sonnet 4.8.

Current levels: low / medium / high / max

Adding xhigh allows for finer control over the cost-performance tradeoff.

Tokenizer Update

Opus 4.7's new tokenizer generates between 1.0x to 1.35x more tokens depending on content type:

Content TypeToken Increase Rate
English Prose~1.0x (unchanged)
Code~1.1-1.2x
Structured Data (JSON, XML)Up to 1.35x

If Sonnet 4.8 adopts this tokenizer, effective costs could rise by 10-35%. Even if the API price remains $3/$15, the same content would be split into more tokens.

Key Improvements Over Sonnet 4.6

1. Dramatically Enhanced Vision Capabilities

Expected visual abilities inherited from Opus 4.7:

  • Image resolution: 1.25MP → 3.75MP
  • Visual accuracy: 98.5% (measured on Opus 4.7)
  • High-accuracy understanding of document screenshots, UI mockups, and charts

2. Improved Coding Performance

Opus 4.7's improvement track record:

  • CursorBench: 58% → 70% (+12 points)
  • 93-task coding benchmark: +13%
  • Production task resolution: 3x increase

Sonnet 4.8 is expected to bring these gains to the $3/$15 price tier.

3. Shift in Prompt Interpretation

Opus 4.7 is described as interpreting prompts "more literally than 4.6." A similar behavioral shift is anticipated for Sonnet 4.8. While this means more precise instruction following, it may reduce flexibility for ambiguous prompts.

Pricing: Stability in the Sonnet Tier

The Sonnet series has consistently maintained the $3/$15 pricing from 3.5 through 4.6:

ModelInput/1MOutput/1M
Claude 3.5 Sonnet$3$15
Claude 4 Sonnet$3$15
Claude Sonnet 4.5$3$15
Claude Sonnet 4.6$3$15
Claude Sonnet 4.8 (Expected)$3$15

Meanwhile, the Opus tier saw a significant price cut with version 4.6 (from $15/$75 to $5/$25). If Sonnet 4.8 maintains $3/$15, it would be a model with 82-84% SWE-bench performance at 60% of Opus 4.7's price.

Related Leak Information

KAIROS

Mentioned over 150 times in leaked Claude Code files, KAIROS appears to be a persistent daemon agent. Features include autonomous monitoring, memory consolidation ("autoDream"), preemptive actions, and push notifications.

Mythos

Released April 7, 2026. Achieved 93.9% on SWE-bench Verified. Its use is limited to cybersecurity (Project Glasswing).

numbat

Mentioned in leaked code but details are unknown for this unreleased model.

Developer Expectations

Consolidating developer hopes for Sonnet 4.8:

  1. More reliable long-form output generation
  2. Reduced over-engineering (less addition of unnecessary complexity)
  3. Improved Time-to-First-Token (TTFT)
  4. Better multi-file awareness in large codebases
  5. Stable JSON/structured output and constrained decoding

Release Timeline

Predictions based on leaked information:

ScenarioTimingProbability
Best CaseLate April 2026Low
Most LikelyMay 5-16, 2026Medium
Worst CaseLate May 2026Medium

Given Opus 4.7 launched on April 16, Sonnet versions typically follow 3-4 weeks later. Mid-May 2026 is the most probable release window.

Competitive Landscape

Sonnet 4.8's positioning:

ModelInput/1MOutput/1MSWE-benchHighlights
Claude Sonnet 4.8 (Expected)$3$1582-84%Vision 3.75MP, 1M context
Claude Sonnet 4.6$3$1579.6%Current workhorse
GPT-5.2$1.25$10.0080.0%OpenAI mainstay
DeepSeek V4 Pro (Max)$1.74$3.4880.6%Most affordable frontier
Gemini 3.1 Pro$2.50$15.0080.6%Google mainstay
MiniMax M2.5$0.15$1.1580.2%Coding-focused, cheapest

If it hits 82-84%, Sonnet 4.8 would become the strongest coding model in the Sonnet tier. It would outperform Opus 4.5 (80.9%) and Opus 4.6 (80.8%) at a fraction of the cost.

Conclusion

Claude Sonnet 4.8 has the potential to be Anthropic's most significant leap in the Sonnet line yet.

It promises to bring Opus 4.7's Vision capabilities, coding performance, and effort levels to the $3/$15 price bracket. Should it achieve 82-84% on SWE-bench Verified, it would deliver Opus-level performance at Sonnet pricing.

While these predictions are based on leaks, reviewing Anthropic's Sonnet series history suggests such an improvement is entirely plausible. We'll have to wait for the official announcement, but this is undoubtedly one of the most anticipated models for developers.

Comments (0)

Share:XHatena

Post a Comment

Loading...